Files
llama-3-8b-base-new-dpo-hh-…/margin_logs/margins.jsonl
ModelHub XC ab10c6b4c0 初始化项目,由ModelHub XC社区提供模型
Model: jackf857/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun
Source: Original Platform
2026-05-10 13:35:35 +08:00

682 lines
1.1 MiB

{"epoch": 0.0, "step": 1, "batch_size": 64, "mean": -0.02287048101425171, "std": 0.42023447155952454, "min": -1.4034271240234375, "p10": -0.46674575805664065, "median": 0.04234886169433594, "p90": 0.4323463439941407, "max": 0.89263916015625, "pos_frac": 0.53125, "sample": [-0.06523895263671875, 0.436798095703125, 0.27811431884765625, -0.9194221496582031, 0.018890380859375, 0.20587158203125, 0.18878173828125, -0.3968696594238281, 0.26206207275390625, 0.2470550537109375, -0.040912628173828125, 0.4394989013671875, -0.44133758544921875, -0.39148712158203125, 0.2764854431152344, 0.89263916015625, -0.42584991455078125, -0.46125030517578125, -0.8638992309570312, -0.3508758544921875, 0.371368408203125, 0.887847900390625, -0.382904052734375, 0.36145782470703125, -0.4890003204345703, 0.052455902099609375, -0.036136627197265625, 0.23079299926757812, 0.2469482421875, 0.1643218994140625, -0.07129669189453125, 0.2790794372558594, 0.3637123107910156, -0.8916168212890625, 0.03298759460449219, -0.2790107727050781, -0.17860984802246094, 0.23892593383789062, 0.05171012878417969, -0.2564239501953125, -0.14655303955078125, 0.27777862548828125, 0.0810394287109375, -1.4034271240234375, -0.28739166259765625, -0.1489429473876953, 0.44918060302734375, 0.1693286895751953, 0.10933303833007812, -0.14766693115234375, -0.40944671630859375, -0.18532562255859375, 0.6261310577392578, -0.20856857299804688, 0.602569580078125, 0.05538177490234375, 0.1505279541015625, 0.1313800811767578, -0.006317138671875, 0.42195892333984375, -0.29936981201171875, -0.4691009521484375, 0.16705322265625, -0.5789260864257812], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000001.npy"}
{"epoch": 0.0014684287812041115, "step": 2, "batch_size": 64, "mean": -0.06572240591049194, "std": 0.3523969054222107, "min": -0.9291305541992188, "p10": -0.46334152221679686, "median": -0.05502510070800781, "p90": 0.3672500610351563, "max": 1.0444793701171875, "pos_frac": 0.4375, "sample": [-0.2829437255859375, 0.3027191162109375, -0.19867706298828125, -0.3062286376953125, 0.10318756103515625, 0.20131683349609375, -0.34906005859375, 0.2802886962890625, 0.1914520263671875, -0.31072998046875, 0.08922195434570312, 0.10284614562988281, -0.03655242919921875, -0.0604095458984375, -0.06208038330078125, 0.32562255859375, -0.37982177734375, 0.2746162414550781, -0.049640655517578125, 0.3752174377441406, -0.103973388671875, 0.0699462890625, 0.36417388916015625, -0.033428192138671875, 0.37265777587890625, -0.3787078857421875, -0.6610565185546875, 0.4720420837402344, 0.47701263427734375, -0.27928924560546875, -0.44719696044921875, -0.0965118408203125, -0.7628555297851562, 0.046764373779296875, 0.06670379638671875, -0.9291305541992188, -0.7122802734375, -0.16554832458496094, 0.1485595703125, -0.07539939880371094, 0.2588920593261719, 0.039890289306640625, 0.201690673828125, 0.0623016357421875, 1.0444793701171875, -0.37696075439453125, -0.02794647216796875, -0.223297119140625, -0.35730743408203125, -0.1309051513671875, -0.3106689453125, -0.11409187316894531, -0.1669769287109375, 0.131317138671875, -0.2361297607421875, 0.4093780517578125, -0.6485977172851562, 0.36856842041015625, -0.1951904296875, -0.4702606201171875, -0.7624168395996094, 0.008928298950195312, -0.31630706787109375, 0.022550582885742188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000002.npy"}
{"epoch": 0.002936857562408223, "step": 3, "batch_size": 64, "mean": 0.005800485610961914, "std": 0.3519676923751831, "min": -1.0667724609375, "p10": -0.43715362548828124, "median": 0.021262168884277344, "p90": 0.4142408370971681, "max": 0.7697525024414062, "pos_frac": 0.53125, "sample": [0.0335540771484375, -0.35327911376953125, -0.00292205810546875, 0.7697525024414062, -0.2964191436767578, 0.008512496948242188, 0.2404327392578125, 0.16851806640625, -0.1563262939453125, 0.009889602661132812, -1.0667724609375, 0.2082538604736328, 0.56268310546875, -0.4325714111328125, -0.08674240112304688, 0.301727294921875, 0.3809356689453125, 0.032634735107421875, 0.24275970458984375, 0.08998489379882812, -0.5494308471679688, 0.033538818359375, 0.7362594604492188, 0.153656005859375, 0.32257080078125, -0.4064655303955078, 0.1155242919921875, 0.4285144805908203, 0.251312255859375, -0.34081268310546875, 0.5052871704101562, 0.3284454345703125, -0.10449409484863281, 0.311676025390625, -0.051868438720703125, -0.0801239013671875, -0.1947021484375, -0.10048675537109375, -0.5119857788085938, -0.09004974365234375, 0.16557693481445312, 0.10590362548828125, -0.5589141845703125, -0.19528961181640625, 0.14806365966796875, 0.17899703979492188, -0.6941757202148438, 0.3196830749511719, 0.14305877685546875, -0.24903106689453125, 0.1640605926513672, -0.3004341125488281, 0.30182647705078125, 0.1408843994140625, -0.105560302734375, 0.45755767822265625, -0.010498046875, -0.16989898681640625, -0.344085693359375, 0.6761932373046875, -0.002391815185546875, -0.439117431640625, -0.2281951904296875, -0.5439529418945312], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000003.npy"}
{"epoch": 0.004405286343612335, "step": 4, "batch_size": 64, "mean": -0.033598393201828, "std": 0.37999019026756287, "min": -0.9428482055664062, "p10": -0.49340057373046875, "median": -0.027309417724609375, "p90": 0.44341983795166023, "max": 1.1188392639160156, "pos_frac": 0.484375, "sample": [-0.4453468322753906, -0.11343765258789062, -0.15898895263671875, -0.9428482055664062, 0.09244537353515625, 0.3075218200683594, 0.457275390625, -0.719146728515625, -0.3920783996582031, -0.32115936279296875, 0.0196990966796875, 1.1188392639160156, -0.19579315185546875, 0.15241241455078125, -0.055816650390625, -0.05597686767578125, -0.09604644775390625, -0.416778564453125, -0.29833984375, 0.016710281372070312, -0.41786956787109375, 0.1021270751953125, -0.492279052734375, -0.3339996337890625, -0.30890655517578125, -0.5379638671875, 0.4896087646484375, 0.13695526123046875, 0.05931854248046875, -0.03564453125, -0.12433242797851562, 0.15996932983398438, -0.55902099609375, -0.01897430419921875, -0.1384449005126953, 0.3156261444091797, -0.4938812255859375, 0.68841552734375, 0.01534271240234375, 0.1091461181640625, 0.1005859375, -0.1928424835205078, 0.45218658447265625, 0.39111328125, 0.09183502197265625, 0.07952499389648438, 0.2761077880859375, 0.3504180908203125, 0.17659950256347656, -0.766876220703125, 0.4792003631591797, 0.20528030395507812, 0.42296409606933594, -0.570159912109375, -0.19019699096679688, 0.12895774841308594, 0.841949462890625, -0.193267822265625, 0.1990966796875, 0.029170989990234375, -0.2363433837890625, -0.2537403106689453, -0.28570556640625, -0.25449371337890625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000004.npy"}
{"epoch": 0.005873715124816446, "step": 5, "batch_size": 64, "mean": 0.06242257356643677, "std": 0.4240538775920868, "min": -1.5327606201171875, "p10": -0.3658599853515625, "median": 0.024437904357910156, "p90": 0.6044029235839844, "max": 0.96142578125, "pos_frac": 0.546875, "sample": [-0.5433425903320312, 0.96142578125, 0.18933868408203125, 0.10531997680664062, -0.3305206298828125, -0.2437896728515625, 0.3372955322265625, 0.18900299072265625, 0.0055027008056640625, -0.07151031494140625, -0.21562957763671875, -0.370758056640625, 0.017278671264648438, 0.070098876953125, 0.008632659912109375, -0.13076019287109375, -0.00884246826171875, -0.246246337890625, -0.055248260498046875, -0.5170135498046875, -0.08971977233886719, 0.888885498046875, 0.21389389038085938, 0.3806915283203125, 0.11779403686523438, 0.3567962646484375, 0.951416015625, 0.031597137451171875, -0.0372161865234375, 0.429107666015625, -0.12975692749023438, 0.2901134490966797, 0.66864013671875, 0.4660167694091797, 0.08536148071289062, 0.09478759765625, 0.9041976928710938, -0.16809844970703125, 0.30997467041015625, 0.5900421142578125, -0.168212890625, -0.01015472412109375, 0.3917999267578125, -1.5327606201171875, 0.9378890991210938, -0.01668548583984375, -0.454345703125, 0.35263824462890625, -0.2663726806640625, -0.8276519775390625, 0.15118026733398438, 0.10655021667480469, 0.6105575561523438, -0.0010223388671875, -0.02933502197265625, -0.1510009765625, 0.1322021484375, -0.38355255126953125, -0.2837677001953125, -0.030033111572265625, -0.35443115234375, 0.08844757080078125, 0.073455810546875, 0.1548919677734375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000005.npy"}
{"epoch": 0.007342143906020558, "step": 6, "batch_size": 64, "mean": -0.00980466604232788, "std": 0.38482362031936646, "min": -1.038604736328125, "p10": -0.5166404724121094, "median": 0.019227981567382812, "p90": 0.4275375366210938, "max": 0.9494476318359375, "pos_frac": 0.515625, "sample": [-0.593353271484375, 0.10240554809570312, -0.06489944458007812, 0.1634674072265625, 0.02890777587890625, 0.29541015625, -0.6636428833007812, -0.9975662231445312, 0.018707275390625, -0.451934814453125, 0.4663581848144531, 0.3572578430175781, -0.0751800537109375, 0.41591835021972656, -0.07769775390625, -0.5072250366210938, 0.1884613037109375, -0.11884498596191406, 0.18921279907226562, 0.5783805847167969, -0.338226318359375, -0.3291645050048828, 0.419921875, -0.009946823120117188, 0.054439544677734375, 0.7589111328125, 0.0549468994140625, 0.4308013916015625, 0.05054473876953125, -0.10179901123046875, 0.16371917724609375, -0.1841583251953125, 0.6753997802734375, 0.019748687744140625, -0.17523193359375, -0.1574554443359375, -0.6075592041015625, -0.17722320556640625, -0.201385498046875, -0.20154571533203125, -0.145782470703125, -0.03216552734375, 0.19234466552734375, 0.1781902313232422, 0.32108116149902344, 0.13829994201660156, 0.3152008056640625, 0.044162750244140625, 0.07403564453125, -0.30670928955078125, -0.5206756591796875, 0.35955047607421875, 0.09241485595703125, 0.26088714599609375, 0.9494476318359375, -0.2233257293701172, -0.2579994201660156, -1.038604736328125, -0.025299072265625, 0.346221923828125, -0.4052276611328125, -0.14322280883789062, 0.4518585205078125, -0.65106201171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000006.npy"}
{"epoch": 0.00881057268722467, "step": 7, "batch_size": 64, "mean": 0.023058056831359863, "std": 0.4159083366394043, "min": -1.0472183227539062, "p10": -0.43124408721923824, "median": 0.0045642852783203125, "p90": 0.4744728088378907, "max": 1.40472412109375, "pos_frac": 0.5, "sample": [-0.11658096313476562, -1.0472183227539062, -0.06951522827148438, 0.16102981567382812, -0.5786056518554688, 0.1803436279296875, 0.10532760620117188, 0.359710693359375, -0.2848663330078125, 0.456512451171875, -0.828125, 0.14754486083984375, 0.06346893310546875, 0.48217010498046875, -0.1818695068359375, -0.1889495849609375, -0.0496826171875, -0.22151947021484375, -0.014596939086914062, -0.646636962890625, 0.35599517822265625, 0.1238250732421875, 0.7468452453613281, 0.38196563720703125, 0.10888671875, 0.04175567626953125, -0.4448986053466797, 0.05414390563964844, -0.37831878662109375, -0.399383544921875, 0.08155441284179688, 0.4122123718261719, 0.6601753234863281, -0.72552490234375, 0.2357940673828125, -0.1737823486328125, -0.1260223388671875, -0.06254768371582031, -0.234130859375, 0.022434234619140625, -0.4911651611328125, -0.221405029296875, 0.31501007080078125, -0.2638092041015625, 0.14744949340820312, 0.8439483642578125, 0.2331695556640625, -0.0276031494140625, 0.0470733642578125, -0.0133056640625, 1.40472412109375, -0.11316680908203125, 0.2669677734375, 0.6204681396484375, -0.01567840576171875, -0.18413543701171875, -0.33535003662109375, 0.2731666564941406, 0.9337120056152344, 0.10991287231445312, 0.22198486328125, -0.154998779296875, -0.19730377197265625, -0.3328704833984375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000007.npy"}
{"epoch": 0.010279001468428781, "step": 8, "batch_size": 64, "mean": 0.013962745666503906, "std": 0.3936125636100769, "min": -1.5029449462890625, "p10": -0.3637277603149414, "median": 0.011749267578125, "p90": 0.46925945281982423, "max": 1.250732421875, "pos_frac": 0.546875, "sample": [0.137176513671875, 0.6202392578125, -0.3318023681640625, -0.054027557373046875, -0.29863739013671875, 0.05381011962890625, -0.010974884033203125, 0.1625213623046875, 0.584136962890625, 0.113372802734375, 0.4683341979980469, 0.06198883056640625, -0.11213302612304688, 0.09533882141113281, -0.09976959228515625, 0.06607818603515625, -0.009675979614257812, 0.3457679748535156, 0.1115570068359375, 0.03717041015625, -0.290924072265625, -1.5029449462890625, 0.0058746337890625, 0.6599578857421875, -0.028203964233398438, -0.46501922607421875, 0.0611419677734375, 0.05724143981933594, 0.246337890625, -0.32181549072265625, -0.12213897705078125, -0.4609527587890625, 0.2611732482910156, -0.09546661376953125, 0.00743865966796875, -0.11254119873046875, 0.01605987548828125, 0.2880134582519531, -0.12088203430175781, -0.10187911987304688, 0.8216781616210938, -0.18140411376953125, -0.08766937255859375, 0.1807861328125, 0.1107940673828125, -0.827484130859375, -0.7289581298828125, 0.0486602783203125, 0.15288734436035156, -0.00270843505859375, -0.3513603210449219, 0.000640869140625, 0.5906982421875, -0.2326812744140625, 1.250732421875, 0.10080718994140625, -0.12099456787109375, -0.17902374267578125, 0.434112548828125, -0.4224395751953125, -0.36902809143066406, 0.46965599060058594, 0.03570556640625, 0.279266357421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000008.npy"}
{"epoch": 0.011747430249632892, "step": 9, "batch_size": 64, "mean": 0.05677860975265503, "std": 0.47101157903671265, "min": -0.8725738525390625, "p10": -0.6059989929199219, "median": 0.06195831298828125, "p90": 0.6589889526367188, "max": 1.2829513549804688, "pos_frac": 0.53125, "sample": [0.271514892578125, -0.33512115478515625, 0.8424854278564453, 0.26886749267578125, 0.31830596923828125, -0.21431732177734375, 0.6517791748046875, -0.11437225341796875, 1.0064163208007812, 0.38811492919921875, -0.4627265930175781, -0.3310546875, -0.1289215087890625, -0.5716323852539062, 0.054107666015625, -0.6207275390625, -0.33603668212890625, -0.035861968994140625, -0.643218994140625, -0.07541275024414062, -0.6832199096679688, 0.11523056030273438, -0.45153045654296875, 0.27978515625, 1.2829513549804688, -0.37804412841796875, 0.2759208679199219, -0.7047119140625, -0.2215862274169922, -0.13326263427734375, 0.062255859375, 0.21598243713378906, 0.2869110107421875, 0.22611236572265625, 0.114105224609375, -0.7261276245117188, -0.1758747100830078, -0.16722488403320312, -0.23919677734375, 0.9144058227539062, -0.7255020141601562, 0.36418914794921875, -0.8725738525390625, -0.024560928344726562, 0.18211936950683594, 0.5279617309570312, 0.169830322265625, 0.28558349609375, 0.5467071533203125, 0.11917686462402344, -0.0506744384765625, -0.19744873046875, 0.4750213623046875, 0.9350528717041016, 0.7753353118896484, 0.5813140869140625, 0.0616607666015625, 0.1014251708984375, -0.05914497375488281, 0.662078857421875, -0.40471649169921875, 0.5289230346679688, 0.1141204833984375, -0.28711700439453125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000009.npy"}
{"epoch": 0.013215859030837005, "step": 10, "batch_size": 64, "mean": 0.004975318908691406, "std": 0.37408262491226196, "min": -1.1198654174804688, "p10": -0.4508584976196289, "median": 0.028249740600585938, "p90": 0.43997268676757834, "max": 0.9347991943359375, "pos_frac": 0.53125, "sample": [0.15612030029296875, -0.10178375244140625, -0.3938407897949219, -0.0147247314453125, -0.4303455352783203, 0.13311004638671875, 0.01573944091796875, 0.38719940185546875, 0.335052490234375, 0.3475990295410156, 0.05132293701171875, 0.13439178466796875, -0.37039947509765625, 0.626617431640625, -0.569091796875, 0.08172607421875, -0.10635757446289062, -0.3698577880859375, 0.09270477294921875, 0.20258522033691406, 0.6090850830078125, 0.4614372253417969, 0.19974517822265625, -0.04021263122558594, -0.05493927001953125, -0.051418304443359375, -1.1198654174804688, 0.2053985595703125, 0.06237983703613281, -0.3593902587890625, 0.3659820556640625, -0.212646484375, -0.14468002319335938, -0.17816925048828125, -0.21663665771484375, -0.5753059387207031, -0.581817626953125, 0.3191680908203125, 0.15062522888183594, 0.18482589721679688, -0.2987709045410156, -0.4779205322265625, -0.38527488708496094, -0.459869384765625, 0.100982666015625, 0.816192626953125, 0.6722526550292969, 0.040760040283203125, -0.1763763427734375, 0.3689117431640625, -0.454864501953125, 0.00478363037109375, 0.06456756591796875, 0.3898887634277344, -0.34569549560546875, -0.4415111541748047, -0.13968658447265625, -0.0935516357421875, 0.31565093994140625, 0.4771728515625, 0.9347991943359375, -0.10943412780761719, 0.16532135009765625, 0.1187591552734375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000010.npy"}
{"epoch": 0.014684287812041116, "step": 11, "batch_size": 64, "mean": 0.047412097454071045, "std": 0.38612228631973267, "min": -0.8685493469238281, "p10": -0.4965469360351562, "median": 0.06329727172851562, "p90": 0.44717178344726577, "max": 0.8984375, "pos_frac": 0.59375, "sample": [-0.517425537109375, 0.7832489013671875, 0.2631244659423828, -0.15761947631835938, -0.6115493774414062, 0.35193634033203125, 0.6544952392578125, 0.1374378204345703, 0.08857345581054688, 0.4179534912109375, -0.10704612731933594, -0.06568145751953125, 0.0035858154296875, 0.45969390869140625, 0.1560821533203125, -0.4478302001953125, 0.32599639892578125, 0.3071937561035156, -0.525238037109375, 0.1998920440673828, 0.22396469116210938, 0.4759521484375, -0.5441379547119141, 0.4092445373535156, -0.12535858154296875, -0.05582427978515625, 0.05272674560546875, -0.2465972900390625, -0.061065673828125, 0.21456146240234375, 0.8319320678710938, -0.8683929443359375, 0.3589897155761719, -0.3668670654296875, 0.0377349853515625, 0.0091400146484375, 0.361419677734375, 0.8984375, 0.34299468994140625, -0.1693897247314453, 0.04089164733886719, 0.18939590454101562, -0.39444732666015625, -0.042022705078125, 0.3458213806152344, 0.27789306640625, 0.3614044189453125, -0.1441192626953125, -0.746795654296875, 0.367431640625, -0.2978668212890625, -0.4408721923828125, 0.165130615234375, -0.8685493469238281, 0.0738677978515625, -0.1598796844482422, -0.25754547119140625, -0.06661224365234375, 0.12679290771484375, 0.6802291870117188, 0.223297119140625, 0.122406005859375, 0.003955841064453125, -0.02172088623046875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000011.npy"}
{"epoch": 0.016152716593245228, "step": 12, "batch_size": 64, "mean": -0.04085996747016907, "std": 0.37206172943115234, "min": -0.854217529296875, "p10": -0.6071647644042969, "median": -0.015848159790039062, "p90": 0.3713495254516602, "max": 0.7443389892578125, "pos_frac": 0.5, "sample": [0.7443389892578125, -0.6785125732421875, 0.09807586669921875, -0.6412200927734375, -0.540863037109375, 0.2344207763671875, 0.21481704711914062, 0.23077774047851562, 0.314178466796875, 0.5028305053710938, -0.2517242431640625, -0.6646881103515625, -0.0571746826171875, 0.0202178955078125, -0.24477005004882812, 0.25034332275390625, -0.11771392822265625, -0.17376708984375, 0.2491912841796875, -0.1399383544921875, -0.34881591796875, -0.34537506103515625, 0.24981689453125, -0.6017227172851562, -0.6300811767578125, 0.3904228210449219, 0.198333740234375, 0.0709381103515625, -0.13858604431152344, -0.6094970703125, -0.4166107177734375, 0.7355499267578125, -0.0351409912109375, 0.12013435363769531, -0.2207183837890625, -0.07977485656738281, -0.10595703125, -0.3355712890625, -0.5860214233398438, -0.10651588439941406, -0.8165359497070312, 0.3034820556640625, -0.09386253356933594, 0.08877182006835938, -0.114501953125, 0.37567138671875, -0.854217529296875, 0.0838470458984375, 0.09562873840332031, 0.3222846984863281, 0.12089347839355469, -0.09921073913574219, -0.482421875, 0.3612651824951172, -0.19854736328125, 0.197540283203125, 0.0307464599609375, -0.035953521728515625, 0.003444671630859375, 0.03052520751953125, 0.6503639221191406, 0.16332244873046875, 0.5933837890625, 0.10541534423828125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000012.npy"}
{"epoch": 0.01762114537444934, "step": 13, "batch_size": 64, "mean": 0.060471951961517334, "std": 0.4298148453235626, "min": -1.564666748046875, "p10": -0.401959228515625, "median": 0.09225082397460938, "p90": 0.4119892120361328, "max": 1.35198974609375, "pos_frac": 0.640625, "sample": [0.8122100830078125, -0.08889961242675781, 0.13982391357421875, -0.01483917236328125, 0.09279632568359375, 0.6008262634277344, 0.17205047607421875, 0.1695556640625, 0.25879669189453125, -0.3642692565917969, 0.18184661865234375, 0.7881393432617188, 0.3498992919921875, -0.26146697998046875, 0.018556594848632812, 0.9858016967773438, -0.425506591796875, 0.091705322265625, 0.08727836608886719, 0.20442962646484375, -0.09586143493652344, 0.31463623046875, 0.13063812255859375, -0.0679168701171875, 0.11096572875976562, -0.6929168701171875, 0.0892181396484375, -0.18294715881347656, -0.22381019592285156, 0.000335693359375, -0.6834945678710938, 0.069854736328125, 0.08778762817382812, 0.19446182250976562, -1.564666748046875, 0.13704681396484375, 0.18032073974609375, -0.004589080810546875, -0.010478973388671875, 0.3236083984375, 0.4087715148925781, 0.1663036346435547, 0.788787841796875, 0.32820892333984375, 0.210113525390625, 0.002727508544921875, -0.09546470642089844, 0.028413772583007812, 0.3551597595214844, -0.14691162109375, -0.142242431640625, 0.41336822509765625, 0.1583404541015625, -0.1908283233642578, -0.821746826171875, -0.6139602661132812, 0.16379547119140625, -0.40547943115234375, 1.35198974609375, -0.2508544921875, 0.379913330078125, 0.0936126708984375, 0.1710052490234375, -0.39374542236328125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000013.npy"}
{"epoch": 0.01908957415565345, "step": 14, "batch_size": 64, "mean": 0.009781479835510254, "std": 0.428362637758255, "min": -1.139678955078125, "p10": -0.5108818054199219, "median": -0.009484291076660156, "p90": 0.5680206298828125, "max": 1.0740966796875, "pos_frac": 0.484375, "sample": [-0.8328475952148438, -0.011350631713867188, 0.40860748291015625, -0.041271209716796875, 1.0740966796875, -0.3880424499511719, -0.370513916015625, -0.2270660400390625, 0.23455047607421875, 0.11051559448242188, -0.07878875732421875, -0.22779464721679688, -0.2107696533203125, 0.5586318969726562, -0.43599700927734375, -0.6060943603515625, -0.28362464904785156, -0.1390380859375, 0.10763168334960938, -0.1128692626953125, 0.7933082580566406, 0.23927688598632812, -0.09074020385742188, 0.20424270629882812, 0.7789459228515625, -0.4923820495605469, -0.557464599609375, 0.4519195556640625, -0.120025634765625, 0.6726303100585938, 0.269622802734375, -0.049243927001953125, 0.5401382446289062, 0.5970802307128906, 0.15477371215820312, 0.18162918090820312, 0.2154388427734375, 0.3665771484375, -0.5188102722167969, -0.08289527893066406, -0.739166259765625, -0.28082275390625, -0.21105194091796875, 0.129486083984375, 0.10802841186523438, -0.007617950439453125, 0.7105140686035156, -0.53607177734375, 0.3782501220703125, -0.27246856689453125, -1.139678955078125, 0.057098388671875, 0.396270751953125, -0.19512176513671875, 0.07364273071289062, 0.4257965087890625, 0.5720443725585938, 0.1854095458984375, -0.3820343017578125, -0.3627166748046875, -0.41340065002441406, 0.3119392395019531, 0.08251953125, -0.3468208312988281], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000014.npy"}
{"epoch": 0.020558002936857563, "step": 15, "batch_size": 64, "mean": -0.010011672973632812, "std": 0.37274840474128723, "min": -0.8534603118896484, "p10": -0.5699829101562499, "median": 0.024845123291015625, "p90": 0.42551708221435547, "max": 0.7226104736328125, "pos_frac": 0.5625, "sample": [-0.023351669311523438, 0.498382568359375, 0.1649646759033203, 0.0001049041748046875, 0.6226425170898438, 0.18100357055664062, -0.057010650634765625, 0.4149627685546875, 0.0717620849609375, 0.0235595703125, -0.8320846557617188, 0.022134780883789062, 0.49383544921875, -0.6096343994140625, 0.29444313049316406, 0.317840576171875, 0.22745895385742188, -0.283935546875, 0.4199695587158203, -0.11997222900390625, 0.20682525634765625, -0.737823486328125, 0.42789459228515625, -0.0176544189453125, -0.8534603118896484, 0.0810394287109375, 0.45246124267578125, 0.7226104736328125, -0.06934356689453125, 0.03499603271484375, -0.5237655639648438, -0.5897903442382812, -0.375457763671875, -0.2995491027832031, -0.19647216796875, 0.306182861328125, -0.43810272216796875, 0.2887420654296875, -0.478179931640625, -0.14802932739257812, -0.06175422668457031, 0.037212371826171875, 0.10670280456542969, 0.21219635009765625, 0.31629180908203125, -0.052051544189453125, 0.094696044921875, -0.236480712890625, 0.25676727294921875, 0.2019500732421875, 0.27184295654296875, -0.24454116821289062, -0.04106903076171875, 0.23616790771484375, 0.012298583984375, 0.5230484008789062, -0.836944580078125, -0.45543670654296875, 0.02613067626953125, -0.043365478515625, -0.1606006622314453, -0.7840728759765625, 0.046916961669921875, 0.31314849853515625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000015.npy"}
{"epoch": 0.022026431718061675, "step": 16, "batch_size": 64, "mean": 0.08432334661483765, "std": 0.38263678550720215, "min": -0.8667144775390625, "p10": -0.40624008178710935, "median": 0.06780815124511719, "p90": 0.6050853729248048, "max": 0.98394775390625, "pos_frac": 0.578125, "sample": [0.32303810119628906, 0.3166961669921875, -0.42022705078125, -0.1241455078125, -0.1933746337890625, 0.3910865783691406, 0.317596435546875, 0.13353729248046875, 0.16170120239257812, 0.47496795654296875, -0.14825439453125, 0.611053466796875, -0.013824462890625, 0.25391387939453125, 0.18192672729492188, 0.1768646240234375, 0.06530380249023438, -0.17523956298828125, -0.47467041015625, -0.44832611083984375, 0.49078369140625, 0.23858070373535156, 0.7451553344726562, -0.0932464599609375, -0.10979461669921875, -0.22956085205078125, -0.37360382080078125, -0.05726051330566406, -0.23018646240234375, -0.18835067749023438, 0.22140884399414062, 0.2533721923828125, 0.0703125, -0.5180702209472656, -0.37340545654296875, 0.8030166625976562, 0.08031463623046875, -0.8667144775390625, 0.5859603881835938, 0.15133285522460938, 0.5911598205566406, 0.03702545166015625, 0.14491653442382812, 0.635009765625, 0.7170639038085938, 0.46427154541015625, 0.008296966552734375, -0.1594524383544922, -0.20024871826171875, 0.00531005859375, -0.08938980102539062, 0.17763519287109375, 0.404571533203125, -0.29045867919921875, -0.08681488037109375, 0.04046630859375, 0.12460708618164062, 0.397216796875, 0.9194107055664062, 0.98394775390625, -0.3072509765625, -0.484283447265625, -0.18487930297851562, -0.4611053466796875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000016.npy"}
{"epoch": 0.023494860499265784, "step": 17, "batch_size": 64, "mean": 0.14771583676338196, "std": 0.3264169692993164, "min": -0.9883575439453125, "p10": -0.1812385559082031, "median": 0.1518087387084961, "p90": 0.4676116943359375, "max": 1.0341796875, "pos_frac": 0.71875, "sample": [-0.189910888671875, 0.2005615234375, 0.47003936767578125, 0.13985443115234375, -0.14126968383789062, -0.19136810302734375, 0.3819427490234375, -0.9883575439453125, 0.08937835693359375, -0.38599395751953125, 0.2432098388671875, 0.12042236328125, 0.238861083984375, 0.26702880859375, 0.22000694274902344, 0.46846771240234375, 0.028530120849609375, 0.29916954040527344, 0.1659870147705078, 0.39946746826171875, -0.0259857177734375, 0.32049560546875, 0.28778839111328125, 0.140960693359375, 0.138885498046875, 0.07529449462890625, 0.46414947509765625, -0.0998077392578125, 0.8208465576171875, 0.4648017883300781, -0.5555038452148438, 0.380340576171875, 0.5941009521484375, 0.2740936279296875, 0.0069713592529296875, 0.6503982543945312, 0.04615020751953125, 0.1626567840576172, -0.0855560302734375, 0.3470458984375, 0.24239349365234375, -0.00164794921875, 0.46561431884765625, 0.1861438751220703, 0.0156402587890625, 0.091278076171875, -0.08019638061523438, -0.16100311279296875, 0.0211639404296875, -0.13692474365234375, 0.223419189453125, 0.045200347900390625, 0.4584503173828125, -0.022342681884765625, 0.10781097412109375, 0.375762939453125, 1.0341796875, -0.021778106689453125, 0.4093017578125, -0.38858795166015625, 0.24181556701660156, -0.5309982299804688, 0.6994552612304688, -0.06449127197265625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000017.npy"}
{"epoch": 0.024963289280469897, "step": 18, "batch_size": 64, "mean": 0.08384177088737488, "std": 0.39051568508148193, "min": -1.0850372314453125, "p10": -0.3893581390380859, "median": 0.08577346801757812, "p90": 0.5394050598144532, "max": 0.801513671875, "pos_frac": 0.578125, "sample": [-0.1910247802734375, -0.6930694580078125, 0.266845703125, 0.2932281494140625, 0.2816200256347656, -0.5477714538574219, -0.3978462219238281, -0.23724365234375, -0.3695526123046875, 0.5953922271728516, -0.2686614990234375, -0.019697189331054688, 0.39884185791015625, 0.3989753723144531, -0.0479278564453125, 0.46067047119140625, 0.2728118896484375, -0.0517730712890625, 0.5998153686523438, -0.03623199462890625, 0.2873058319091797, -0.18251419067382812, 0.06787109375, -0.05344581604003906, 0.7386131286621094, -0.019069671630859375, 0.30437469482421875, 0.5503616333007812, -0.265350341796875, 0.10527992248535156, 0.434051513671875, 0.07892227172851562, -0.4161796569824219, -0.31575775146484375, 0.801513671875, -0.18100357055664062, 0.00321197509765625, 0.16968727111816406, -0.6077957153320312, 0.4000091552734375, 0.49233436584472656, 0.509185791015625, -0.6964569091796875, 0.40602874755859375, 0.2927703857421875, 0.5138397216796875, -0.10344314575195312, -0.07878684997558594, 0.7828369140625, 0.8004875183105469, 0.058010101318359375, -0.14745521545410156, -0.1723175048828125, 0.2809562683105469, -0.08658599853515625, 0.4951324462890625, 0.2779388427734375, 0.09262466430664062, -1.0850372314453125, 0.14000892639160156, 0.050312042236328125, -0.298370361328125, 0.1290130615234375, 0.10536003112792969], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000018.npy"}
{"epoch": 0.02643171806167401, "step": 19, "batch_size": 64, "mean": 0.08694282174110413, "std": 0.36367225646972656, "min": -0.6956329345703125, "p10": -0.3243719100952148, "median": 0.030633926391601562, "p90": 0.4941173553466798, "max": 1.0932769775390625, "pos_frac": 0.53125, "sample": [0.4412384033203125, -0.09857559204101562, -0.18407440185546875, 0.19596099853515625, 0.3468170166015625, -0.01214599609375, -0.186737060546875, -0.12999343872070312, 0.1401996612548828, 0.5716705322265625, -0.0664825439453125, -0.10306167602539062, 0.5054206848144531, -0.5822982788085938, 0.342071533203125, -0.4810028076171875, 0.9364013671875, -0.1296234130859375, -0.6956329345703125, 0.3844451904296875, -0.2158203125, 0.2927360534667969, -0.026277542114257812, -0.0635223388671875, -0.02252197265625, 0.22168350219726562, 0.0310211181640625, 0.20795440673828125, -0.26982879638671875, 0.02960968017578125, 0.0587158203125, 0.5351829528808594, -0.0016651153564453125, 0.050037384033203125, 0.30611419677734375, 0.3330249786376953, 0.0612335205078125, -0.12748336791992188, -0.44167327880859375, -0.256591796875, -0.1717700958251953, 1.0932769775390625, 0.467742919921875, -0.25299072265625, 0.15015220642089844, 0.3230438232421875, 0.326202392578125, 0.35833740234375, 0.2395782470703125, 0.030246734619140625, -0.08330535888671875, -0.032489776611328125, -0.0521392822265625, -0.3282909393310547, 0.8639297485351562, 0.09527969360351562, 0.21189117431640625, 0.3780555725097656, 0.4266204833984375, -0.5654296875, 0.9220733642578125, -0.3152275085449219, -0.3810997009277344, -0.0358734130859375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000019.npy"}
{"epoch": 0.027900146842878122, "step": 20, "batch_size": 64, "mean": 0.05212897062301636, "std": 0.3416326642036438, "min": -1.1411895751953125, "p10": -0.29332275390625, "median": 0.09351444244384766, "p90": 0.41756515502929686, "max": 0.760772705078125, "pos_frac": 0.609375, "sample": [0.09257698059082031, 0.47820091247558594, 0.300567626953125, 0.5473175048828125, -0.15290069580078125, 0.10997390747070312, -0.10147857666015625, 0.1025848388671875, 0.418853759765625, -0.0467529296875, -0.9891891479492188, -0.25839996337890625, 0.3178138732910156, 0.0964813232421875, -0.184051513671875, 0.13983154296875, -0.2990264892578125, 0.3314628601074219, -0.5553531646728516, -0.3108673095703125, 0.25104522705078125, -1.1411895751953125, 0.01383209228515625, -0.10884666442871094, 0.25826263427734375, -0.0819854736328125, 0.0660400390625, 0.17408180236816406, -0.2075653076171875, -0.04683876037597656, 0.2287731170654297, -0.11189651489257812, -0.5431785583496094, 0.18646621704101562, 0.3230476379394531, -0.23947906494140625, 0.1300506591796875, -0.32186126708984375, 0.03951263427734375, -0.2800140380859375, 0.3954010009765625, 0.41455841064453125, -0.15836334228515625, 0.1358642578125, 0.0811309814453125, -0.04122734069824219, 0.094451904296875, 0.51519775390625, 0.3256072998046875, -0.21800994873046875, 0.39414215087890625, -0.2606658935546875, 0.3085041046142578, 0.325347900390625, 0.63507080078125, -0.10081291198730469, 0.19329833984375, 0.3292675018310547, 0.2565116882324219, 0.03223228454589844, -0.17367935180664062, 0.760772705078125, 0.46286773681640625, 0.0028839111328125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000020.npy"}
{"epoch": 0.02936857562408223, "step": 21, "batch_size": 64, "mean": 0.11469615995883942, "std": 0.32359999418258667, "min": -0.5274562835693359, "p10": -0.27943878173828124, "median": 0.09116649627685547, "p90": 0.5085479736328127, "max": 1.0457763671875, "pos_frac": 0.59375, "sample": [-0.05864715576171875, -0.1893463134765625, 0.19130706787109375, -0.25823974609375, 0.34590911865234375, 0.1922922134399414, -0.12420272827148438, 0.08484840393066406, 1.0457763671875, 0.4739837646484375, 0.19537353515625, 0.5233612060546875, 0.1521129608154297, 0.09748458862304688, 0.5769462585449219, 0.41546630859375, 0.04391288757324219, -0.025251388549804688, -0.15084266662597656, 0.020236968994140625, -0.09588623046875, 0.3808746337890625, 0.22600555419921875, -0.16678619384765625, -0.0124359130859375, 0.8024749755859375, 0.4361724853515625, -0.04599952697753906, -0.4456901550292969, 0.22853851318359375, -0.01120758056640625, 0.29813385009765625, -0.12007522583007812, 0.1665515899658203, -0.26043701171875, 0.08002853393554688, -0.4532966613769531, 0.2136707305908203, 0.7312774658203125, -0.2875823974609375, 0.35771942138671875, 0.13360595703125, -0.10448455810546875, -0.07402801513671875, -0.4429206848144531, -0.5274562835693359, 0.2746925354003906, 0.32497406005859375, 0.11684036254882812, -0.3944549560546875, 0.1470317840576172, 0.4720306396484375, 0.6046180725097656, 0.0615997314453125, 0.04017829895019531, -0.00045013427734375, 0.8183135986328125, -0.1569366455078125, 0.1854248046875, 0.3783283233642578, -0.30338287353515625, -0.029575347900390625, 0.25691986083984375, -0.0148468017578125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000021.npy"}
{"epoch": 0.030837004405286344, "step": 22, "batch_size": 64, "mean": 0.18543151021003723, "std": 0.4194871485233307, "min": -0.49565887451171875, "p10": -0.3287391662597656, "median": 0.09654998779296875, "p90": 0.7067947387695317, "max": 1.1383056640625, "pos_frac": 0.640625, "sample": [-0.3786163330078125, 0.504638671875, -0.24729156494140625, 0.06119728088378906, 0.55780029296875, -0.32946014404296875, -0.3675537109375, 0.08546829223632812, -0.15160751342773438, 0.32196044921875, 0.3994560241699219, 1.1383056640625, -0.081817626953125, 1.0774154663085938, -0.3020954132080078, -0.49565887451171875, 1.0742874145507812, 0.12480354309082031, -0.04120635986328125, -0.4365234375, 0.2072906494140625, 0.8833465576171875, -0.327056884765625, -0.1647491455078125, -0.16130447387695312, 0.16539382934570312, 1.0857315063476562, 0.7482452392578125, -0.04869842529296875, 0.0553436279296875, 0.35341644287109375, 0.02938079833984375, -0.10713958740234375, -0.2635536193847656, 0.26633262634277344, 0.6051254272460938, 0.2748756408691406, -0.232147216796875, -0.14162826538085938, -0.0321197509765625, -0.0053253173828125, 0.5378761291503906, 0.0382232666015625, 0.06816482543945312, 0.018602371215820312, 0.6006240844726562, 0.055065155029296875, -0.48311614990234375, 1.05517578125, 0.22521591186523438, 0.04559135437011719, 0.48917388916015625, 0.10763168334960938, -0.11097908020019531, 0.4150390625, 0.2125244140625, 0.14658355712890625, 0.38643646240234375, 0.41439056396484375, 0.610076904296875, 0.5858001708984375, 0.588653564453125, 0.570068359375, -0.4134674072265625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000022.npy"}
{"epoch": 0.032305433186490456, "step": 23, "batch_size": 64, "mean": 0.24831706285476685, "std": 0.4868762195110321, "min": -1.550048828125, "p10": -0.19577827453613275, "median": 0.21927642822265625, "p90": 0.8769014358520509, "max": 1.4889678955078125, "pos_frac": 0.6875, "sample": [-0.14325714111328125, -0.0881805419921875, 1.1219558715820312, -0.032688140869140625, 0.4900627136230469, 0.3324623107910156, 0.6836700439453125, -0.23163604736328125, -0.21715927124023438, -0.728729248046875, 0.510498046875, 1.4038162231445312, 0.1803131103515625, -0.10288238525390625, -0.035472869873046875, -0.0898284912109375, 0.3287467956542969, 0.21109771728515625, 0.3175506591796875, -0.07439804077148438, 0.576629638671875, 0.02725982666015625, 1.4889678955078125, 0.23350143432617188, -0.11982154846191406, -0.0733184814453125, 0.0917510986328125, -0.1458892822265625, -0.21846771240234375, 0.8959197998046875, 0.2197723388671875, 0.75927734375, 0.405029296875, 0.7826595306396484, 0.34944725036621094, 0.3314208984375, -0.22296905517578125, 0.09391021728515625, 0.23616600036621094, 0.8325252532958984, 0.08736228942871094, 0.06339645385742188, 0.263671875, 0.6520805358886719, 0.16897964477539062, 0.9912567138671875, 0.18117904663085938, -1.550048828125, 0.24710464477539062, -0.4751434326171875, -0.045989990234375, 0.3074531555175781, 0.39519500732421875, -0.013446807861328125, 0.04112434387207031, 1.170623779296875, 0.35108184814453125, 1.080780029296875, 0.16591835021972656, 0.4395294189453125, 0.218780517578125, -0.1386566162109375, 0.5832061767578125, 0.32714080810546875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000023.npy"}
{"epoch": 0.033773861967694566, "step": 24, "batch_size": 64, "mean": 0.21542611718177795, "std": 0.39663538336753845, "min": -0.4170989990234375, "p10": -0.2505500793457031, "median": 0.23830318450927734, "p90": 0.6646711349487305, "max": 1.3483123779296875, "pos_frac": 0.640625, "sample": [0.48915863037109375, 0.619873046875, -0.203582763671875, 0.12694549560546875, 0.6673107147216797, -0.17405319213867188, 1.3483123779296875, -0.25852203369140625, 1.0257568359375, 0.338592529296875, 0.36797142028808594, -0.07790374755859375, -0.0729522705078125, 0.24932861328125, 0.395751953125, 0.3017997741699219, -0.1775970458984375, 0.41951942443847656, 0.2708892822265625, -0.36248016357421875, 0.6585121154785156, 0.1259613037109375, -0.1250743865966797, -0.3079986572265625, 0.3638763427734375, 0.9229965209960938, 0.5899810791015625, -0.40692138671875, 0.3221435546875, 0.48065948486328125, 1.263824462890625, 0.046783447265625, 0.02036285400390625, -0.21503448486328125, 0.3585472106933594, -0.38238525390625, 0.4948577880859375, 0.6244049072265625, 0.0123443603515625, -0.21099853515625, 0.751739501953125, 0.34457969665527344, -0.11997604370117188, 0.32135772705078125, -0.11681747436523438, 0.5883255004882812, -0.2319488525390625, -0.0247344970703125, 0.34405517578125, 0.2272777557373047, 0.4715995788574219, -0.03186988830566406, 0.19911575317382812, 0.29911041259765625, 0.1523895263671875, 0.0234375, 0.40534400939941406, -0.03423309326171875, -0.029169082641601562, -0.346954345703125, 0.744873046875, -0.4170989990234375, -0.0911407470703125, 0.4270477294921875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000024.npy"}
{"epoch": 0.03524229074889868, "step": 25, "batch_size": 64, "mean": 0.31733620166778564, "std": 0.4653257131576538, "min": -1.2067718505859375, "p10": -0.2589731216430664, "median": 0.300567626953125, "p90": 0.8281471252441409, "max": 1.7274169921875, "pos_frac": 0.765625, "sample": [0.5458259582519531, 0.38370704650878906, 0.496795654296875, 0.28075408935546875, 0.24093246459960938, -0.08319664001464844, -0.28560638427734375, 1.2943115234375, 0.700408935546875, -0.00862884521484375, 0.32038116455078125, 0.85821533203125, 0.25041961669921875, 0.7579879760742188, 1.7274169921875, -0.5659160614013672, 0.24933242797851562, 0.9130859375, 0.08061599731445312, 0.7132682800292969, -0.35109710693359375, 0.07844924926757812, 0.03760337829589844, 0.34346771240234375, 0.718048095703125, 0.6373977661132812, 0.416473388671875, 0.7486190795898438, -0.48604583740234375, -0.23229026794433594, -0.02977752685546875, -0.00270843505859375, 0.38861846923828125, 0.2598876953125, 0.6109619140625, 0.2585601806640625, -0.15836524963378906, 1.0211334228515625, 0.227447509765625, 0.9266128540039062, -0.08185386657714844, 0.6328659057617188, 1.0708847045898438, 0.5537147521972656, -0.2801399230957031, 0.40283966064453125, 0.19930267333984375, 0.4108734130859375, 0.7010116577148438, 0.49921417236328125, 0.2365570068359375, -0.27040863037109375, 0.2216033935546875, -1.2067718505859375, 0.5683746337890625, 0.049083709716796875, -0.1966552734375, 0.5657539367675781, 0.4755401611328125, 0.33881378173828125, 0.2004070281982422, 0.622833251953125, 0.092529296875, 0.22003555297851562], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000025.npy"}
{"epoch": 0.03671071953010279, "step": 26, "batch_size": 64, "mean": 0.4308694005012512, "std": 0.540145218372345, "min": -0.65936279296875, "p10": -0.13279247283935547, "median": 0.3834238052368164, "p90": 1.0970275878906253, "max": 1.789581298828125, "pos_frac": 0.796875, "sample": [0.9716796875, 0.15354537963867188, 0.3894329071044922, 0.35491371154785156, 0.6494140625, 0.6677322387695312, 0.42539215087890625, 0.7227363586425781, 0.0499267578125, 0.11325454711914062, 1.784881591796875, -0.02089691162109375, 1.789581298828125, 0.02214813232421875, 0.46723175048828125, 0.9854736328125, 1.0128173828125, 0.43824005126953125, 0.3236083984375, -0.2405853271484375, 0.38364601135253906, 0.18225860595703125, 0.5571060180664062, 0.18682098388671875, -0.4562835693359375, 0.992523193359375, 0.20829010009765625, -0.65936279296875, 0.35103797912597656, 0.5509719848632812, -0.22955322265625, 0.18133544921875, 0.3119525909423828, 0.23038482666015625, -0.06036567687988281, 0.532745361328125, 1.543212890625, -0.1084136962890625, -0.0301055908203125, 1.7755813598632812, 0.5339431762695312, -0.141815185546875, 0.5350189208984375, 0.605499267578125, 0.00183868408203125, 0.6924896240234375, -0.6348609924316406, 0.7289276123046875, 0.2364940643310547, -0.1333904266357422, 0.008174896240234375, 0.502288818359375, -0.07805061340332031, 0.505462646484375, 0.48123931884765625, 0.581390380859375, 1.1419525146484375, 1.7684249877929688, 1.13311767578125, -0.13139724731445312, 0.38320159912109375, 0.3597412109375, 0.6356086730957031, 0.3560314178466797], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000026.npy"}
{"epoch": 0.0381791483113069, "step": 27, "batch_size": 64, "mean": 0.504896879196167, "std": 0.4791494309902191, "min": -0.4965534210205078, "p10": -0.05844058990478513, "median": 0.49291229248046875, "p90": 1.0404758453369143, "max": 2.3171844482421875, "pos_frac": 0.84375, "sample": [-0.10955619812011719, 0.7095832824707031, 0.6816482543945312, 0.2533721923828125, -0.14199066162109375, 0.5672836303710938, 0.716156005859375, 0.5012969970703125, 1.4628753662109375, -0.4965534210205078, 0.4755992889404297, -0.09052276611328125, 1.0039329528808594, 0.3856353759765625, 1.0561370849609375, 1.0812530517578125, 2.3171844482421875, 0.5956192016601562, 0.3724822998046875, 0.3848075866699219, 0.2753562927246094, 0.3486328125, 0.6993331909179688, -0.1469898223876953, 0.29526519775390625, 0.33535194396972656, 0.09220314025878906, -0.03184318542480469, 0.6495323181152344, 1.6354217529296875, 0.138519287109375, 0.12757110595703125, -0.009037017822265625, 1.27191162109375, 0.42368316650390625, 0.3309326171875, 0.7408218383789062, 0.8241043090820312, 0.3844795227050781, 0.11067962646484375, 0.127593994140625, 0.6194686889648438, 0.086151123046875, 0.540252685546875, 0.7388877868652344, 0.6855621337890625, -0.0698394775390625, 0.4050483703613281, 0.7645797729492188, 0.484527587890625, 0.92645263671875, 0.07559967041015625, 0.2572174072265625, 0.6191024780273438, 0.6764678955078125, 0.65216064453125, 1.4225234985351562, -2.288818359375e-05, 0.568115234375, -0.30361175537109375, 0.5287933349609375, 0.8934249877929688, 0.5312423706054688, 0.8615264892578125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000027.npy"}
{"epoch": 0.039647577092511016, "step": 28, "batch_size": 64, "mean": 0.30226144194602966, "std": 0.6013567447662354, "min": -1.793487548828125, "p10": -0.20055160522460938, "median": 0.3547401428222656, "p90": 0.9831741333007813, "max": 1.933837890625, "pos_frac": 0.6875, "sample": [0.9867172241210938, -0.0402679443359375, -0.35459136962890625, 0.42778587341308594, 0.8968830108642578, -0.0825653076171875, -0.06051063537597656, 0.2176074981689453, -0.11760330200195312, 0.06607437133789062, 0.05686187744140625, 0.700439453125, 1.6917266845703125, -0.06153106689453125, 0.5449752807617188, 0.06267547607421875, 0.4957275390625, 0.99139404296875, 0.39752960205078125, 0.593414306640625, 0.13014984130859375, -0.1221466064453125, -0.00431060791015625, 1.234771728515625, 0.33242034912109375, 0.39361572265625, 0.3804435729980469, 0.9749069213867188, 1.0714645385742188, 0.29572296142578125, 0.6409339904785156, 0.14061737060546875, 0.1352996826171875, -0.176177978515625, 0.8443870544433594, 0.3945503234863281, 0.7332057952880859, 0.39023590087890625, -0.19443130493164062, -0.3749237060546875, 0.4958534240722656, 0.6433734893798828, 0.1339855194091797, 0.6972198486328125, 1.933837890625, 0.03924560546875, 0.64422607421875, -0.09826469421386719, 0.42493438720703125, -0.07872772216796875, 0.6667919158935547, -0.9405441284179688, -0.05609130859375, 0.4321441650390625, -0.20317459106445312, 0.4818115234375, -0.11856842041015625, 0.9163665771484375, 1.4457244873046875, -0.3058815002441406, 0.3770599365234375, -1.793487548828125, -1.1559524536132812, 0.12937164306640625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000028.npy"}
{"epoch": 0.041116005873715125, "step": 29, "batch_size": 64, "mean": 0.4999760091304779, "std": 0.5782369375228882, "min": -0.6080322265625, "p10": -0.19221839904785154, "median": 0.48360252380371094, "p90": 1.2405723571777345, "max": 2.1335906982421875, "pos_frac": 0.78125, "sample": [1.0900039672851562, -0.13866043090820312, 0.765625, -0.2568016052246094, 0.062774658203125, 0.7848987579345703, -0.22956085205078125, 1.3343658447265625, 0.8445892333984375, -0.6080322265625, 0.48492431640625, -0.12355995178222656, 1.4241409301757812, 0.37677764892578125, -0.48406982421875, 0.211883544921875, 0.7736663818359375, -0.12711715698242188, -0.14442825317382812, 1.026824951171875, 0.3794708251953125, 1.0219306945800781, 1.2080001831054688, 0.384002685546875, 1.439788818359375, 0.5386581420898438, 0.031383514404296875, 0.7854690551757812, -0.208099365234375, 0.198486328125, -0.031106948852539062, 1.1511611938476562, 2.1335906982421875, 0.0355224609375, 0.1857147216796875, 0.4103546142578125, 0.5326080322265625, 0.13272476196289062, 1.2686767578125, 1.9459228515625, 0.7611846923828125, 1.2545318603515625, 0.4822807312011719, 0.5451278686523438, 0.6208858489990234, -0.15516281127929688, 0.7426223754882812, 0.0741424560546875, 0.0358734130859375, -0.41051292419433594, 0.12267112731933594, 0.22265052795410156, 0.9191131591796875, 0.7704963684082031, 0.5628929138183594, 0.47216796875, -0.0045928955078125, 0.36003875732421875, 0.64959716796875, 0.6299934387207031, 1.1165771484375, 0.7868270874023438, -0.27771759033203125, 1.1042709350585938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000029.npy"}
{"epoch": 0.042584434654919234, "step": 30, "batch_size": 64, "mean": 0.6536521911621094, "std": 0.6181467175483704, "min": -0.5793380737304688, "p10": -0.08195304870605469, "median": 0.6604385375976562, "p90": 1.489009094238282, "max": 2.338470458984375, "pos_frac": 0.84375, "sample": [1.1189384460449219, 0.2426300048828125, 0.2073516845703125, 0.9851226806640625, 1.866607666015625, -0.490570068359375, 0.05694580078125, -0.0431365966796875, -0.1340618133544922, 0.6803817749023438, 0.2639007568359375, 1.0892868041992188, 1.271209716796875, 0.8556785583496094, 0.474853515625, 0.879730224609375, 0.8862762451171875, 0.9157791137695312, 0.5445899963378906, 1.6529083251953125, 0.6723480224609375, 1.6539154052734375, 2.338470458984375, 0.8719635009765625, 0.731414794921875, 0.47783660888671875, 0.122894287109375, 0.49491119384765625, -0.10280799865722656, 0.2756061553955078, 0.3157806396484375, 0.22139358520507812, 0.6574325561523438, -0.07502365112304688, 0.1366252899169922, 0.48139190673828125, 1.8187255859375, 2.3374786376953125, 0.514404296875, 0.8658885955810547, -0.04216575622558594, 0.88275146484375, 0.21719741821289062, 0.5014877319335938, -0.5793380737304688, 0.9074630737304688, 1.576904296875, 0.6634445190429688, 0.9893875122070312, 1.144378662109375, 1.1825332641601562, 0.9787521362304688, 0.16225242614746094, -0.15886688232421875, 0.20500946044921875, 0.7445526123046875, 0.9485282897949219, -0.08492279052734375, 0.8678131103515625, 0.1760234832763672, -0.1388397216796875, 1.2839202880859375, 0.9776382446289062, 0.29276084899902344], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000030.npy"}
{"epoch": 0.04405286343612335, "step": 31, "batch_size": 64, "mean": 0.5278059244155884, "std": 0.6378049254417419, "min": -0.3424530029296875, "p10": -0.13241653442382811, "median": 0.4535350799560547, "p90": 1.312902069091797, "max": 2.79595947265625, "pos_frac": 0.78125, "sample": [0.04804229736328125, 1.3292465209960938, -0.11536407470703125, 0.514404296875, 1.8332366943359375, 0.4668426513671875, 0.0726470947265625, -0.21984100341796875, 0.36708831787109375, -0.3424530029296875, 2.79595947265625, 0.1697998046875, 0.8090591430664062, 0.3611717224121094, 0.142547607421875, -0.0464019775390625, 0.9605560302734375, 0.2052440643310547, 0.7740249633789062, 0.8057918548583984, -0.089324951171875, -0.2617225646972656, 0.5131607055664062, 0.04773712158203125, 0.12950515747070312, 0.22139930725097656, 0.9051322937011719, 0.6318817138671875, 0.7182197570800781, -0.06510543823242188, 2.3754425048828125, 0.5022125244140625, 1.6714935302734375, 0.03094482421875, 0.48072052001953125, -0.225555419921875, 0.32622528076171875, 0.957794189453125, -0.020294189453125, 0.77716064453125, 0.7632369995117188, 0.77337646484375, -0.14467239379882812, 0.39708709716796875, 0.4753570556640625, -0.1397247314453125, 0.6873779296875, 0.8592147827148438, 0.2158355712890625, -0.1695575714111328, 0.09766960144042969, 1.47882080078125, 1.20391845703125, 0.6946258544921875, -0.05684661865234375, 0.3780364990234375, 0.11928176879882812, 0.4627532958984375, -0.0213470458984375, 0.4443168640136719, 1.2747650146484375, 2.027172088623047, 0.7799758911132812, 0.6202774047851562], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000031.npy"}
{"epoch": 0.04552129221732746, "step": 32, "batch_size": 64, "mean": 0.8104609251022339, "std": 0.9149099588394165, "min": -0.4446525573730469, "p10": -0.1684200286865234, "median": 0.6273603439331055, "p90": 2.143382263183594, "max": 4.57861328125, "pos_frac": 0.828125, "sample": [0.5509414672851562, 0.2288055419921875, 0.820068359375, 1.4767608642578125, 1.0502853393554688, 1.5856170654296875, -0.1223602294921875, 0.9353179931640625, 0.5858974456787109, 1.1458282470703125, 1.2393074035644531, 0.4883689880371094, -0.006771087646484375, 0.22141265869140625, -0.202392578125, 1.1538238525390625, -0.4446525573730469, 1.1626968383789062, -0.18815994262695312, 2.7368621826171875, 0.11688995361328125, 1.3090286254882812, 2.1716766357421875, 0.5358505249023438, 0.7638320922851562, 0.47576141357421875, 0.34111785888671875, 0.036102294921875, -0.23753738403320312, 1.0588912963867188, 1.0245895385742188, 0.24110794067382812, 0.4422340393066406, 2.3834381103515625, 1.5521202087402344, 0.24616241455078125, 1.00543212890625, 0.8378372192382812, -0.07495498657226562, 0.36727142333984375, 2.6358795166015625, 0.6384716033935547, -0.31920623779296875, 0.01318359375, 4.57861328125, 1.0416336059570312, 0.6162490844726562, 2.077362060546875, 2.801361083984375, -0.3644866943359375, 0.06668663024902344, 0.36661529541015625, 2.233245849609375, 1.0928535461425781, -0.2817840576171875, 0.20126724243164062, 0.7040901184082031, 0.9840240478515625, -0.07599067687988281, 0.691070556640625, 0.1864910125732422, 1.1668624877929688, 1.244049072265625, 0.5564460754394531], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000032.npy"}
{"epoch": 0.04698972099853157, "step": 33, "batch_size": 64, "mean": 0.5591914653778076, "std": 0.6299973726272583, "min": -0.8059539794921875, "p10": -0.2490028381347656, "median": 0.5254192352294922, "p90": 1.2724822998046876, "max": 2.20330810546875, "pos_frac": 0.8125, "sample": [-0.22382736206054688, 0.3521137237548828, 0.622711181640625, 0.5469436645507812, 2.20330810546875, -0.08452606201171875, -0.35965728759765625, 0.14825439453125, 0.2892417907714844, 0.6386966705322266, 0.34104156494140625, 0.674224853515625, 1.8077392578125, 1.0077438354492188, 0.927978515625, 0.2728118896484375, 0.630767822265625, -0.5905685424804688, 0.2716217041015625, -0.8059539794921875, 0.16790771484375, 1.3467864990234375, 1.2883987426757812, 1.1845169067382812, 0.9774246215820312, 0.2318878173828125, 1.2353439331054688, -0.03212738037109375, 0.23203277587890625, 1.63690185546875, 0.054157257080078125, -0.33359527587890625, 2.1041717529296875, -0.11698722839355469, 0.16141510009765625, 0.6980819702148438, 0.4073638916015625, -0.293060302734375, 1.1995010375976562, 0.7092666625976562, 0.32877349853515625, 1.0755882263183594, 0.4740142822265625, 0.33882904052734375, 0.4272613525390625, 1.079681396484375, 0.27921295166015625, 1.6368637084960938, 1.139047622680664, -0.5903778076171875, 0.48558807373046875, -0.2597923278808594, 0.9446449279785156, -0.19951438903808594, 0.58135986328125, 0.7232208251953125, 1.2064018249511719, 0.13185882568359375, 0.7261505126953125, 0.7031135559082031, 0.5038948059082031, 0.77752685546875, 0.7638607025146484, 0.9809913635253906], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000033.npy"}
{"epoch": 0.048458149779735685, "step": 34, "batch_size": 64, "mean": 0.8371996879577637, "std": 0.9117099642753601, "min": -0.49461936950683594, "p10": -0.14145622253417967, "median": 0.5702590942382812, "p90": 2.2900146484375, "max": 2.956878662109375, "pos_frac": 0.84375, "sample": [2.54534912109375, -0.19385910034179688, 1.3483123779296875, 0.10477066040039062, 2.9362640380859375, 0.5037651062011719, 1.393707275390625, 1.2188873291015625, -0.10935211181640625, 0.12265777587890625, 0.170806884765625, -0.10466575622558594, 0.595550537109375, -0.24260711669921875, 2.3004150390625, 2.4824676513671875, 0.3662128448486328, -0.15301132202148438, 2.4519195556640625, 0.1464080810546875, 1.1649398803710938, -0.3670692443847656, 1.3865852355957031, 1.7410240173339844, 0.803955078125, 1.6150360107421875, 2.2302017211914062, -0.18486404418945312, 0.1221160888671875, 0.9453964233398438, 0.29901123046875, 0.715240478515625, 1.0806121826171875, 1.2359447479248047, 2.2657470703125, 0.9637832641601562, 0.6056919097900391, 2.956878662109375, 0.47049713134765625, 0.20177078247070312, 0.16502761840820312, 0.3957843780517578, 1.0206260681152344, 1.0928268432617188, 0.8583602905273438, 0.96240234375, 0.1305084228515625, -0.11449432373046875, 1.377044677734375, 0.15060806274414062, 0.4169197082519531, -0.380340576171875, 0.8203811645507812, 0.412567138671875, 0.5449676513671875, -0.49461936950683594, 2.1190414428710938, 0.24597549438476562, 0.21524810791015625, 0.2562103271484375, 2.0002288818359375, 2.904052734375, 0.029558181762695312, 0.3153953552246094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000034.npy"}
{"epoch": 0.049926578560939794, "step": 35, "batch_size": 64, "mean": 1.1504449844360352, "std": 1.1679655313491821, "min": -0.30419158935546875, "p10": 0.059680557250976624, "median": 0.8149681091308594, "p90": 2.2430915832519536, "max": 7.2916412353515625, "pos_frac": 0.921875, "sample": [1.5773468017578125, 0.6674423217773438, 3.3729248046875, 1.4806404113769531, 0.13006591796875, 0.7673110961914062, 0.544830322265625, 0.0109405517578125, -0.20986557006835938, 1.462188720703125, 1.613677978515625, 0.12137222290039062, 0.8863143920898438, 0.811614990234375, 0.582550048828125, 0.14618682861328125, 1.9706039428710938, 2.89306640625, 0.3936614990234375, 0.27474212646484375, 0.580169677734375, 1.8367691040039062, 2.5351715087890625, 0.03324127197265625, 0.518646240234375, 0.8173294067382812, 0.7893619537353516, 1.42694091796875, 0.5371017456054688, 0.48419952392578125, 7.2916412353515625, 1.171875, 0.740509033203125, 1.6380538940429688, 0.6296977996826172, 3.28533935546875, 0.24159622192382812, 1.5403518676757812, 2.2702484130859375, 2.0918121337890625, 0.164306640625, 2.1702651977539062, 0.667236328125, 1.05413818359375, 1.0571327209472656, -0.18233489990234375, 1.9612884521484375, 2.02325439453125, 1.0499420166015625, 0.16693115234375, 0.8813419342041016, -0.006072998046875, 0.5820198059082031, 2.1790618896484375, 0.6634674072265625, 1.6082839965820312, 0.9022979736328125, -0.06116294860839844, 1.3050613403320312, -0.30419158935546875, 0.8126068115234375, 2.621856689453125, 2.1797256469726562, 0.174346923828125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000035.npy"}
{"epoch": 0.0513950073421439, "step": 36, "batch_size": 64, "mean": 1.2142744064331055, "std": 1.3078432083129883, "min": -1.7821273803710938, "p10": -0.0653816223144531, "median": 0.8655338287353516, "p90": 3.209806823730469, "max": 4.2769775390625, "pos_frac": 0.859375, "sample": [1.2704849243164062, -0.20292091369628906, 0.7975368499755859, 1.001983642578125, 1.3590164184570312, 0.6060829162597656, 1.375213623046875, -0.28436279296875, -0.14644813537597656, 2.1266632080078125, 2.0906753540039062, 0.20123291015625, -0.24487686157226562, 0.5390739440917969, 0.1728839874267578, 1.1143035888671875, -0.020099639892578125, 1.9020919799804688, 1.544708251953125, 0.0094451904296875, -0.0369110107421875, -0.07758331298828125, 0.31890106201171875, 0.07578468322753906, 4.106513977050781, 3.8984718322753906, 1.6428947448730469, 0.0695953369140625, 3.5704498291015625, -0.33820343017578125, 0.5583057403564453, 1.86444091796875, 0.2568359375, 2.6768264770507812, 0.386566162109375, 0.04229164123535156, 4.092193603515625, 0.42431640625, 1.9055328369140625, 0.6483268737792969, 2.3743438720703125, 1.0696334838867188, 3.2357711791992188, 0.9810161590576172, 3.07440185546875, 0.92962646484375, 0.22908401489257812, 4.2769775390625, 0.7677440643310547, -1.7821273803710938, 2.456493377685547, 0.7375030517578125, 2.3437728881835938, 0.3382720947265625, 3.1492233276367188, 1.163665771484375, 0.2241973876953125, 0.3601551055908203, 2.05291748046875, 0.5647201538085938, 1.8456802368164062, 3.9124908447265625, 1.3083229064941406, 0.8014411926269531], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000036.npy"}
{"epoch": 0.05286343612334802, "step": 37, "batch_size": 64, "mean": 1.002803087234497, "std": 1.3508623838424683, "min": -1.88897705078125, "p10": -0.1365543365478515, "median": 0.5529546737670898, "p90": 2.7565582275390637, "max": 6.1043701171875, "pos_frac": 0.859375, "sample": [2.0650863647460938, 0.03960609436035156, 0.9141998291015625, 0.97723388671875, -0.15493011474609375, 4.57763671875e-05, 0.30365753173828125, 0.8201370239257812, 1.7774124145507812, 0.033107757568359375, 0.7181320190429688, 3.0110931396484375, 0.40660858154296875, 0.02960205078125, 0.242218017578125, 0.02588653564453125, 0.7228775024414062, 0.1272449493408203, 1.972585678100586, 6.1043701171875, 0.8519058227539062, 4.62762451171875, 0.9907150268554688, 0.8384742736816406, 0.33126068115234375, 1.5161056518554688, 0.3656463623046875, 0.16481781005859375, 1.3177261352539062, 1.3244590759277344, 0.4539375305175781, 0.6345443725585938, 0.4762592315673828, -0.8927688598632812, 0.33109283447265625, -0.002513885498046875, 0.3585205078125, 1.262054443359375, 2.2098464965820312, 3.4092254638671875, -0.25794410705566406, 2.8863754272460938, 0.53289794921875, 1.77197265625, -0.2758369445800781, 0.5730113983154297, 0.014379501342773438, -0.282989501953125, 0.4692840576171875, 1.9825210571289062, 3.0199127197265625, -0.09367752075195312, 2.0963668823242188, 2.4536514282226562, 0.22751998901367188, 2.4078521728515625, 1.33709716796875, -1.88897705078125, 0.0704803466796875, -0.22788238525390625, 2.207775115966797, 4.1947174072265625, 0.02623748779296875, 0.22954559326171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000037.npy"}
{"epoch": 0.05433186490455213, "step": 38, "batch_size": 64, "mean": 1.3876228332519531, "std": 1.6507039070129395, "min": -1.0857009887695312, "p10": 0.052504730224609414, "median": 0.9444561004638672, "p90": 2.81158676147461, "max": 8.544769287109375, "pos_frac": 0.921875, "sample": [0.31188201904296875, 8.544769287109375, 4.369781494140625, 1.9714851379394531, 0.7492103576660156, 0.22418785095214844, 1.359222412109375, 0.1316070556640625, 0.8114471435546875, 1.337249755859375, 1.8223991394042969, 0.8173370361328125, -0.0848388671875, 2.6341781616210938, 1.2757949829101562, -0.027812957763671875, 0.7707748413085938, 0.09455108642578125, 0.20917320251464844, 2.6266021728515625, 0.9663200378417969, 2.1779632568359375, 2.145263671875, 0.1536998748779297, 3.232635498046875, 0.1873035430908203, 2.3003005981445312, 2.49652099609375, 1.6053962707519531, 0.6592788696289062, 0.41335105895996094, 2.5658416748046875, -1.0857009887695312, 1.1927413940429688, 0.6907634735107422, 0.1706085205078125, 1.1247024536132812, 5.8410491943359375, 0.8716316223144531, 0.25505828857421875, 2.8876190185546875, 5.138519287109375, 1.5813674926757812, 0.8176078796386719, 0.7527523040771484, 0.0240936279296875, 0.16131591796875, 0.1442413330078125, 1.0738906860351562, 1.6947021484375, -0.829437255859375, 0.6335411071777344, 0.03448486328125, 1.56060791015625, 1.5084552764892578, 1.9561996459960938, 1.3065948486328125, 5.72967529296875, 0.9225921630859375, 0.16378021240234375, 1.8425216674804688, -0.4916191101074219, 0.34307861328125, 1.9375495910644531], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000038.npy"}
{"epoch": 0.055800293685756244, "step": 39, "batch_size": 64, "mean": 1.624021291732788, "std": 1.4317415952682495, "min": -2.7746734619140625, "p10": 0.28542556762695315, "median": 1.3284835815429688, "p90": 3.6499526977539065, "max": 5.3980712890625, "pos_frac": 0.921875, "sample": [2.3950424194335938, 2.2357635498046875, 3.5483856201171875, 2.1169357299804688, 1.0769577026367188, -0.7726974487304688, 2.7997283935546875, 2.784717559814453, 1.1154193878173828, 1.9776477813720703, 0.42412567138671875, 3.211273193359375, 1.9175262451171875, 1.3986396789550781, 3.7495040893554688, 1.4274711608886719, -0.06902313232421875, 0.81585693359375, 0.2394561767578125, 1.2583274841308594, 0.5269622802734375, 2.619049072265625, 2.4416961669921875, 0.7931728363037109, 0.9603614807128906, 2.7545394897460938, -0.06439208984375, 2.049999237060547, 4.627716064453125, 0.4390296936035156, 0.64813232421875, 0.36491966247558594, 3.4205780029296875, 2.177703857421875, -2.7746734619140625, 2.1403465270996094, 0.4458904266357422, 1.5281105041503906, 3.787942886352539, 5.3980712890625, 1.1016674041748047, 1.157806396484375, 3.921222686767578, 0.71807861328125, 1.1292686462402344, 0.4741973876953125, 0.2994232177734375, 0.4143486022949219, 1.6555404663085938, 1.0440521240234375, 0.36653900146484375, 3.6934814453125, 2.8914260864257812, 2.94451904296875, 0.27942657470703125, 4.04327392578125, 2.9500274658203125, 0.7205390930175781, 2.1811485290527344, 0.4242744445800781, 0.8304233551025391, -0.10865974426269531, 0.5554962158203125, 2.313627243041992], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000039.npy"}
{"epoch": 0.05726872246696035, "step": 40, "batch_size": 64, "mean": 1.5779554843902588, "std": 1.7397431135177612, "min": -0.8597412109375, "p10": -0.27761611938476555, "median": 1.2738933563232422, "p90": 3.8157829284667977, "max": 7.669830322265625, "pos_frac": 0.796875, "sample": [-0.20497894287109375, -0.0091705322265625, 1.2615966796875, 1.1434974670410156, 1.4681396484375, 0.6407470703125, 0.75201416015625, 0.3161163330078125, 0.29976844787597656, 2.218503952026367, 6.5370941162109375, 3.4131011962890625, 0.6885356903076172, 3.9077682495117188, 1.2523880004882812, -0.8425254821777344, 2.087879180908203, 1.9978790283203125, 0.2266826629638672, -0.3824005126953125, 3.0673370361328125, -0.44434165954589844, 1.2241668701171875, -0.8597412109375, 2.4816665649414062, 1.736724853515625, 0.4916343688964844, 1.2331867218017578, 1.2861900329589844, -0.40673828125, 4.397377014160156, -0.1468334197998047, 3.6011505126953125, -0.3789215087890625, 0.5229988098144531, 5.046516418457031, 0.008396148681640625, 3.4162864685058594, 0.74658203125, 1.414358139038086, 4.827239990234375, 1.2540836334228516, 1.4571952819824219, 7.669830322265625, -0.11339950561523438, 1.8907699584960938, -0.308746337890625, 1.9105453491210938, 3.2201156616210938, 1.7362556457519531, 1.813507080078125, 2.2549819946289062, 3.1475448608398438, 1.6597480773925781, 1.2567520141601562, 1.5189208984375, 4.685211181640625, -0.20459747314453125, 1.7259445190429688, 0.23889923095703125, -0.11898040771484375, 1.91656494140625, 1.8656387329101562, 0.4744987487792969], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000040.npy"}
{"epoch": 0.05873715124816446, "step": 41, "batch_size": 64, "mean": 1.9229004383087158, "std": 2.0603528022766113, "min": -3.3866729736328125, "p10": 0.1093921661376955, "median": 1.4770278930664062, "p90": 4.653224182128907, "max": 8.865776062011719, "pos_frac": 0.90625, "sample": [1.0543556213378906, 0.87396240234375, 3.503753662109375, 1.5547714233398438, 0.7775344848632812, 2.3517837524414062, 3.4772491455078125, 2.8938674926757812, -0.2342376708984375, 3.013641357421875, 6.0533599853515625, 0.37036895751953125, 2.9459228515625, 8.865776062011719, 0.29669952392578125, 0.5700836181640625, 1.2990875244140625, 1.0698966979980469, 0.029117584228515625, -2.303985595703125, 1.4080581665039062, 3.3346099853515625, 4.7222137451171875, 1.0909805297851562, 6.421783447265625, 1.6803016662597656, 4.49224853515625, 4.77947998046875, 1.115274429321289, 0.5677680969238281, 2.1392059326171875, 0.325927734375, 7.012664794921875, 0.7000827789306641, 1.2709999084472656, 3.9896392822265625, 1.7053489685058594, 2.707977294921875, -0.569854736328125, 2.1851959228515625, 4.051582336425781, 1.5722484588623047, -3.3866729736328125, 3.5701904296875, 1.1008377075195312, 0.886199951171875, 1.9883956909179688, 2.1129112243652344, 2.288328170776367, 0.3230266571044922, 0.7173004150390625, 1.0555953979492188, 1.546600341796875, 1.2141036987304688, 0.7223854064941406, 1.5459976196289062, 0.49423980712890625, 3.5487747192382812, 1.19512939453125, 0.40964698791503906, -0.1854095458984375, -0.4606819152832031, 2.3848342895507812, 4.8271484375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000041.npy"}
{"epoch": 0.06020558002936858, "step": 42, "batch_size": 64, "mean": 2.6237030029296875, "std": 2.418861150741577, "min": -1.6228561401367188, "p10": 0.126861572265625, "median": 2.0554981231689453, "p90": 6.050041961669923, "max": 11.061981201171875, "pos_frac": 0.921875, "sample": [1.3543319702148438, 0.1268157958984375, 1.9598407745361328, 10.255462646484375, 11.061981201171875, 2.5243682861328125, 0.1269683837890625, 6.1186676025390625, 2.115894317626953, 3.1248703002929688, 2.9061660766601562, 6.079780578613281, 3.0254974365234375, 7.067718505859375, 0.7193107604980469, 2.1375274658203125, 2.9178924560546875, -0.04923248291015625, 3.1516494750976562, 3.5929183959960938, 1.5741539001464844, 1.3550548553466797, 3.9994277954101562, 0.6582374572753906, 1.1545639038085938, 3.0059165954589844, -0.259674072265625, 2.3561553955078125, 1.9575881958007812, 0.10410308837890625, 2.019855499267578, 6.4029388427734375, 0.3504066467285156, 5.98065185546875, 2.7554092407226562, 1.8640060424804688, 0.8923397064208984, 1.6789169311523438, 2.0911407470703125, -0.49950408935546875, 3.6254730224609375, 0.5700225830078125, 1.8493003845214844, 5.7718963623046875, 1.6748695373535156, 3.9892654418945312, 1.0592041015625, 0.2844085693359375, 1.1731433868408203, 1.347808837890625, 0.7155113220214844, 6.4544677734375, -1.6228561401367188, 0.8297443389892578, 2.5782546997070312, 3.4154281616210938, -0.15273284912109375, 1.9812164306640625, 5.2830810546875, 2.107330322265625, 4.3374786376953125, 5.200843811035156, 1.0429267883300781, 4.640777587890625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000042.npy"}
{"epoch": 0.06167400881057269, "step": 43, "batch_size": 64, "mean": 2.405916929244995, "std": 1.9180521965026855, "min": -1.1945877075195312, "p10": 0.16683998107910158, "median": 2.3344039916992188, "p90": 4.667335510253907, "max": 9.842727661132812, "pos_frac": 0.953125, "sample": [0.19163894653320312, 0.3238182067871094, 0.15621185302734375, 3.0367279052734375, 3.89691162109375, 1.0716552734375, 4.193817138671875, 1.239349365234375, 2.3957138061523438, 6.3384246826171875, 0.0673980712890625, 2.4126338958740234, 0.9527816772460938, 1.837646484375, 3.4085731506347656, 1.4872894287109375, 3.50689697265625, 5.33892822265625, 1.599639892578125, 0.1335468292236328, 6.471405029296875, 2.7543869018554688, 4.7398223876953125, 1.0062446594238281, 3.1927108764648438, 2.3371124267578125, 0.2713584899902344, 2.331695556640625, -0.04780769348144531, 4.498199462890625, 1.8490829467773438, 2.4954261779785156, 3.821136474609375, 2.944427490234375, 3.1513519287109375, 3.794330596923828, 0.7084732055664062, 2.376251220703125, 1.5268573760986328, 1.07720947265625, 1.6629257202148438, 2.4084701538085938, 6.4244537353515625, 3.59765625, 0.6251983642578125, -1.1945877075195312, 1.1985950469970703, 3.0181427001953125, 2.120208740234375, 3.815643310546875, 2.266490936279297, 3.7764663696289062, 1.8654022216796875, 0.47469139099121094, 1.81549072265625, 9.842727661132812, 2.507965087890625, 1.6194229125976562, -0.03399848937988281, 0.7051734924316406, 2.9730453491210938, 4.818878173828125, 2.7653656005859375, 0.015581130981445312], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000043.npy"}
{"epoch": 0.0631424375917768, "step": 44, "batch_size": 64, "mean": 2.789440155029297, "std": 2.217756748199463, "min": -1.1635894775390625, "p10": 0.4771303176879883, "median": 2.4576492309570312, "p90": 5.902138519287111, "max": 9.56182861328125, "pos_frac": 0.96875, "sample": [4.247283935546875, 2.5247268676757812, 1.8234195709228516, 0.5173301696777344, 6.9090576171875, 2.7783355712890625, 2.9699478149414062, 1.0966415405273438, 2.134113311767578, 5.527435302734375, 6.535308837890625, 0.3207588195800781, 4.383087158203125, 5.329681396484375, 5.455902099609375, 3.8463668823242188, 0.6472511291503906, 2.041046142578125, 1.9776992797851562, 6.06121826171875, 0.4599018096923828, 2.6744384765625, 2.4533824920654297, 8.228424072265625, 3.06298828125, 2.81854248046875, 0.5799484252929688, 1.8523712158203125, 5.530952453613281, 1.8918704986572266, 3.1443519592285156, 0.04879951477050781, 1.7602787017822266, 0.9300918579101562, 1.3696117401123047, 3.164154052734375, 0.5837860107421875, 1.4825935363769531, 4.245849609375, 2.5235443115234375, 0.8649444580078125, 2.235321044921875, 3.7813873291015625, 3.7265472412109375, 1.3959388732910156, 3.1631011962890625, 3.325653076171875, 2.4534454345703125, 0.6812572479248047, 0.2910652160644531, 6.065277099609375, 4.6079864501953125, 2.46185302734375, 3.60205078125, 1.2253456115722656, 9.56182861328125, 0.6131248474121094, -1.1635894775390625, 2.9570465087890625, 1.9388847351074219, -0.7976646423339844, 0.045257568359375, 1.545867919921875, 8.009742736816406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000044.npy"}
{"epoch": 0.06461086637298091, "step": 45, "batch_size": 64, "mean": 2.2871103286743164, "std": 2.663897752761841, "min": -1.5081138610839844, "p10": -0.17716217041015617, "median": 1.6286468505859375, "p90": 6.0298095703125005, "max": 13.398681640625, "pos_frac": 0.859375, "sample": [3.5306243896484375, 1.1875762939453125, 0.2482757568359375, 2.222370147705078, 0.13756942749023438, 1.7202892303466797, 3.4458236694335938, 6.50689697265625, 5.113372802734375, 3.5869369506835938, 0.5608673095703125, 6.977119445800781, 3.770465850830078, 0.5482597351074219, 0.11708259582519531, 1.6718635559082031, -0.21039581298828125, 2.3400344848632812, 0.045440673828125, 5.559787750244141, -0.07843780517578125, 0.3016624450683594, 3.3573379516601562, 0.9402580261230469, 0.7402763366699219, 5.145221710205078, 7.052436828613281, 4.337860107421875, 0.25493812561035156, 0.5376911163330078, 5.2854461669921875, 2.5675125122070312, 6.0689849853515625, 0.0172882080078125, -0.09961700439453125, 1.649322509765625, 5.9384002685546875, 3.897552490234375, -0.2995414733886719, 1.60797119140625, 1.5351829528808594, 0.6136322021484375, -1.5081138610839844, -1.1857757568359375, 1.19366455078125, 13.398681640625, 0.845672607421875, 7.101829528808594, 2.9421310424804688, -0.7696304321289062, 2.235443115234375, 1.41094970703125, 1.7638626098632812, 0.4517383575439453, 1.0258598327636719, 6.4640960693359375, 1.1543121337890625, 1.83074951171875, 2.1107215881347656, 2.0028839111328125, -0.31266212463378906, 4.412258148193359, 0.6457996368408203, -1.2910614013671875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000045.npy"}
{"epoch": 0.06607929515418502, "step": 46, "batch_size": 64, "mean": 2.9119834899902344, "std": 3.1643407344818115, "min": -2.5914306640625, "p10": -0.16278076171874997, "median": 2.292971611022949, "p90": 6.710679626464844, "max": 14.501190185546875, "pos_frac": 0.828125, "sample": [0.3971595764160156, 0.39073944091796875, 5.6704864501953125, 7.1026611328125, 4.1945648193359375, -2.5156631469726562, 2.0252304077148438, 1.4185943603515625, 6.5512847900390625, 3.4636783599853516, 6.77899169921875, 1.4636497497558594, 0.26633453369140625, 2.327098846435547, 4.8927001953125, -0.09119033813476562, -0.5305862426757812, 1.0321693420410156, 4.1393585205078125, 12.260772705078125, -0.49677467346191406, 0.02557373046875, -0.4260063171386719, 4.062744140625, 0.26482391357421875, 4.0708160400390625, 4.2060394287109375, 1.3720645904541016, 4.394340515136719, 2.2588443756103516, 4.252952575683594, 0.6723365783691406, -2.5914306640625, 1.0887298583984375, 3.3945159912109375, 6.241916656494141, 5.333244323730469, -0.1376800537109375, 7.996147155761719, -0.1735382080078125, 0.27344322204589844, 3.6848373413085938, 5.7450408935546875, -0.0051422119140625, 14.501190185546875, 0.18647003173828125, 6.2798309326171875, 5.8690948486328125, 0.9604949951171875, 2.4742889404296875, 1.9920463562011719, 7.81219482421875, 3.0369319915771484, -0.06441497802734375, 2.327167510986328, 6.886985778808594, 1.4187698364257812, 2.5943775177001953, 3.4933242797851562, 5.351341247558594, 1.8378143310546875, 1.422922134399414, 1.536325454711914, -0.2960853576660156], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000046.npy"}
{"epoch": 0.06754772393538913, "step": 47, "batch_size": 64, "mean": 3.060344934463501, "std": 2.8008270263671875, "min": -2.5197525024414062, "p10": 0.11969203948974613, "median": 2.307880401611328, "p90": 6.858503723144532, "max": 11.18701171875, "pos_frac": 0.90625, "sample": [1.960693359375, 3.3262481689453125, 0.5270462036132812, 1.5976543426513672, 5.105068206787109, 1.7127265930175781, 0.7673568725585938, 2.251007080078125, 7.4882354736328125, 3.3297977447509766, -0.15422821044921875, 6.629524230957031, 6.480232238769531, 1.2445411682128906, 3.0599136352539062, 2.7222442626953125, 0.16485595703125, 1.9193878173828125, 2.3647537231445312, 5.819526672363281, 7.278717041015625, 1.74395751953125, 0.7684421539306641, 3.228130340576172, 3.0604782104492188, 1.5680599212646484, -0.9763641357421875, 6.8740081787109375, 7.433197021484375, 2.903888702392578, -0.3267631530761719, 3.1876964569091797, 5.979572296142578, 4.381866455078125, -2.5197525024414062, 0.10033607482910156, 0.38080596923828125, 1.2371807098388672, 1.65093994140625, 11.18701171875, 6.538330078125, 2.1343536376953125, 0.9791431427001953, 3.267974853515625, 9.332687377929688, 6.82232666015625, 0.6497039794921875, 3.6551456451416016, 9.95733642578125, -1.7289505004882812, 1.8259658813476562, 2.108062744140625, -0.9446563720703125, 1.709585189819336, 2.99310302734375, 0.9958038330078125, 5.5706024169921875, 1.8319625854492188, 4.710601806640625, 2.120473861694336, 4.5002593994140625, 4.1732940673828125, 2.2463340759277344, 2.9546375274658203], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000047.npy"}
{"epoch": 0.06901615271659324, "step": 48, "batch_size": 64, "mean": 3.1601364612579346, "std": 3.2292840480804443, "min": -7.8262786865234375, "p10": -0.008636665344238181, "median": 2.371309280395508, "p90": 7.888081359863283, "max": 12.470184326171875, "pos_frac": 0.890625, "sample": [4.008544921875, 0.9931983947753906, 8.085922241210938, 8.772972106933594, 7.04705810546875, 2.9623947143554688, -0.757537841796875, 2.3315811157226562, 5.523307800292969, 4.6721649169921875, 1.81329345703125, 6.5029144287109375, 2.340057373046875, -0.083709716796875, 0.0872039794921875, 1.1278152465820312, 7.42645263671875, -0.04971122741699219, 1.6405868530273438, -0.13170242309570312, 1.6308135986328125, 9.235130310058594, -0.3378582000732422, 1.8754501342773438, 6.2332763671875, 1.726806640625, 4.370765686035156, 3.425596237182617, 0.2744789123535156, 3.8145370483398438, 1.9729804992675781, 2.407726287841797, 3.71124267578125, 2.2638072967529297, 9.97906494140625, 3.8642444610595703, 4.8915252685546875, 9.353935241699219, 3.6309051513671875, 0.9370880126953125, 0.83685302734375, 1.5305233001708984, 1.4886856079101562, 3.046344757080078, 0.912872314453125, 4.064826965332031, 3.0646915435791016, 2.1189231872558594, -2.084959030151367, 8.364036560058594, 3.091989517211914, 3.081838607788086, 2.1609954833984375, 12.470184326171875, 4.944023132324219, -7.8262786865234375, 6.414318084716797, 1.5514183044433594, 1.077463150024414, 0.57135009765625, 5.3615875244140625, 1.97137451171875, 2.4025611877441406, 2.0587844848632812], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000048.npy"}
{"epoch": 0.07048458149779736, "step": 49, "batch_size": 64, "mean": 3.8101086616516113, "std": 4.062777519226074, "min": -4.034782409667969, "p10": 0.009857368469238301, "median": 3.317779541015625, "p90": 8.62259979248047, "max": 16.87786865234375, "pos_frac": 0.90625, "sample": [3.4553909301757812, 0.4583930969238281, 0.2804126739501953, 3.7614517211914062, 2.6819820404052734, 6.329341888427734, 0.633392333984375, 2.768756866455078, -1.3594818115234375, 0.6339263916015625, 0.001819610595703125, 2.4627037048339844, 3.2299346923828125, 9.599235534667969, 8.154830932617188, 2.6596832275390625, 4.213024139404297, 1.8989143371582031, 3.4056243896484375, 1.6718673706054688, 2.300567626953125, -3.2815704345703125, 3.66595458984375, 0.11030960083007812, 14.91314697265625, 10.385787963867188, 0.8646697998046875, 5.885711669921875, 3.882537841796875, 0.36527252197265625, 8.403305053710938, 16.87786865234375, 4.459770202636719, 1.40997314453125, -1.3172760009765625, -4.034782409667969, 3.8847999572753906, 3.763124465942383, 3.9807090759277344, 4.935573577880859, 3.076213836669922, 0.028612136840820312, -0.57720947265625, 4.498809814453125, 5.116233825683594, 15.178131103515625, 3.1364784240722656, 2.8535537719726562, 8.716583251953125, 6.888336181640625, 6.903106689453125, 3.2266502380371094, 4.880435943603516, 5.0567626953125, 13.198074340820312, 0.7402915954589844, -1.029937744140625, 3.6076889038085938, 1.9302692413330078, 4.254150390625, 5.3489532470703125, 2.0074920654296875, 5.3362579345703125, 1.0743560791015625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000049.npy"}
{"epoch": 0.07195301027900147, "step": 50, "batch_size": 64, "mean": 4.095892906188965, "std": 4.009570121765137, "min": -3.220518112182617, "p10": -0.4634956359863279, "median": 3.177978515625, "p90": 9.835353469848634, "max": 14.655105590820312, "pos_frac": 0.875, "sample": [6.695404052734375, 0.726318359375, -2.15875244140625, 10.29339599609375, 1.1957817077636719, 2.0250320434570312, 7.6593780517578125, 2.0738983154296875, 4.166629791259766, 1.950333595275879, -0.5574684143066406, 7.646629333496094, 3.2349853515625, 14.655105590820312, 0.508636474609375, 2.861236572265625, 7.691856384277344, 3.479635238647461, 10.75946044921875, 5.59674072265625, 0.9261150360107422, 5.237579345703125, 10.005706787109375, 2.804473876953125, 2.0336761474609375, 5.382680892944336, 0.043163299560546875, -0.24619293212890625, 3.0640716552734375, 9.460372924804688, 7.112884521484375, 5.52313232421875, 9.99605941772461, 14.522247314453125, -3.220518112182617, 0.6122226715087891, 0.5020523071289062, 1.7460556030273438, 0.345672607421875, 3.943359375, 2.3019866943359375, 1.3831424713134766, 7.4548187255859375, 4.3457489013671875, 7.564216613769531, 7.0042724609375, -1.5491485595703125, 2.1392059326171875, 5.736030578613281, -0.5566253662109375, 5.9418792724609375, 3.1209716796875, 4.619964599609375, 6.536445617675781, 2.7991561889648438, 5.023284912109375, 8.279014587402344, 1.5855178833007812, 12.537384033203125, -0.8258056640625, 6.96484375, 1.6804885864257812, -2.3624496459960938, 0.11372947692871094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000050.npy"}
{"epoch": 0.07342143906020558, "step": 51, "batch_size": 64, "mean": 4.060206413269043, "std": 5.273400783538818, "min": -6.6989593505859375, "p10": -0.8058500289916991, "median": 3.0884017944335938, "p90": 10.344725799560548, "max": 20.197174072265625, "pos_frac": 0.796875, "sample": [10.227676391601562, 4.159233093261719, 0.8600807189941406, -1.0146560668945312, 0.7215042114257812, -0.693359375, 9.819381713867188, -0.9255142211914062, 2.0161514282226562, -0.2212066650390625, 12.05670166015625, 7.959774017333984, 2.0098648071289062, 4.218284606933594, 1.9478912353515625, 4.58551025390625, 4.310871124267578, -0.8473033905029297, 9.565391540527344, 6.459934234619141, 4.606620788574219, -0.9302978515625, -0.3289794921875, 3.8473358154296875, 1.2746658325195312, 8.15633773803711, 0.5698223114013672, 5.8748931884765625, 0.8051490783691406, 5.6356201171875, 20.197174072265625, 3.288370132446289, 0.4280967712402344, 5.9186553955078125, 3.1152687072753906, 15.070541381835938, 1.1489982604980469, 16.234390258789062, 5.3534393310546875, 1.9549503326416016, 3.2986602783203125, 1.3782463073730469, 3.999774932861328, 10.394889831542969, 1.3906631469726562, 0.4544525146484375, 6.6032867431640625, -4.17901611328125, -0.6797714233398438, 19.495162963867188, 14.894058227539062, 8.024589538574219, 7.843311309814453, -0.7091255187988281, 1.327840805053711, 5.955741882324219, 0.06601333618164062, 0.3157997131347656, 2.5179901123046875, -6.6989593505859375, 3.3235626220703125, -1.58917236328125, 3.061534881591797, -0.07358551025390625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000051.npy"}
{"epoch": 0.07488986784140969, "step": 52, "batch_size": 64, "mean": 6.09666633605957, "std": 5.132640361785889, "min": -2.6499786376953125, "p10": 0.7115547180175783, "median": 5.045331954956055, "p90": 13.102807617187501, "max": 22.241790771484375, "pos_frac": 0.921875, "sample": [1.9003849029541016, 12.374954223632812, 4.414264678955078, 12.583831787109375, 7.490234375, 10.953536987304688, 3.036029815673828, 7.539306640625, 8.147674560546875, 14.406753540039062, -0.10881805419921875, 13.325225830078125, 5.132072448730469, 5.5634002685546875, 9.361234664916992, 8.312875747680664, 2.8367385864257812, 5.273967742919922, 18.271339416503906, 9.8907470703125, 3.206615447998047, 0.618896484375, 11.302650451660156, 3.2607898712158203, 0.1594409942626953, 8.139629364013672, 2.7479324340820312, 1.895547866821289, 9.438497543334961, 22.241790771484375, 10.276885986328125, 1.6356143951416016, 10.363975524902344, 4.925128936767578, 4.039710998535156, 13.427978515625, 4.2861480712890625, -1.1961822509765625, 5.402275085449219, -2.6499786376953125, -2.0530548095703125, 2.42620849609375, 6.225942611694336, 6.115760803222656, 1.0473594665527344, 9.734779357910156, 4.613767623901367, 13.837554931640625, 7.736438751220703, 1.691061019897461, 1.0940208435058594, 3.388671875, 3.8365097045898438, 4.958591461181641, 3.037841796875, 9.890960693359375, 6.558769226074219, 0.944183349609375, 17.994102478027344, -1.301483154296875, 6.241249084472656, 4.446266174316406, 2.564281463623047, 0.9277572631835938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000052.npy"}
{"epoch": 0.0763582966226138, "step": 53, "batch_size": 64, "mean": 6.241450309753418, "std": 6.861721992492676, "min": -3.814727783203125, "p10": 0.5249324798583985, "median": 4.288445472717285, "p90": 15.858927917480472, "max": 30.436798095703125, "pos_frac": 0.921875, "sample": [-1.9339542388916016, 4.061668395996094, 4.044441223144531, 1.939483642578125, 1.0831184387207031, 4.244062423706055, 13.622871398925781, 0.6609916687011719, 7.72198486328125, -2.1970367431640625, 2.927001953125, 29.93414306640625, -3.1084442138671875, 0.1373138427734375, 6.707038879394531, 8.160530090332031, 1.6587066650390625, 11.443252563476562, 3.0882034301757812, 12.686172485351562, 4.671382904052734, 5.6699371337890625, 13.984176635742188, 7.316871643066406, 16.885528564453125, 1.3294525146484375, 5.045433044433594, 0.70599365234375, 5.78533935546875, 1.9035091400146484, 18.757598876953125, 8.550891876220703, 7.583339691162109, 8.392875671386719, 4.070278167724609, 0.46662139892578125, 1.2346115112304688, 4.332828521728516, 1.0230178833007812, 4.673978805541992, 3.749481201171875, 7.958253860473633, 1.5055961608886719, 15.110702514648438, -3.814727783203125, 17.638351440429688, 19.07196044921875, 3.0138626098632812, 5.08038330078125, 3.4071102142333984, 2.2216739654541016, 4.434307098388672, 2.9775848388671875, 2.140594482421875, 3.7802963256835938, 4.538665771484375, 3.164175033569336, 30.436798095703125, 2.9484825134277344, 15.002777099609375, 5.298725128173828, 4.838081359863281, 16.179595947265625, -0.4951171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000053.npy"}
{"epoch": 0.07782672540381791, "step": 54, "batch_size": 64, "mean": 5.395318031311035, "std": 5.037128448486328, "min": -0.9480018615722656, "p10": 0.15600242614746093, "median": 4.346921920776367, "p90": 12.15518569946289, "max": 20.933746337890625, "pos_frac": 0.921875, "sample": [8.230064392089844, 5.3034210205078125, 9.075550079345703, 1.3350372314453125, 0.806243896484375, 3.0143890380859375, 0.1595458984375, 2.7049598693847656, 6.102439880371094, 8.189411163330078, 0.9750556945800781, 6.203369140625, 12.20623779296875, 2.3429489135742188, 3.962261199951172, 12.036064147949219, 0.11888694763183594, 5.418422698974609, 5.615194320678711, 5.710670471191406, -0.000980377197265625, 1.1833953857421875, 17.456153869628906, 0.7854156494140625, 7.798301696777344, -0.0847015380859375, 4.793895721435547, 5.271156311035156, 6.277153015136719, 6.835273742675781, 1.7797222137451172, -0.21719741821289062, 7.9988250732421875, -0.22411727905273438, 0.8623485565185547, 3.7146472930908203, 12.869163513183594, 2.8742218017578125, 6.610786437988281, 13.801498413085938, 1.800771713256836, 2.997699737548828, 20.933746337890625, 10.661911010742188, -0.9480018615722656, 3.062091827392578, 0.15448379516601562, 13.907989501953125, 0.5898151397705078, 4.7315826416015625, 1.3263626098632812, 5.02311897277832, 11.300872802734375, 2.154052734375, 2.7460784912109375, 19.806610107421875, 1.0163707733154297, 5.287853240966797, 2.5781936645507812, 10.915145874023438, 2.5703201293945312, 11.19598388671875, 1.35382080078125, 10.23834228515625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000054.npy"}
{"epoch": 0.07929515418502203, "step": 55, "batch_size": 64, "mean": 6.4762959480285645, "std": 7.390641212463379, "min": -10.81390380859375, "p10": -1.6101850509643554, "median": 5.805845260620117, "p90": 16.65812377929688, "max": 28.67547607421875, "pos_frac": 0.8125, "sample": [7.683631896972656, 10.323577880859375, 5.2548065185546875, -3.71038818359375, 24.63336181640625, -2.1441421508789062, 7.459445953369141, 4.685787200927734, 17.554580688476562, -1.4823074340820312, 6.6049346923828125, 6.235206604003906, 2.1849899291992188, 11.998611450195312, 20.666961669921875, 9.9378662109375, 5.005073547363281, 9.9241943359375, 28.67547607421875, 3.640911102294922, 11.039962768554688, -1.658721923828125, 4.906059265136719, 9.498321533203125, 23.665969848632812, 0.08424949645996094, 17.0743408203125, -0.0750579833984375, 0.5345592498779297, 8.08087158203125, 4.904754638671875, -2.534576416015625, 6.786144256591797, 0.08695602416992188, 12.964385986328125, 4.8006134033203125, -1.631174087524414, 0.168365478515625, 0.6921119689941406, -5.578025817871094, 10.37982177734375, 3.6281967163085938, 2.77899169921875, 15.49993896484375, 6.755119323730469, -1.5612106323242188, 1.0234432220458984, 11.368011474609375, 6.039205551147461, -10.81390380859375, 2.8301315307617188, 15.68695068359375, 6.6484222412109375, 4.47052001953125, 8.193977355957031, 5.954013824462891, -1.1013031005859375, 7.913475036621094, 8.123512268066406, 13.988754272460938, 5.657676696777344, -0.28215789794921875, 5.2464752197265625, 17.1121826171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000055.npy"}
{"epoch": 0.08076358296622614, "step": 56, "batch_size": 64, "mean": 6.132702827453613, "std": 7.380139350891113, "min": -9.856311798095703, "p10": -1.6858526229858393, "median": 5.102407455444336, "p90": 16.182032012939462, "max": 26.589111328125, "pos_frac": 0.796875, "sample": [-1.9467086791992188, 7.694873809814453, -4.17950439453125, 8.951766967773438, -0.6482906341552734, -9.856311798095703, 11.31103515625, -1.077188491821289, 3.683124542236328, -3.889606475830078, 3.9625473022460938, 2.420703887939453, 4.544609069824219, 17.40283203125, 2.2201271057128906, 0.3900604248046875, -0.2720909118652344, 12.091072082519531, 10.413978576660156, 8.412422180175781, -2.1337966918945312, 11.557510375976562, 2.7139739990234375, 24.091796875, 8.611358642578125, 17.08837890625, 1.239156723022461, -0.2768669128417969, 9.402406692504883, 2.646860122680664, 7.336616516113281, 7.6100311279296875, 9.585128784179688, 1.5526180267333984, 10.8961181640625, 12.365543365478516, -0.14928817749023438, 19.900344848632812, 7.816585540771484, 6.7369384765625, 5.989774703979492, 4.9041290283203125, 2.8646621704101562, 21.4559326171875, 7.108512878417969, 14.067222595214844, 6.095609664916992, 10.565658569335938, 0.40924072265625, 26.589111328125, 1.3121833801269531, 0.013479232788085938, 4.753837585449219, 14.06597900390625, -7.4295501708984375, 3.207366943359375, -0.327239990234375, 22.59825897216797, 5.300685882568359, -3.22210693359375, 3.09332275390625, 1.9920082092285156, 6.630195617675781, 10.233833312988281], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000056.npy"}
{"epoch": 0.08223201174743025, "step": 57, "batch_size": 64, "mean": 6.95832633972168, "std": 6.658350944519043, "min": -10.85540771484375, "p10": 0.2182605743408205, "median": 6.879890441894531, "p90": 15.486381912231446, "max": 25.005386352539062, "pos_frac": 0.921875, "sample": [9.08639144897461, 3.2658138275146484, 0.5655593872070312, -2.9779052734375, 10.875503540039062, -10.85540771484375, 13.715957641601562, 0.3958320617675781, 14.48101806640625, 6.972023010253906, 17.169189453125, 3.9001121520996094, 25.005386352539062, 2.620513916015625, 8.162002563476562, 7.5242462158203125, -1.1379566192626953, 3.988321304321289, 6.787757873535156, 20.901168823242188, 10.619255065917969, 18.1004638671875, 12.089157104492188, 4.162498474121094, 7.2073516845703125, 1.3900527954101562, 15.217288970947266, 5.516939163208008, 4.383613586425781, 2.352630615234375, 0.672271728515625, 3.887399673461914, 3.719757080078125, 0.14215850830078125, 8.04461669921875, 8.313140869140625, 0.10787200927734375, 1.6896781921386719, 3.180135726928711, 9.617866516113281, 12.3953857421875, 2.1619644165039062, 10.463829040527344, 5.9813995361328125, 10.493759155273438, -3.0207252502441406, 9.0299072265625, 7.7675628662109375, 8.61712646484375, 3.5162525177001953, 2.688447952270508, 18.44085693359375, 5.522151947021484, 3.6314125061035156, 7.694427490234375, 5.055938720703125, 1.1098136901855469, 12.992874145507812, 13.748649597167969, -7.0962066650390625, 15.601707458496094, 21.471664428710938, 8.231595993041992, 7.973419189453125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000057.npy"}
{"epoch": 0.08370044052863436, "step": 58, "batch_size": 64, "mean": 7.706109523773193, "std": 8.075557708740234, "min": -9.13629150390625, "p10": -0.08036422729492174, "median": 5.779872894287109, "p90": 18.51816787719727, "max": 28.395538330078125, "pos_frac": 0.890625, "sample": [16.086318969726562, -2.0962162017822266, 0.5531234741210938, 5.086235046386719, 3.991058349609375, 4.645040512084961, 21.795654296875, 2.4297561645507812, 16.25336456298828, 2.780885696411133, 7.195159912109375, 7.482734680175781, 8.835861206054688, 4.1461944580078125, 2.427947998046875, 12.694766998291016, 11.4295654296875, 9.561391830444336, 4.661828994750977, 3.0711231231689453, 8.907875061035156, 15.803176879882812, 14.409523010253906, -2.0460891723632812, 12.504707336425781, 1.6723861694335938, 10.142845153808594, 6.71537971496582, 23.433242797851562, 4.758872985839844, 28.395538330078125, 9.060405731201172, 6.847141265869141, 0.048503875732421875, 16.901206970214844, 1.1627769470214844, 15.859001159667969, 4.481960296630859, 2.506072998046875, -1.7817230224609375, 2.457122802734375, 2.6900672912597656, 25.345870971679688, -1.8964920043945312, 10.947509765625, -0.13559341430664062, 1.6113662719726562, 7.2356414794921875, -6.71405029296875, 25.747955322265625, 1.8898067474365234, 6.4735107421875, 0.9672908782958984, 15.828052520751953, 12.033161163330078, -9.13629150390625, 23.471542358398438, 2.248748779296875, 3.3345260620117188, 7.482765197753906, 16.884742736816406, 1.7924346923828125, 0.6055564880371094, 19.211151123046875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000058.npy"}
{"epoch": 0.08516886930983847, "step": 59, "batch_size": 64, "mean": 8.487316131591797, "std": 8.810294151306152, "min": -8.268905639648438, "p10": -1.210660934448242, "median": 7.304224014282227, "p90": 21.778811645507815, "max": 36.03564453125, "pos_frac": 0.875, "sample": [9.4990234375, 9.181022644042969, 18.396286010742188, 21.443359375, 17.45484161376953, 4.25201416015625, 36.03564453125, 22.603256225585938, 7.962123870849609, 19.274826049804688, 0.6663818359375, 7.247333526611328, -4.301822662353516, 2.9239654541015625, 25.369781494140625, -0.9879722595214844, 2.4530582427978516, 15.706756591796875, 6.773834228515625, 22.587020874023438, 1.4577350616455078, 2.5598602294921875, 1.131744384765625, 1.827362060546875, 4.719367980957031, 28.222320556640625, 2.606649398803711, 2.0587196350097656, 8.774993896484375, -8.268905639648438, 10.652053833007812, -2.3153038024902344, 9.160972595214844, 8.289073944091797, 10.573631286621094, 19.64427947998047, 4.1063690185546875, 13.091934204101562, 8.308685302734375, 4.147819519042969, -6.86212158203125, -4.346763610839844, 5.953168869018555, -1.3060989379882812, 4.974111557006836, 10.566570281982422, 7.820774078369141, 3.62408447265625, 8.638481140136719, 3.0013046264648438, 2.2901382446289062, 7.361114501953125, 3.7391529083251953, 10.549213409423828, 6.292510986328125, 16.379730224609375, 10.739913940429688, 5.1657257080078125, 4.403167724609375, -2.5944442749023438, 19.32925033569336, 22.28326416015625, 7.9733123779296875, 21.922576904296875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000059.npy"}
{"epoch": 0.08663729809104258, "step": 60, "batch_size": 64, "mean": 6.6707658767700195, "std": 7.790846347808838, "min": -17.661529541015625, "p10": -2.0464210510253906, "median": 6.928343772888184, "p90": 13.679589843750001, "max": 27.87012481689453, "pos_frac": 0.828125, "sample": [8.741287231445312, -5.67730712890625, 2.1245594024658203, 12.854568481445312, 1.3159523010253906, 3.2620162963867188, 12.37127685546875, 9.031929016113281, 1.4269561767578125, -17.661529541015625, 6.960641860961914, -2.5782623291015625, 2.1846961975097656, 4.286006927490234, 12.049430847167969, 6.7025909423828125, 4.8378448486328125, 12.889617919921875, 4.8877410888671875, 11.287017822265625, 13.193527221679688, 13.766921997070312, 21.46185302734375, 9.915687561035156, 9.466133117675781, 8.542003631591797, -0.8888263702392578, 7.203529357910156, 12.36444091796875, 0.20617103576660156, 11.148426055908203, 19.974327087402344, 3.821758270263672, 6.049945831298828, 9.12417984008789, -1.9334945678710938, 12.623214721679688, 7.885009765625, 2.6958694458007812, -6.416961669921875, 13.475814819335938, 3.7625083923339844, 6.395561218261719, 9.676597595214844, 4.9229888916015625, -1.11370849609375, -12.368206024169922, 19.182601928710938, 22.350997924804688, 12.503265380859375, 13.874244689941406, 2.8347396850585938, -2.094818115234375, -1.5386810302734375, 12.129600524902344, 6.896045684814453, 27.87012481689453, 10.958671569824219, 2.2538604736328125, 10.511171340942383, -2.278656005859375, 9.232765197753906, 4.802560806274414, 1.1881904602050781], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000060.npy"}
{"epoch": 0.0881057268722467, "step": 61, "batch_size": 64, "mean": 8.314947128295898, "std": 11.443917274475098, "min": -11.147026062011719, "p10": -1.8104660034179685, "median": 5.896364212036133, "p90": 21.33400497436524, "max": 63.52325439453125, "pos_frac": 0.796875, "sample": [3.29949951171875, 1.7227516174316406, -1.6099700927734375, 1.0603713989257812, 16.436439514160156, -0.1636371612548828, 0.3686065673828125, 8.986320495605469, -0.8556442260742188, 30.086471557617188, 24.960769653320312, 22.678512573242188, 28.75756072998047, 14.470451354980469, 8.726997375488281, 10.75412368774414, -1.9724655151367188, 1.6993274688720703, 12.33718490600586, 13.121658325195312, 3.2184524536132812, 5.491283416748047, 18.260482788085938, 10.385265350341797, 7.045066833496094, 7.205333709716797, 6.212806701660156, 6.6367034912109375, 10.423309326171875, 9.370513916015625, -0.353912353515625, -1.896392822265625, 13.733589172363281, 18.682891845703125, 19.544677734375, 25.230743408203125, 20.054931640625, 1.83074951171875, -2.3427734375, 14.550857543945312, 3.1072006225585938, 0.28751182556152344, 15.53143310546875, 3.2446861267089844, 0.7982196807861328, 21.882179260253906, 1.0992279052734375, 4.379341125488281, 63.52325439453125, 14.6668701171875, 13.87890625, -1.5753059387207031, -11.147026062011719, 13.203483581542969, 2.439706802368164, -5.847541809082031, 5.579921722412109, 3.8706512451171875, -8.828178405761719, 2.7970008850097656, 7.65435791015625, -6.9351806640625, -1.5149116516113281, 1.9109630584716797], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000061.npy"}
{"epoch": 0.08957415565345081, "step": 62, "batch_size": 64, "mean": 6.243136405944824, "std": 8.393269538879395, "min": -11.689666748046875, "p10": -3.999816322326659, "median": 5.882681846618652, "p90": 17.593785095214844, "max": 34.23078918457031, "pos_frac": 0.859375, "sample": [9.921951293945312, 1.600006103515625, 0.6514883041381836, 11.833419799804688, 5.836679458618164, 13.462982177734375, 8.593620300292969, 0.9745769500732422, 10.335067749023438, -6.231048583984375, 17.029296875, 7.0326690673828125, 1.7169265747070312, 0.13126754760742188, 11.618461608886719, 3.7744293212890625, 11.671890258789062, 6.597320556640625, 0.3636455535888672, 5.286979675292969, 7.916252136230469, 7.893867492675781, 1.8572845458984375, 6.183422088623047, -5.281787872314453, 10.0162353515625, 4.827249526977539, 8.866561889648438, -0.42580223083496094, 34.23078918457031, 18.640113830566406, 9.425651550292969, 6.801933288574219, 2.5331993103027344, 0.7325897216796875, 6.475395202636719, 1.3705291748046875, 20.473251342773438, -11.689666748046875, -5.259407043457031, -7.3909912109375, 0.6701984405517578, 1.3505096435546875, 0.9488410949707031, 1.6718101501464844, 9.977567672729492, 17.98461151123047, 10.206829071044922, -4.342321395874023, 17.835708618164062, 9.074832916259766, 5.928684234619141, -8.63140869140625, 4.8181304931640625, 14.300201416015625, 4.126655578613281, 3.362834930419922, 4.237495422363281, 30.2095947265625, 22.952194213867188, 7.283733367919922, 1.7321014404296875, -3.2006378173828125, 6.6642303466796875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000062.npy"}
{"epoch": 0.09104258443465492, "step": 63, "batch_size": 64, "mean": 8.424701690673828, "std": 9.105888366699219, "min": -14.760383605957031, "p10": -0.7347524642944327, "median": 8.03676986694336, "p90": 19.12708435058594, "max": 36.806396484375, "pos_frac": 0.890625, "sample": [27.83747100830078, 8.857376098632812, 5.090154647827148, 9.544715881347656, 2.85302734375, 1.0513687133789062, 5.7808074951171875, -8.391036987304688, 7.5949859619140625, 7.4211273193359375, -14.760383605957031, 18.41899871826172, 9.577224731445312, 8.539398193359375, 3.2665328979492188, -1.1033992767333984, 19.73046875, 4.766469955444336, 14.937164306640625, 23.209060668945312, 13.507843017578125, 15.455696105957031, 7.314979553222656, 0.2528533935546875, -1.335174560546875, 11.551752090454102, -3.7524566650390625, 0.12542343139648438, 5.75860595703125, 3.5446529388427734, 13.338699340820312, 9.359962463378906, 8.478553771972656, 11.185619354248047, 12.632369995117188, 13.8765869140625, 16.60334014892578, 15.449836730957031, 5.529106140136719, 0.9879264831542969, 0.3772430419921875, 13.912490844726562, 3.1974048614501953, 4.927833557128906, -13.233383178710938, 14.584585189819336, 9.079338073730469, 2.0792236328125, 1.846221923828125, 6.8182830810546875, 3.737213134765625, 16.60173797607422, 20.195266723632812, 8.78121566772461, 19.43054962158203, 0.9671440124511719, 9.69593620300293, 36.806396484375, -3.7497406005859375, 16.562297821044922, 4.8470306396484375, 29.05866241455078, 5.298553466796875, 13.27166748046875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000063.npy"}
{"epoch": 0.09251101321585903, "step": 64, "batch_size": 64, "mean": 9.953544616699219, "std": 11.583884239196777, "min": -21.050430297851562, "p10": -1.346813583374023, "median": 8.18783187866211, "p90": 24.34029693603516, "max": 54.379150390625, "pos_frac": 0.8125, "sample": [1.3795700073242188, 20.18145751953125, 30.237045288085938, 8.323310852050781, 5.874479293823242, 13.409530639648438, 3.8891544342041016, 17.62139129638672, -0.14348411560058594, 7.693300247192383, 2.7641983032226562, 11.758880615234375, 0.5086441040039062, 5.414146423339844, 13.358146667480469, 35.06065368652344, 5.3053741455078125, -2.2532901763916016, 20.569808959960938, 8.35103988647461, 12.705780029296875, -3.8452377319335938, 31.91375732421875, 8.052352905273438, -0.7145061492919922, 5.693660736083984, -1.677877426147461, 7.888212203979492, -21.050430297851562, -0.7977142333984375, 8.499801635742188, -1.5821418762207031, 54.379150390625, 6.052951812744141, 2.1260948181152344, 11.802440643310547, 19.75684356689453, 23.489212036132812, 12.138946533203125, 6.457927703857422, 6.673397064208984, 26.539077758789062, 14.039634704589844, 10.986381530761719, -0.3777618408203125, -10.339225769042969, 28.907928466796875, 3.5342559814453125, 24.705047607421875, 0.17630767822265625, 16.856658935546875, 13.667984008789062, 14.665306091308594, -1.8311805725097656, 14.395988464355469, -0.4858722686767578, 5.582387924194336, 8.96224594116211, 15.230308532714844, 13.857803344726562, 1.3902130126953125, 12.71624755859375, 20.40435028076172, 6.176769256591797], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000064.npy"}
{"epoch": 0.09397944199706314, "step": 65, "batch_size": 64, "mean": 10.083711624145508, "std": 11.472171783447266, "min": -26.60980987548828, "p10": -1.176770210266113, "median": 8.233183860778809, "p90": 25.62964935302735, "max": 35.81349182128906, "pos_frac": 0.84375, "sample": [35.81349182128906, 28.69610595703125, 20.902694702148438, 12.922599792480469, 6.213796615600586, 11.145866394042969, 0.41580963134765625, 14.715278625488281, 3.8472442626953125, 23.244239807128906, 18.551979064941406, 1.9545726776123047, -0.3898639678955078, 26.254638671875, 32.48826599121094, 1.1809349060058594, 11.099929809570312, 2.6908092498779297, -4.777008056640625, 5.50482177734375, 1.5136775970458984, 21.303665161132812, 6.452323913574219, -26.60980987548828, 19.741653442382812, 9.681991577148438, -4.383270263671875, 8.423954010009766, 15.508956909179688, 5.536964416503906, 5.986907958984375, 5.8074188232421875, -2.7767601013183594, 5.99848747253418, 32.11834716796875, 8.191604614257812, 23.69170379638672, 16.90550994873047, 29.21556854248047, 8.71282958984375, -0.8884124755859375, -14.928741455078125, 10.607086181640625, 2.004800796508789, 7.722198486328125, 21.373802185058594, 28.16370391845703, -2.082183837890625, 7.507904052734375, 23.542007446289062, 12.934440612792969, 5.52337646484375, 24.171340942382812, -1.3003520965576172, 13.966949462890625, 10.306060791015625, 6.182487487792969, 2.7509307861328125, 18.297134399414062, 10.421585083007812, -0.6957550048828125, 8.274763107299805, 0.1982555389404297, 7.80620002746582], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000065.npy"}
{"epoch": 0.09544787077826726, "step": 66, "batch_size": 64, "mean": 9.253350257873535, "std": 11.441926002502441, "min": -10.580841064453125, "p10": -1.4261718749999996, "median": 7.1121320724487305, "p90": 21.814537048339847, "max": 52.88410949707031, "pos_frac": 0.828125, "sample": [8.510971069335938, 52.88410949707031, -0.17928314208984375, 17.436309814453125, 14.804153442382812, 13.925033569335938, 2.2384586334228516, -0.5548973083496094, 11.067453384399414, 7.15203857421875, -10.208786010742188, 2.5243148803710938, 7.449726104736328, -0.9665298461914062, 13.315296173095703, 4.285640716552734, 5.083932876586914, 0.2037029266357422, 2.3514461517333984, 22.301483154296875, 10.089469909667969, 11.352933883666992, 0.3906383514404297, 12.241981506347656, 19.85302734375, 2.4149093627929688, 6.5312652587890625, -3.75140380859375, -4.571496963500977, -10.580841064453125, 13.007728576660156, 18.81395721435547, 20.678329467773438, 8.604209899902344, 3.7016220092773438, 13.468704223632812, 6.390296936035156, -0.6210498809814453, 2.7401561737060547, 15.889961242675781, 11.888114929199219, 3.7065200805664062, -1.6231613159179688, 7.061920166015625, 7.072225570678711, -10.508682250976562, 9.980499267578125, 1.4283905029296875, -4.782325744628906, 15.242088317871094, 4.94189453125, 23.394866943359375, 31.535446166992188, 23.65442657470703, 5.411380767822266, 19.370201110839844, 2.9700393676757812, 9.623516082763672, 17.297271728515625, 2.4697532653808594, 38.39422607421875, 1.67047119140625, 36.66705322265625, 15.079307556152344], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000066.npy"}
{"epoch": 0.09691629955947137, "step": 67, "batch_size": 64, "mean": 9.497722625732422, "std": 11.531469345092773, "min": -15.198028564453125, "p10": -1.5380809783935545, "median": 6.967222213745117, "p90": 23.91813888549805, "max": 54.619049072265625, "pos_frac": 0.8125, "sample": [25.360679626464844, 4.462089538574219, 24.128456115722656, -0.134246826171875, 2.5804805755615234, -1.3032913208007812, 17.999557495117188, 5.900754928588867, -7.287681579589844, 2.373748779296875, 19.633220672607422, 11.121330261230469, 4.356605529785156, 18.453102111816406, -0.61859130859375, -4.862892150878906, 10.071685791015625, 11.137649536132812, 1.4626274108886719, -15.198028564453125, 23.427398681640625, 9.614341735839844, -1.4407463073730469, -0.057952880859375, 6.898338317871094, 2.0591659545898438, 24.60363006591797, 7.332233428955078, -1.5797958374023438, 6.308509826660156, 1.169036865234375, 16.832561492919922, 8.119583129882812, 3.015848159790039, 3.2044906616210938, 10.823715209960938, 16.195222854614258, 6.317071914672852, 1.6092720031738281, -6.1591033935546875, 7.030609130859375, 19.80615234375, 4.967678070068359, 4.242704391479492, 30.517578125, 16.705612182617188, 6.903835296630859, 19.455291748046875, 15.92205810546875, 16.127151489257812, 13.181892395019531, 6.15498161315918, 12.960409164428711, 37.287933349609375, 7.190837860107422, 1.2061328887939453, 54.619049072265625, 3.280488967895508, -2.9858169555664062, 29.70977783203125, 22.688812255859375, 8.511493682861328, -4.6474456787109375, 9.086965560913086], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000067.npy"}
{"epoch": 0.09838472834067548, "step": 68, "batch_size": 64, "mean": 9.00277328491211, "std": 9.503949165344238, "min": -9.692108154296875, "p10": -0.011645126342772899, "median": 6.615828514099121, "p90": 22.728830718994146, "max": 43.078094482421875, "pos_frac": 0.890625, "sample": [-5.3827056884765625, 2.1502532958984375, 28.202178955078125, 9.663311004638672, 6.288810729980469, -7.7975616455078125, 10.145614624023438, 25.47601318359375, 2.0374298095703125, 0.512054443359375, 14.9058837890625, 17.36767578125, 20.3780517578125, 12.652080535888672, 10.70404052734375, -1.263885498046875, 7.80615234375, 13.90450668334961, 4.259284973144531, 9.291389465332031, -1.3791961669921875, 5.239994049072266, 5.313852310180664, 1.104278564453125, 8.439605712890625, 12.2154541015625, 21.303436279296875, 8.794578552246094, 10.58816909790039, 20.895538330078125, 23.33971405029297, 25.890758514404297, 1.3260650634765625, 6.910530090332031, 11.323455810546875, 1.8495292663574219, -0.23608779907226562, 6.391073226928711, 2.2541580200195312, 6.289588928222656, 19.069995880126953, 2.785198211669922, 11.169466018676758, -9.692108154296875, 6.33245849609375, 6.840583801269531, 4.089286804199219, 23.425735473632812, 2.3985366821289062, 1.6443958282470703, 1.9629192352294922, 43.078094482421875, 27.703048706054688, 17.705734252929688, 15.374603271484375, 1.0029678344726562, 2.854248046875, 4.996101379394531, 12.192092895507812, -1.9776763916015625, 6.120452880859375, 3.788846969604492, 5.162017822265625, 8.995399475097656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000068.npy"}
{"epoch": 0.09985315712187959, "step": 69, "batch_size": 64, "mean": 11.27790641784668, "std": 11.517322540283203, "min": -7.104400634765625, "p10": -0.25604515075683576, "median": 7.574008941650391, "p90": 24.60918426513672, "max": 52.812591552734375, "pos_frac": 0.859375, "sample": [4.7735443115234375, 7.739116668701172, 35.81669616699219, 11.9290771484375, 14.224166870117188, 19.739501953125, -7.104400634765625, -2.543487548828125, -0.34378814697265625, 5.793281555175781, 6.3664398193359375, 10.716686248779297, 0.064453125, 10.681747436523438, 5.904216766357422, 17.062923431396484, 33.166900634765625, 7.8973541259765625, 19.930938720703125, 18.684715270996094, -2.670684814453125, 30.761016845703125, 14.503181457519531, 10.01223373413086, 21.176773071289062, 16.56707763671875, 4.6562042236328125, 17.697891235351562, 14.634254455566406, 6.971025466918945, 0.8041038513183594, -0.051311492919921875, 21.925918579101562, 52.812591552734375, 5.773414611816406, 24.528060913085938, 6.814140319824219, 15.560455322265625, 21.896835327148438, 1.2130203247070312, 1.7226982116699219, 19.190338134765625, 11.567363739013672, 2.7152099609375, -1.168436050415039, 17.01177978515625, 7.408901214599609, 41.1951904296875, 5.1090087890625, 9.39790153503418, 6.470806121826172, -5.527824401855469, 7.0421142578125, 3.7729644775390625, 2.3513565063476562, 6.2047882080078125, 24.83087158203125, 24.643951416015625, -3.0517578125e-05, 23.6824951171875, 2.2760562896728516, 1.428955078125, 5.661842346191406, -1.2885894775390625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000069.npy"}
{"epoch": 0.1013215859030837, "step": 70, "batch_size": 64, "mean": 12.264190673828125, "std": 13.555784225463867, "min": -14.128631591796875, "p10": -2.6378849029541014, "median": 11.470687866210938, "p90": 29.272846984863293, "max": 50.015472412109375, "pos_frac": 0.84375, "sample": [7.481203079223633, 0.9892578125, 13.37002944946289, 22.027076721191406, 24.932472229003906, 14.228410720825195, -11.333511352539062, -3.7298736572265625, 15.731712341308594, 4.400796890258789, -3.4905853271484375, 17.03697967529297, -3.4662342071533203, 11.72314453125, 46.59040832519531, 1.770925521850586, 10.328338623046875, 31.400619506835938, 13.236108779907227, 50.015472412109375, 14.346450805664062, 11.218231201171875, 1.8670234680175781, 35.02166748046875, 0.5562362670898438, 20.085006713867188, 23.693359375, 14.700239181518555, -2.9463443756103516, 8.326459884643555, 7.608154296875, 30.30516815185547, 2.5038719177246094, 0.7229423522949219, 1.4870452880859375, -1.7441329956054688, 16.910457611083984, 4.102897644042969, 15.10784912109375, -2.6699867248535156, 4.541473388671875, 4.613626480102539, -2.5629806518554688, 0.955535888671875, 12.477363586425781, 2.4263763427734375, 22.67749786376953, 26.411636352539062, -14.128631591796875, 5.700492858886719, 22.046218872070312, 4.780574798583984, 1.7285270690917969, 40.602081298828125, 16.340538024902344, -1.4243354797363281, 41.540924072265625, 12.309051513671875, 25.23377227783203, 19.63359832763672, 5.0039825439453125, 26.864097595214844, 16.424774169921875, 26.266708374023438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000070.npy"}
{"epoch": 0.1027900146842878, "step": 71, "batch_size": 64, "mean": 14.531265258789062, "std": 15.077340126037598, "min": -11.10357666015625, "p10": 0.6260429382324239, "median": 9.662269592285156, "p90": 28.388186645507822, "max": 64.44268798828125, "pos_frac": 0.890625, "sample": [-6.448143005371094, 52.867523193359375, 6.236019134521484, 3.3199386596679688, 7.928264617919922, 15.10744857788086, 6.671298980712891, 13.968902587890625, 19.196331024169922, 7.170219421386719, 26.283294677734375, 9.382970809936523, 23.030914306640625, 5.058233261108398, 4.4932403564453125, 6.553871154785156, 12.027965545654297, 19.472274780273438, -0.3141021728515625, 2.599407196044922, 4.38092041015625, 55.12733459472656, 9.253063201904297, 8.679069519042969, 19.629638671875, 18.853271484375, 6.879055023193359, 16.039932250976562, 15.409860610961914, -11.10357666015625, 2.615692138671875, 9.476043701171875, 16.256515502929688, -0.21968460083007812, 5.97175407409668, 21.266159057617188, 23.62073516845703, -0.650299072265625, 24.95306396484375, 22.028915405273438, 3.9902572631835938, 17.10124969482422, 6.7474517822265625, 15.676788330078125, 19.315567016601562, 22.74059295654297, 3.8106155395507812, 44.23420715332031, 58.393829345703125, 64.44268798828125, 3.131072998046875, 29.290283203125, -10.783760070800781, -1.5883331298828125, 18.465717315673828, 9.140060424804688, 8.421844482421875, 12.657173156738281, 22.345123291015625, 20.277870178222656, 37.561553955078125, 4.464080810546875, 7.243257522583008, 9.848495483398438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000071.npy"}
{"epoch": 0.10425844346549193, "step": 72, "batch_size": 64, "mean": 15.254532814025879, "std": 17.02393913269043, "min": -22.360107421875, "p10": -3.06773300170898, "median": 13.769417762756348, "p90": 39.27478942871095, "max": 70.77267456054688, "pos_frac": 0.890625, "sample": [21.96881103515625, 18.511648178100586, 6.7499542236328125, 15.962631225585938, 4.402069091796875, 5.655586242675781, 11.425468444824219, 32.724510192871094, -19.92462158203125, 30.281585693359375, 18.578624725341797, 17.816024780273438, 8.210821151733398, 1.087005615234375, 19.675277709960938, 10.587276458740234, 16.199295043945312, 32.98565673828125, 19.512596130371094, 41.107757568359375, 47.023834228515625, -6.408552169799805, 8.908210754394531, 11.543146133422852, 22.36516571044922, 3.1557769775390625, 21.662662506103516, 13.25396728515625, 15.663261413574219, 20.695289611816406, 4.565406799316406, 5.727470397949219, 15.573440551757812, -20.467361450195312, 45.056884765625, 34.99786376953125, 22.694290161132812, 8.180158615112305, 25.833877563476562, -6.615692138671875, 11.020015716552734, 52.75007629394531, 70.77267456054688, -8.683563232421875, 11.609939575195312, 24.19232177734375, 8.19598388671875, 10.121509552001953, 1.5480327606201172, 45.378173828125, -22.360107421875, 4.833736419677734, 1.9540138244628906, 18.359085083007812, 17.79163360595703, 22.741615295410156, 11.545021057128906, 18.36284637451172, 46.146759033203125, 14.284868240356445, 6.78778076171875, 5.124107360839844, -4.848335266113281, 1.7348060607910156], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000072.npy"}
{"epoch": 0.10572687224669604, "step": 73, "batch_size": 64, "mean": 15.381328582763672, "std": 23.727678298950195, "min": -35.602874755859375, "p10": -13.453687477111815, "median": 12.1298828125, "p90": 40.35156860351563, "max": 83.63969421386719, "pos_frac": 0.78125, "sample": [36.31092071533203, -20.09070587158203, 9.956684112548828, 19.83959197998047, 8.965263366699219, 45.172447204589844, -10.043777465820312, 26.971420288085938, 10.11962890625, -0.17649078369140625, 37.273895263671875, 37.58274459838867, 29.862045288085938, -1.5146026611328125, 0.29803466796875, -19.763092041015625, 70.57952880859375, -2.1070175170898438, 12.367271423339844, 23.65190887451172, 20.264724731445312, 9.594429016113281, 13.918594360351562, 1.3303604125976562, 2.8967208862304688, 12.02880859375, -24.43204116821289, 12.23095703125, 38.1578369140625, 17.8826904296875, 16.084392547607422, 4.429132461547852, 72.85678100585938, -16.386734008789062, 4.437305450439453, 37.095489501953125, 38.59967803955078, -12.183675765991211, 5.896694183349609, 1.936777114868164, -5.683128356933594, -7.152732849121094, -13.997978210449219, 7.627527236938477, 18.075485229492188, 18.96253204345703, 3.6981563568115234, 6.013402938842773, 25.36748504638672, 53.98933410644531, 19.72802734375, 7.621700286865234, 30.4014892578125, 5.714138031005859, 18.850440979003906, -14.134353637695312, 15.152999877929688, 72.25764465332031, 37.60960388183594, 83.63969421386719, 3.2356433868408203, 41.102378845214844, 20.033767700195312, -35.602874755859375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000073.npy"}
{"epoch": 0.10719530102790015, "step": 74, "batch_size": 64, "mean": 16.83268928527832, "std": 26.327777862548828, "min": -49.37757110595703, "p10": -4.24322624206543, "median": 12.859125137329102, "p90": 39.16424865722657, "max": 158.296875, "pos_frac": 0.796875, "sample": [30.925094604492188, 10.526557922363281, 16.449356079101562, 0.03997039794921875, 27.2015380859375, 9.536590576171875, 12.909587860107422, 67.17363739013672, 2.243640899658203, 23.387962341308594, 20.800048828125, 11.0330810546875, 23.745067596435547, -0.69793701171875, 18.555801391601562, 158.296875, 10.150245666503906, 37.18955993652344, 17.995166778564453, 4.363922119140625, 2.349956512451172, 62.86241149902344, 6.1019439697265625, 40.01054382324219, 28.41790771484375, 12.808662414550781, -4.242321014404297, -4.606170654296875, -0.9512901306152344, -24.12708282470703, 18.956748962402344, -3.0938720703125, 0.8432540893554688, 10.225021362304688, -49.37757110595703, 25.17681121826172, -3.6931228637695312, 32.249446868896484, 4.82044792175293, -6.996334075927734, -12.823814392089844, 13.015800476074219, 23.780197143554688, 26.795928955078125, 0.9557685852050781, 60.263336181640625, 16.19927978515625, 0.724700927734375, 12.108139038085938, 10.864627838134766, 12.354667663574219, 43.690338134765625, 23.225936889648438, 29.2127685546875, 46.438987731933594, 17.668411254882812, 5.767431259155273, 22.752708435058594, 34.82029724121094, 32.69304656982422, -1.6176910400390625, -4.243614196777344, -4.541385650634766, 17.625137329101562], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000074.npy"}
{"epoch": 0.10866372980910426, "step": 75, "batch_size": 64, "mean": 23.66238021850586, "std": 24.678485870361328, "min": -16.822364807128906, "p10": -2.374828338623045, "median": 18.202049255371094, "p90": 58.14931030273438, "max": 89.9029541015625, "pos_frac": 0.859375, "sample": [-0.48175811767578125, 30.889205932617188, 19.815399169921875, 14.310028076171875, 22.068862915039062, 13.60809326171875, 12.43105697631836, 6.168067932128906, 55.94329833984375, 83.073486328125, 36.23814392089844, 17.844501495361328, -11.585922241210938, 33.05272674560547, 89.9029541015625, 18.55959701538086, 50.291015625, 73.73576354980469, 10.964010238647461, -0.6465377807617188, 52.305633544921875, 29.972930908203125, 17.66061782836914, 10.321365356445312, 13.837799072265625, 56.76560974121094, 58.74232482910156, 3.738008499145508, 9.852836608886719, 36.21756362915039, 21.960052490234375, 20.867420196533203, 2.1205902099609375, 77.80690002441406, -7.718208312988281, -3.1155242919921875, 30.727584838867188, -16.822364807128906, 24.434555053710938, 23.279911041259766, 8.519046783447266, 13.965938568115234, -11.492218017578125, 10.724235534667969, -7.027400970458984, 35.072845458984375, 3.562541961669922, 15.503433227539062, 48.840606689453125, 33.13189697265625, 4.75080680847168, -10.486785888671875, 9.324529647827148, 9.004051208496094, 18.963851928710938, 6.044288635253906, 6.93719482421875, 28.856414794921875, 17.75843048095703, 20.219970703125, 46.14257049560547, 73.7457275390625, 73.97076416015625, 19.221839904785156], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000075.npy"}
{"epoch": 0.11013215859030837, "step": 76, "batch_size": 64, "mean": 14.961589813232422, "std": 19.148834228515625, "min": -53.86041259765625, "p10": -4.067306900024413, "median": 14.047431945800781, "p90": 36.604196929931646, "max": 79.78627014160156, "pos_frac": 0.84375, "sample": [15.53154182434082, 5.247699737548828, 3.3006820678710938, 29.785125732421875, 4.3550567626953125, 17.227813720703125, 7.3668975830078125, 30.205657958984375, 37.1191291809082, 17.64264678955078, 19.623497009277344, 25.964309692382812, 11.132102966308594, 35.959327697753906, 4.541831970214844, 6.125232696533203, 9.62324333190918, 79.78627014160156, 41.216835021972656, 6.899940490722656, -9.558753967285156, 13.997268676757812, -0.44809913635253906, 15.264839172363281, -18.399627685546875, 6.466165542602539, 6.597593307495117, 19.463577270507812, 7.036888122558594, 6.497829437255859, 34.85206604003906, 19.37070083618164, 25.3299560546875, 10.75877571105957, 33.860443115234375, 4.812530517578125, -4.731319427490234, 32.06971740722656, 2.5345077514648438, -13.961471557617188, 7.835639953613281, 3.43878173828125, -8.968202590942383, -11.780426025390625, 20.24059295654297, 36.88056945800781, 58.701019287109375, 33.33384704589844, 10.165294647216797, 7.464500427246094, 13.062492370605469, -2.5179443359375, 16.46099853515625, 22.173057556152344, 19.340179443359375, 14.812999725341797, -0.7216205596923828, 18.31964874267578, 42.05201721191406, 27.985145568847656, 27.82598876953125, 40.73150634765625, 14.09759521484375, -53.86041259765625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000076.npy"}
{"epoch": 0.11160058737151249, "step": 77, "batch_size": 64, "mean": 24.377437591552734, "std": 26.402814865112305, "min": -53.5345458984375, "p10": -0.6880558013915994, "median": 25.063209533691406, "p90": 65.7834945678711, "max": 90.84963989257812, "pos_frac": 0.890625, "sample": [20.556442260742188, 49.24250030517578, 35.86796951293945, 47.20515823364258, 49.30271911621094, -1.6093063354492188, 68.61395263671875, 25.608577728271484, 10.117074966430664, 11.284812927246094, 7.417779922485352, 41.61347961425781, 45.54606628417969, 5.945896148681641, 2.9213104248046875, 32.47273254394531, -53.5345458984375, 27.353775024414062, 44.470298767089844, 37.19383239746094, 15.620542526245117, -3.3342018127441406, 11.395088195800781, 1.4615287780761719, 4.708099365234375, 64.96229553222656, 4.338384628295898, 31.523902893066406, 13.651840209960938, -6.511884689331055, 37.1386604309082, 32.874237060546875, 14.855880737304688, 75.11613464355469, 39.074195861816406, 11.365238189697266, 70.66079711914062, 90.84963989257812, 24.12781524658203, 17.003738403320312, -37.577606201171875, 10.377822875976562, -22.403106689453125, 3.1191558837890625, 14.01609992980957, 27.558998107910156, 35.72547149658203, 27.73640251159668, 29.655517578125, 29.261932373046875, 9.18341064453125, -10.195571899414062, 5.173574447631836, 30.86463165283203, 31.613807678222656, 26.373870849609375, 4.946479797363281, 66.13543701171875, 78.27122497558594, 15.039993286132812, 28.06240463256836, 67.36640930175781, 24.517841339111328, 6.859375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000077.npy"}
{"epoch": 0.1130690161527166, "step": 78, "batch_size": 64, "mean": 18.202960968017578, "std": 22.19403648376465, "min": -27.006980895996094, "p10": -5.1070922851562495, "median": 17.623817443847656, "p90": 48.274826812744145, "max": 100.38885498046875, "pos_frac": 0.78125, "sample": [-0.9275665283203125, 2.18450927734375, 47.135986328125, 19.4324951171875, -3.4685897827148438, 52.91188049316406, 18.066986083984375, 100.38885498046875, 29.37677001953125, 23.81784439086914, 23.31159210205078, 10.19693374633789, -5.407543182373047, 14.962160110473633, 23.610332489013672, 22.92209243774414, -0.913787841796875, 10.122562408447266, -5.17755126953125, 4.65777587890625, 6.686777114868164, 19.331804275512695, 30.39229965209961, 28.92412567138672, 22.407581329345703, 14.902729034423828, 19.51746368408203, 21.11407470703125, 28.716339111328125, 5.981241226196289, 48.762901306152344, 17.60332679748535, 6.433616638183594, -14.12722396850586, 29.08005142211914, -15.106048583984375, 64.35922241210938, 24.686264038085938, 49.39482116699219, -4.94268798828125, 40.753536224365234, 4.00726318359375, 56.751617431640625, 35.755680084228516, 33.703468322753906, -1.446502685546875, 12.439380645751953, 4.028072357177734, 14.39687728881836, 8.357917785644531, 7.194601058959961, -1.6217269897460938, 17.64430809020996, -4.2799835205078125, 36.40382385253906, -27.006980895996094, 54.68609619140625, 5.575038909912109, -23.37908935546875, 29.289138793945312, -7.253082275390625, 42.543174743652344, 2.9029159545898438, 32.221527099609375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000078.npy"}
{"epoch": 0.1145374449339207, "step": 79, "batch_size": 64, "mean": 19.941024780273438, "std": 26.51508140563965, "min": -28.920272827148438, "p10": -9.20103721618652, "median": 14.398792266845703, "p90": 56.52076721191408, "max": 87.22076416015625, "pos_frac": 0.765625, "sample": [-10.572441101074219, 60.098365783691406, 6.837972640991211, 80.16178894042969, 3.9690093994140625, 0.06404876708984375, -3.006105422973633, -12.83709716796875, 48.35437774658203, 39.72978210449219, 20.519683837890625, -28.920272827148438, 27.083084106445312, 81.05364990234375, 9.855220794677734, 29.88677215576172, 82.12911987304688, 57.649658203125, 19.32782745361328, -14.646759033203125, 87.22076416015625, 19.589988708496094, -4.171470642089844, -13.1871337890625, 31.854400634765625, 0.51031494140625, -3.12591552734375, 19.29184341430664, 15.577262878417969, 70.62410736083984, 2.9335174560546875, 9.503213882446289, 13.220321655273438, 30.197952270507812, -14.083465576171875, 11.316802978515625, 0.19927597045898438, 7.363883972167969, 48.914306640625, 48.46965026855469, 24.606632232666016, 36.97935485839844, 39.64422607421875, 21.008792877197266, 30.96007537841797, -5.788175582885742, 2.5655517578125, 36.86228942871094, 42.152435302734375, 12.148006439208984, 18.029438018798828, -6.001094818115234, -1.3833045959472656, 44.21162414550781, -1.4949417114257812, 24.3387451171875, 6.286861419677734, 7.774200439453125, 0.5007476806640625, -12.351730346679688, 7.084978103637695, 53.886688232421875, -4.88824462890625, 20.135025024414062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000079.npy"}
{"epoch": 0.11600587371512482, "step": 80, "batch_size": 64, "mean": 18.735233306884766, "std": 27.950864791870117, "min": -30.47199249267578, "p10": -9.5616886138916, "median": 12.129197120666504, "p90": 49.9573081970215, "max": 115.24966430664062, "pos_frac": 0.765625, "sample": [25.921058654785156, 9.19015121459961, -9.5625, 25.81543731689453, 6.526922225952148, 3.903230667114258, -4.522930145263672, -11.1893310546875, 37.704803466796875, -0.7787933349609375, 4.610441207885742, 17.633922576904297, -17.510299682617188, -24.688232421875, -10.624099731445312, 27.49095916748047, 115.24966430664062, 44.489295959472656, 94.26483154296875, 51.22551727294922, 29.759017944335938, 7.285499572753906, 46.99815368652344, 52.21399688720703, 39.451812744140625, 17.144893646240234, -1.6062660217285156, 39.75798034667969, 16.592294692993164, 8.750778198242188, 5.398994445800781, -19.293426513671875, 5.061267852783203, -7.686820983886719, 28.800552368164062, -30.47199249267578, 66.16474914550781, 36.5128173828125, -9.559795379638672, 103.72515869140625, 10.970361709594727, 4.9575042724609375, 2.1036720275878906, 3.0179901123046875, 53.90967559814453, 15.03518295288086, 26.17266082763672, 36.56169128417969, 17.05439567565918, 10.843490600585938, -4.885923385620117, -3.7003917694091797, 1.0285205841064453, 8.4029541015625, 25.33673667907715, 25.119140625, 7.964359283447266, 3.8994522094726562, 37.02284240722656, 28.561111450195312, 13.288032531738281, 46.554351806640625, -5.458715438842773, 15.14605712890625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000080.npy"}
{"epoch": 0.11747430249632893, "step": 81, "batch_size": 64, "mean": 25.64330291748047, "std": 32.318992614746094, "min": -36.7327880859375, "p10": -10.791339874267578, "median": 21.387798309326172, "p90": 66.37551574707031, "max": 135.57147216796875, "pos_frac": 0.84375, "sample": [49.434303283691406, 35.90643310546875, -36.7327880859375, 43.086570739746094, 28.302024841308594, 71.67475891113281, 22.94501495361328, -13.046276092529297, 8.201313018798828, 36.551185607910156, 35.99357604980469, -29.23980712890625, 3.0106887817382812, 29.751785278320312, 100.91796875, 13.308187484741211, 27.372787475585938, 43.158935546875, 66.57278442382812, -18.046424865722656, 1.4090137481689453, 8.196670532226562, 22.714218139648438, 80.23947143554688, 27.612937927246094, 12.141443252563477, 18.409637451171875, 9.887317657470703, 8.199762344360352, -10.622657775878906, 29.329139709472656, 5.7776947021484375, 6.313068389892578, 13.758834838867188, 29.862998962402344, 135.57147216796875, 48.572750091552734, 20.061378479003906, 2.5524215698242188, 65.91522216796875, 47.85406494140625, 56.080841064453125, 18.15662384033203, -2.3873138427734375, 4.621940612792969, -2.3880271911621094, -23.66183090209961, 16.868364334106445, 10.229042053222656, 34.94685363769531, 16.755653381347656, 13.032754898071289, 24.390213012695312, -18.055953979492188, 42.332183837890625, 49.96527099609375, 121.836181640625, 68.31396484375, 9.942512512207031, 38.243247985839844, 2.2499923706054688, 31.576889038085938, -10.863632202148438, 36.10577392578125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000081.npy"}
{"epoch": 0.11894273127753303, "step": 82, "batch_size": 64, "mean": 20.451387405395508, "std": 34.92342758178711, "min": -69.28187561035156, "p10": -19.867961502075193, "median": 18.055081367492676, "p90": 58.84064865112305, "max": 140.68350219726562, "pos_frac": 0.78125, "sample": [-12.62564468383789, 38.073097229003906, -25.01141357421875, 20.15186309814453, 23.58338165283203, 20.54183006286621, -31.57262420654297, 16.587291717529297, 28.25164031982422, 64.03192901611328, 79.60467529296875, 13.799110412597656, -17.936176300048828, 10.859634399414062, 18.124216079711914, 121.50245666503906, 14.899826049804688, -59.17131042480469, -12.385574340820312, 37.699859619140625, 2.5149154663085938, 35.94866943359375, 58.852928161621094, 4.471899032592773, -20.69586944580078, -8.535514831542969, 4.993202209472656, 28.05518341064453, 23.75238800048828, 14.301422119140625, 24.359703063964844, 40.090057373046875, 42.571739196777344, -0.9851360321044922, 85.44406127929688, 21.159198760986328, 58.81199645996094, 17.985946655273438, 9.924076080322266, 8.961845397949219, -2.6417236328125, 33.99040222167969, 70.59698486328125, -22.041412353515625, 33.505638122558594, 25.055450439453125, 7.555307388305664, 51.69575881958008, 41.317359924316406, -29.789794921875, 11.173347473144531, 7.3005218505859375, -69.28187561035156, 8.680793762207031, 37.63969421386719, 23.476959228515625, 5.53265380859375, 28.225006103515625, 17.45061492919922, 140.68350219726562, 18.162277221679688, 17.53608512878418, -0.8716278076171875, 52.94606018066406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000082.npy"}
{"epoch": 0.12041116005873716, "step": 83, "batch_size": 64, "mean": 24.361614227294922, "std": 30.451425552368164, "min": -30.444854736328125, "p10": -11.484876251220703, "median": 19.19930648803711, "p90": 72.37353973388674, "max": 98.61836242675781, "pos_frac": 0.796875, "sample": [23.61009979248047, 30.031497955322266, 18.194942474365234, 78.78274536132812, 51.02257537841797, 14.046260833740234, -8.57720947265625, 81.38690185546875, 67.41952514648438, 4.107879638671875, 14.2308349609375, 18.99187469482422, 0.666107177734375, -6.754791259765625, 20.29174041748047, 78.54473876953125, -15.97540283203125, 98.61836242675781, 11.3203125, 85.21717071533203, 45.642730712890625, 14.401512145996094, 55.71675109863281, -23.267005920410156, -30.444854736328125, -11.895584106445312, 62.326904296875, -24.381248474121094, 18.540624618530273, 74.49668884277344, 9.685325622558594, 16.621795654296875, 5.107107162475586, -0.7332000732421875, 28.615310668945312, 13.357521057128906, 29.175800323486328, -0.03758811950683594, 44.86658477783203, 17.034427642822266, 24.774978637695312, 24.311573028564453, 51.57044982910156, 82.13546752929688, 20.46929931640625, 55.237396240234375, 5.330371856689453, -21.084075927734375, -3.3769397735595703, 3.9097366333007812, -27.64837646484375, 61.8702392578125, 9.923606872558594, 46.18812561035156, 28.64183807373047, 28.624221801757812, -10.526557922363281, 8.711566925048828, 20.22026824951172, 39.167945861816406, 47.157569885253906, 19.40673828125, 31.704246520996094, 2.4178314208984375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000083.npy"}
{"epoch": 0.12187958883994127, "step": 84, "batch_size": 64, "mean": 20.55937385559082, "std": 23.165847778320312, "min": -50.8099365234375, "p10": -2.288106346130371, "median": 18.0060396194458, "p90": 47.43138122558594, "max": 95.41282653808594, "pos_frac": 0.84375, "sample": [6.756416320800781, 17.763717651367188, 3.681671142578125, 6.077726364135742, 52.176422119140625, 6.81561279296875, 6.162834167480469, 23.158355712890625, 18.248361587524414, 8.057914733886719, 32.7152214050293, 40.53923797607422, -22.67388916015625, 13.682815551757812, 23.646404266357422, 32.529693603515625, 11.513561248779297, 95.41282653808594, 33.33528518676758, 1.7821044921875, 7.073486328125, 31.44622802734375, 5.325431823730469, 7.078250885009766, -1.0532455444335938, 46.78143310546875, 29.524032592773438, 32.55995178222656, -50.8099365234375, -2.3192577362060547, 58.41309356689453, 3.0066909790039062, 41.774261474609375, 16.723793029785156, 22.191734313964844, 12.689437866210938, 55.316436767578125, 47.709930419921875, 35.831298828125, 21.27924346923828, 3.4737930297851562, 33.93890380859375, -5.254873275756836, 17.73990249633789, 29.88946533203125, 31.747425079345703, -2.2154197692871094, -4.898918151855469, 37.20512390136719, 35.45564270019531, -4.689064025878906, 46.488319396972656, 63.19398498535156, 20.28081512451172, 66.32708740234375, 16.483062744140625, 34.52484893798828, 9.765899658203125, -20.89160919189453, 33.346954345703125, 13.475130081176758, 21.363311767578125, 7.333950042724609, -0.20843505859375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000084.npy"}
{"epoch": 0.12334801762114538, "step": 85, "batch_size": 64, "mean": 21.296226501464844, "std": 27.59166145324707, "min": -35.4468994140625, "p10": -6.949440002441405, "median": 19.512788772583008, "p90": 64.70830001831057, "max": 78.84951782226562, "pos_frac": 0.734375, "sample": [47.62705993652344, 17.1937255859375, 31.545997619628906, 4.072689056396484, -5.0789337158203125, 23.715774536132812, 32.14320373535156, 1.450164794921875, 66.33228302001953, 10.241729736328125, -2.6150951385498047, 6.4087677001953125, 74.81614685058594, 30.16002655029297, 27.56916618347168, 22.119674682617188, -6.19390869140625, -20.88141632080078, 21.109580993652344, 0.8408279418945312, -35.4468994140625, 78.84951782226562, 21.791000366210938, -2.736734390258789, -7.2318115234375, 48.801612854003906, 7.3379058837890625, -12.364486694335938, 20.174232482910156, 10.663101196289062, 54.03218078613281, -6.2905731201171875, -19.90655517578125, -1.6622467041015625, 15.632858276367188, 5.534292221069336, 69.64387512207031, 18.85134506225586, 42.83589172363281, 0.03725242614746094, -5.553901672363281, 74.31680297851562, -2.406238555908203, 49.951507568359375, 68.01596069335938, 16.66582489013672, 60.91900634765625, 74.57249450683594, 12.475494384765625, 6.057708740234375, 37.25237274169922, 55.36443328857422, -0.94793701171875, -9.111774444580078, 20.5380859375, 20.6778621673584, 56.47596740722656, 23.456993103027344, 32.37937927246094, 41.54111862182617, -25.935256958007812, 20.446277618408203, -1.286367416381836, 45.969573974609375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000085.npy"}
{"epoch": 0.12481644640234948, "step": 86, "batch_size": 64, "mean": 22.428207397460938, "std": 42.78868865966797, "min": -65.83668518066406, "p10": -16.842080307006835, "median": 15.845415115356445, "p90": 73.47634887695314, "max": 165.56829833984375, "pos_frac": 0.71875, "sample": [-14.679283142089844, 20.809627532958984, 76.20387268066406, -23.774017333984375, 38.08345031738281, 34.184783935546875, 29.290388107299805, -1.3202056884765625, 119.04421997070312, 12.408279418945312, 6.853250503540039, -8.362815856933594, -0.6160659790039062, 24.442123413085938, -12.050918579101562, 22.11859893798828, 55.002105712890625, 19.306800842285156, -3.678028106689453, -62.384033203125, 69.41708374023438, 24.546680450439453, 33.81912612915039, 11.5257568359375, 10.205902099609375, -17.768993377685547, -6.877754211425781, 75.21603393554688, -5.128303527832031, 8.859115600585938, 41.3718147277832, 19.613265991210938, 8.632635116577148, -65.83668518066406, 51.827423095703125, 59.386871337890625, 30.557647705078125, -47.34815216064453, 83.66508483886719, 2.5351810455322266, 165.56829833984375, 19.282550811767578, 24.910202026367188, 12.27542495727539, -5.648067474365234, 142.7860107421875, 7.309497833251953, 136.21688842773438, 52.57685852050781, -10.701629638671875, -35.22814178466797, 42.87446594238281, 10.33232307434082, 36.92150115966797, 32.48512268066406, 37.635860443115234, 38.98225402832031, 6.85188102722168, 0.45754241943359375, 3.3063297271728516, -5.017936706542969, -22.262191772460938, 2.5782318115234375, 21.810138702392578], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000086.npy"}
{"epoch": 0.1262848751835536, "step": 87, "batch_size": 64, "mean": 26.637374877929688, "std": 37.68668746948242, "min": -89.42831420898438, "p10": -10.53235168457031, "median": 22.264020919799805, "p90": 75.25146331787111, "max": 141.9814910888672, "pos_frac": 0.8125, "sample": [24.258169174194336, 39.92706298828125, 0.23342323303222656, 10.664628982543945, 39.994384765625, -11.531768798828125, 35.1524658203125, -0.6251697540283203, 0.7331619262695312, -6.875396728515625, -26.853424072265625, 53.49298095703125, -7.709831237792969, 77.50465393066406, 36.416954040527344, 67.77558898925781, 58.54316711425781, 11.414205551147461, 100.62521362304688, 69.9940185546875, 22.983688354492188, 22.709014892578125, 8.46108627319336, 45.85528564453125, 43.51502990722656, 2.045248031616211, 43.144989013671875, 141.9814910888672, 13.878768920898438, 4.20904541015625, 13.365617752075195, 83.6256103515625, 64.89407348632812, 11.548763275146484, 63.63861083984375, 29.666885375976562, 15.628633499145508, 42.5848503112793, 21.819026947021484, 86.80683898925781, 0.7566947937011719, 90.84440612792969, 11.854606628417969, 20.861328125, -89.42831420898438, 13.089559555053711, 44.241844177246094, -36.391876220703125, 42.71147155761719, 79.8070068359375, -19.062042236328125, -20.13379669189453, 9.840251922607422, 21.740234375, -8.20037841796875, 66.04190063476562, 30.015899658203125, 38.62577819824219, -44.398681640625, 27.562828063964844, 15.745651245117188, 0.9772052764892578, 57.788818359375, -5.595455169677734], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000087.npy"}
{"epoch": 0.1277533039647577, "step": 88, "batch_size": 64, "mean": 20.091384887695312, "std": 36.25615310668945, "min": -69.35037231445312, "p10": -12.693922424316405, "median": 12.286921501159668, "p90": 65.30374145507815, "max": 148.13958740234375, "pos_frac": 0.765625, "sample": [9.946159362792969, 10.648330688476562, 70.46185302734375, 72.5057373046875, 36.93513488769531, 55.25697326660156, 4.332036972045898, 68.65145874023438, -13.458831787109375, 14.520771026611328, 38.15641784667969, 8.045547485351562, 2.7000885009765625, 53.00782775878906, -25.919601440429688, 35.275390625, -21.331939697265625, 15.326858520507812, -54.14701843261719, -69.35037231445312, -6.196323394775391, 83.74502563476562, 44.68250274658203, -4.386161804199219, 45.868247985839844, 5.215147018432617, 33.50115966796875, -10.909133911132812, 24.932226181030273, 22.046920776367188, 44.82292175292969, 8.127113342285156, -2.1454219818115234, 28.68744659423828, 1.8860511779785156, 108.01510620117188, 6.617912292480469, 57.492401123046875, 5.067451477050781, 148.13958740234375, 7.16217041015625, 49.21003723144531, 3.465147018432617, 14.577375411987305, 12.109420776367188, 27.232833862304688, 0.17069244384765625, -3.3555221557617188, 39.10546875, 88.6439208984375, 37.21331024169922, 16.471405029296875, -1.3479042053222656, 11.279045104980469, -57.778961181640625, -5.065582275390625, 11.456613540649414, -1.3954200744628906, 12.464422225952148, 15.719873428344727, 19.861677169799805, -22.910858154296875, 7.7480010986328125, 47.03852844238281], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000088.npy"}
{"epoch": 0.12922173274596183, "step": 89, "batch_size": 64, "mean": 22.48392105102539, "std": 30.76645278930664, "min": -75.01654052734375, "p10": -2.4744611740112306, "median": 18.115185737609863, "p90": 54.07527542114259, "max": 138.7556610107422, "pos_frac": 0.84375, "sample": [7.69573974609375, 1.3352794647216797, 6.163980484008789, 22.362051010131836, -11.116024017333984, 55.65557861328125, 47.04851531982422, 17.695186614990234, -9.867012023925781, 62.118690490722656, 64.16210174560547, 25.187435150146484, 23.901290893554688, 50.387901306152344, 41.48712921142578, 30.901874542236328, 13.741226196289062, 10.309200286865234, -2.501932144165039, 20.57364273071289, 16.574119567871094, 4.189741134643555, 17.218286514282227, 12.689990997314453, -11.219650268554688, 1.3257789611816406, 1.5105056762695312, 10.402444839477539, 84.6602783203125, 21.69414520263672, 31.980857849121094, 94.63427734375, 17.759685516357422, 47.900299072265625, 23.69556427001953, -1.02532958984375, 12.95564079284668, 138.7556610107422, 6.198925018310547, -2.8068504333496094, 32.203914642333984, -2.4103622436523438, 25.405492782592773, 33.13923263549805, 16.455400466918945, 5.771553039550781, 10.203510284423828, 4.620983123779297, 2.9521942138671875, 39.39788818359375, 23.905494689941406, 45.214324951171875, -2.329235076904297, 20.246810913085938, 48.47521209716797, -75.01654052734375, -29.123870849609375, 28.92357635498047, 18.470685958862305, 26.455612182617188, 26.082931518554688, 25.411033630371094, 101.18878173828125, 6.990074157714844], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000089.npy"}
{"epoch": 0.13069016152716592, "step": 90, "batch_size": 64, "mean": 30.72348403930664, "std": 46.04564666748047, "min": -58.90483093261719, "p10": -12.972170257568353, "median": 19.798625946044922, "p90": 102.45020904541018, "max": 159.91595458984375, "pos_frac": 0.796875, "sample": [76.85205078125, -48.61237335205078, 31.306427001953125, 104.27630615234375, 23.254966735839844, 38.55302429199219, 1.0359382629394531, -5.538852691650391, 61.756805419921875, -2.1270198822021484, 49.234405517578125, 19.00055694580078, -1.2820396423339844, 12.83245849609375, 42.677215576171875, 93.96511840820312, -5.1832122802734375, 159.91595458984375, 58.19456481933594, 1.1377792358398438, 2.051839828491211, 11.935218811035156, 22.536361694335938, -16.05628204345703, 129.2547607421875, 16.857234954833984, 27.661094665527344, 133.98138427734375, 4.998699188232422, 24.457279205322266, 150.7833251953125, 25.107666015625, -33.009857177734375, 12.79196548461914, -5.775909423828125, -58.90483093261719, 29.31045913696289, 8.444107055664062, 69.20475769042969, 6.797935485839844, 16.89875030517578, 13.85430908203125, -22.751914978027344, -5.563741683959961, 24.694595336914062, 98.18931579589844, 108.7762451171875, -39.103355407714844, -36.28334045410156, 12.377799987792969, 39.955360412597656, 1.5835342407226562, 19.811222076416016, 41.81957244873047, 74.9671630859375, 36.60063934326172, 10.98747444152832, 12.263833999633789, 68.59786987304688, 19.786029815673828, 25.734817504882812, 19.159517288208008, 105.85101318359375, 44.419036865234375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000090.npy"}
{"epoch": 0.13215859030837004, "step": 91, "batch_size": 64, "mean": 24.99896240234375, "std": 37.75916290283203, "min": -37.66044616699219, "p10": -23.000447082519532, "median": 23.13644790649414, "p90": 76.68040466308598, "max": 135.6436767578125, "pos_frac": 0.75, "sample": [20.06237030029297, 56.07984924316406, -37.66044616699219, 9.24832534790039, 81.53828430175781, 0.9243431091308594, -23.329917907714844, 47.10317611694336, 4.051843643188477, 22.89691925048828, 46.53594207763672, 43.47022247314453, 46.89258575439453, 14.2916259765625, -36.19452667236328, 119.86308288574219, -12.206527709960938, 2.5678024291992188, 9.16278076171875, 91.7857666015625, 9.146598815917969, 35.53181076049805, 135.6436767578125, -19.55413818359375, 28.377418518066406, 59.701881408691406, -23.060134887695312, 3.911245346069336, 24.847061157226562, -6.6627197265625, -22.861175537109375, 23.980356216430664, 65.34535217285156, -23.593795776367188, 19.242156982421875, 24.01580810546875, 61.69458770751953, -14.432197570800781, 22.153533935546875, -4.928506851196289, 10.202621459960938, 23.3759765625, 63.55968475341797, 3.2601547241210938, -21.642959594726562, 42.385833740234375, 24.929046630859375, 84.54759216308594, 24.955795288085938, 45.802940368652344, 9.432327270507812, -28.316299438476562, 0.4082317352294922, 31.673324584960938, -2.6318283081054688, 43.61939239501953, -0.6603469848632812, 91.63739013671875, 49.46967315673828, 27.1676025390625, 58.497772216796875, -31.254539489746094, 52.817047119140625, 91.11478424072266], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000091.npy"}
{"epoch": 0.13362701908957417, "step": 92, "batch_size": 64, "mean": 25.541980743408203, "std": 36.66453170776367, "min": -51.923431396484375, "p10": -13.076917266845703, "median": 18.88584613800049, "p90": 64.91728668212892, "max": 138.87640380859375, "pos_frac": 0.75, "sample": [-5.491306304931641, 6.553840637207031, 7.0913848876953125, 52.204261779785156, -8.223987579345703, 6.355796813964844, -51.923431396484375, -17.202192306518555, 54.066650390625, 25.27123260498047, 72.24003601074219, 0.3371410369873047, 48.479400634765625, 3.094930648803711, -6.8103179931640625, 120.75729370117188, 57.448326110839844, 138.87640380859375, -46.758575439453125, -5.3006744384765625, 27.283832550048828, -11.973548889160156, 29.587631225585938, -9.88821792602539, -19.74437713623047, 63.79096984863281, -13.549789428710938, 77.1376953125, 32.15485382080078, 4.348894119262695, 77.81212615966797, -2.752023696899414, 36.325843811035156, 0.6383647918701172, 54.73961639404297, 9.391761779785156, 36.732086181640625, 118.30320739746094, 10.268129348754883, 62.12855529785156, 47.332130432128906, 65.39999389648438, -1.6764907836914062, -6.2979278564453125, 9.742298126220703, 27.743480682373047, 8.445877075195312, 43.64665222167969, 37.495933532714844, 11.940071105957031, 0.009380340576171875, -13.574352264404297, 54.90153121948242, 17.696447372436523, 20.075244903564453, 13.344749450683594, 39.20812225341797, 56.74986267089844, 47.7032470703125, 27.01324462890625, 12.558782577514648, -13.985588073730469, 59.411407470703125, 36.000885009765625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000092.npy"}
{"epoch": 0.13509544787077826, "step": 93, "batch_size": 64, "mean": 25.397796630859375, "std": 32.590415954589844, "min": -41.47327423095703, "p10": -5.812244987487792, "median": 18.034427642822266, "p90": 61.655249023437506, "max": 163.17428588867188, "pos_frac": 0.8125, "sample": [20.70099639892578, 9.516204833984375, 28.224319458007812, 67.11223602294922, 61.91570281982422, 60.047119140625, 4.6855926513671875, 24.23735809326172, 49.63355255126953, -16.73291778564453, 83.07192993164062, 22.882064819335938, 57.640777587890625, 54.10871124267578, 163.17428588867188, 14.268810272216797, -6.1972808837890625, 2.6200714111328125, 0.8910369873046875, 96.01190948486328, -9.252704620361328, 14.401958465576172, 22.692123413085938, -0.1682281494140625, 36.117218017578125, 10.755203247070312, 37.390892028808594, 5.709403991699219, -1.1397113800048828, 11.375350952148438, -0.6852207183837891, 14.805572509765625, 9.37844467163086, 20.46314239501953, 34.945823669433594, 58.039154052734375, 71.15052032470703, -25.556289672851562, -4.913827896118164, 55.30128479003906, 49.811058044433594, 37.25907516479492, 41.064125061035156, 7.529500961303711, 10.564020156860352, 9.988700866699219, 31.810577392578125, 2.7706737518310547, 61.047523498535156, 0.11655044555664062, -9.467262268066406, 63.343292236328125, 4.319000244140625, 18.562034606933594, 6.66398811340332, 5.9984893798828125, -19.823097229003906, 32.16291046142578, 49.727806091308594, 36.12809753417969, 51.26493835449219, 17.506820678710938, -41.47327423095703, -0.03908729553222656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000093.npy"}
{"epoch": 0.13656387665198239, "step": 94, "batch_size": 64, "mean": 29.349872589111328, "std": 34.11307144165039, "min": -23.55682373046875, "p10": -8.930377769470214, "median": 24.6575984954834, "p90": 73.90499191284181, "max": 125.672607421875, "pos_frac": 0.78125, "sample": [27.858718872070312, -19.655723571777344, 47.049095153808594, 20.490982055664062, 6.0628509521484375, -16.227725982666016, 10.907894134521484, 63.1135368347168, 5.536172866821289, -23.55682373046875, 24.855554580688477, 12.110122680664062, 27.316307067871094, 6.908058166503906, 12.276565551757812, 0.21504592895507812, 63.72013854980469, 12.406646728515625, -8.025260925292969, 125.672607421875, 5.472393035888672, 33.419010162353516, 33.89137268066406, 56.86390686035156, 87.10794067382812, 33.35235595703125, 13.660491943359375, 32.226646423339844, 23.785762786865234, -2.8411178588867188, -5.736555099487305, 18.45445442199707, 39.506866455078125, 67.74905395507812, 45.42250442504883, 21.267658233642578, 37.36619567871094, 74.99280548095703, 35.03046798706055, 24.45964241027832, 5.791521072387695, 47.69843292236328, -5.059638977050781, -19.82025909423828, 71.36676025390625, 80.81817626953125, -8.470598220825195, 61.958282470703125, 122.29702758789062, -9.127426147460938, -23.537689208984375, 92.65966796875, 35.05509948730469, 63.03221130371094, -3.5028839111328125, 21.658172607421875, -15.950820922851562, -6.445545196533203, 39.88256072998047, 16.985198974609375, 38.05537033081055, 61.0977783203125, 78.7071533203125, 58.75663757324219], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000094.npy"}
{"epoch": 0.13803230543318648, "step": 95, "batch_size": 64, "mean": 23.649900436401367, "std": 32.5994873046875, "min": -83.02732849121094, "p10": -8.270219421386717, "median": 17.425247192382812, "p90": 70.4343444824219, "max": 100.55891418457031, "pos_frac": 0.8125, "sample": [18.053436279296875, 47.206016540527344, 4.886116027832031, 25.85332489013672, 21.24076271057129, 73.43679809570312, 9.26162338256836, 6.885896682739258, -8.953445434570312, 14.771049499511719, 19.329862594604492, 4.169887542724609, 37.829952239990234, 89.45744323730469, 53.460052490234375, -6.676025390625, 39.997108459472656, 3.3383922576904297, 16.56963348388672, 50.77776336669922, 26.634864807128906, 43.21061706542969, 20.334989547729492, 16.057512283325195, 10.545743942260742, -14.003204345703125, 11.378753662109375, 60.338218688964844, 22.520015716552734, 31.095932006835938, -10.950302124023438, -10.097518920898438, 56.84197998046875, 100.55891418457031, 81.65438842773438, 63.428619384765625, 41.25856018066406, 1.2623043060302734, -0.3427276611328125, 75.05624389648438, 48.514102935791016, 89.565185546875, 77.09774780273438, 3.230560302734375, 15.452735900878906, 9.481014251708984, -1.6535511016845703, 34.29786682128906, 33.98834991455078, -83.02732849121094, 2.631956100463867, 2.0633182525634766, -34.150482177734375, 51.992279052734375, 45.961280822753906, 13.816146850585938, 44.03224182128906, 12.307628631591797, 17.888259887695312, 16.962234497070312, -6.249704360961914, -3.407827377319336, 15.632732391357422, -40.512672424316406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000095.npy"}
{"epoch": 0.1395007342143906, "step": 96, "batch_size": 64, "mean": 32.10420227050781, "std": 39.126991271972656, "min": -60.42955017089844, "p10": -0.7142311096191406, "median": 19.487199783325195, "p90": 82.0933937072754, "max": 170.0359344482422, "pos_frac": 0.875, "sample": [68.81478881835938, 18.004562377929688, 39.27500915527344, 33.26181411743164, 19.142959594726562, 71.44964599609375, 6.949798583984375, 3.013151168823242, 0.9721889495849609, 24.38140106201172, 122.01612854003906, 17.620361328125, 8.355430603027344, -6.299762725830078, 3.6003761291503906, 22.211673736572266, 30.614547729492188, 47.122825622558594, 90.88347625732422, 98.50131225585938, 10.393447875976562, 47.61894226074219, 39.59795379638672, 170.0359344482422, 80.1535415649414, 13.20468521118164, 19.831439971923828, 9.767143249511719, 17.78533935546875, 16.446441650390625, 9.7255859375, 146.29208374023438, 34.57942199707031, 29.455368041992188, 6.738395690917969, 93.41818237304688, 63.11360168457031, 6.325828552246094, 39.70307922363281, 2.3300209045410156, 51.54865264892578, 27.070261001586914, 24.66259765625, -12.012687683105469, 15.926261901855469, 1.1022529602050781, -3.9665069580078125, 43.53654479980469, -1.6803951263427734, 26.89063262939453, 12.811965942382812, 64.73602294921875, -15.650764465332031, 7.500244140625, 75.76519775390625, 13.190597534179688, 18.97508430480957, 71.16450500488281, 82.92475891113281, -0.728790283203125, -0.6802597045898438, 33.8609619140625, -60.42955017089844, 1.743255615234375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000096.npy"}
{"epoch": 0.14096916299559473, "step": 97, "batch_size": 64, "mean": 24.823837280273438, "std": 30.394813537597656, "min": -65.14102935791016, "p10": -8.63835411071777, "median": 24.02078628540039, "p90": 62.38342514038087, "max": 115.74588012695312, "pos_frac": 0.75, "sample": [18.856399536132812, 57.12957763671875, -0.491790771484375, 1.0914230346679688, -5.774467468261719, -4.6615142822265625, 17.230175018310547, 30.00482749938965, 2.605236053466797, 50.207000732421875, 25.925800323486328, 45.5655517578125, -0.9894866943359375, 19.219139099121094, 83.840576171875, 68.55268859863281, -25.026397705078125, 40.039642333984375, -0.19245147705078125, 14.324174880981445, 27.434913635253906, 6.796451568603516, 17.771514892578125, -22.372085571289062, 77.83111572265625, 19.888566970825195, 27.021209716796875, 31.763137817382812, 44.684478759765625, 15.371341705322266, 30.160377502441406, -6.456085205078125, 17.80450439453125, -13.862947463989258, 53.782188415527344, 46.70306396484375, 52.67576599121094, 32.235355377197266, -65.14102935791016, -10.76123046875, 53.32361602783203, 33.7579460144043, -10.549942016601562, 9.020751953125, 58.71562957763672, 36.29826736450195, 0.48937416076660156, 22.56854820251465, 25.473024368286133, 19.48639678955078, -5.395711898803711, 63.95533752441406, -9.573612213134766, 54.478328704833984, -3.548870086669922, 69.02204895019531, -2.547607421875, 42.98081970214844, 50.715484619140625, 68.39588928222656, 34.84125518798828, 10.842388153076172, 115.74588012695312, 29.443649291992188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000097.npy"}
{"epoch": 0.14243759177679882, "step": 98, "batch_size": 64, "mean": 28.452003479003906, "std": 40.56512451171875, "min": -114.55239868164062, "p10": -14.991331863403317, "median": 31.754433631896973, "p90": 71.99261703491212, "max": 141.97384643554688, "pos_frac": 0.796875, "sample": [43.54747772216797, 6.069786071777344, 68.25399780273438, 48.7330322265625, 6.910793304443359, 22.092477798461914, 26.369354248046875, 27.506019592285156, 33.20704650878906, 8.141487121582031, 31.483400344848633, -11.287078857421875, 21.789779663085938, 27.37773895263672, 71.58219909667969, 54.392120361328125, -16.578868865966797, 0.48504638671875, -2.0380821228027344, -24.45391845703125, 42.73611831665039, 38.13835906982422, -33.50444030761719, -44.156005859375, 60.13373565673828, 57.98918533325195, 32.12078857421875, 72.16851043701172, 76.04206848144531, 27.907527923583984, 42.86399841308594, -3.520397186279297, 68.0179214477539, 141.97384643554688, -114.55239868164062, 90.59556579589844, 63.9459228515625, 32.837493896484375, 49.54823303222656, 20.223068237304688, -11.074909210205078, 42.436553955078125, -61.65861511230469, 28.75897216796875, 38.23291015625, 4.11309814453125, 41.157386779785156, 87.63601684570312, 65.11663818359375, 39.81875991821289, 17.106857299804688, 13.128013610839844, 84.31732177734375, 113.876953125, 36.74486541748047, 35.661434173583984, 32.02546691894531, 15.65568733215332, 5.626155853271484, -10.000541687011719, 57.736114501953125, -4.255731582641602, 3.8233394622802734, -20.14733123779297], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000098.npy"}
{"epoch": 0.14390602055800295, "step": 99, "batch_size": 64, "mean": 35.01633834838867, "std": 44.858585357666016, "min": -37.27013397216797, "p10": -8.910260009765624, "median": 23.246341705322266, "p90": 91.89490051269532, "max": 194.51637268066406, "pos_frac": 0.828125, "sample": [-1.723794937133789, -9.683029174804688, 6.330234527587891, 22.786373138427734, 1.5654335021972656, 45.053306579589844, 30.37053680419922, 32.91920471191406, 116.68359375, 22.639448165893555, 25.707054138183594, 134.01922607421875, 5.954044342041016, -33.855560302734375, 92.74238586425781, -7.1071319580078125, 23.47479248046875, 39.97455596923828, -0.8589897155761719, 18.87580108642578, 83.93966674804688, 15.860992431640625, 12.88884162902832, 168.43954467773438, 10.395509719848633, 65.8941650390625, 1.3657760620117188, -28.46454620361328, 11.33265495300293, -37.27013397216797, -10.776542663574219, 56.493736267089844, 41.45927429199219, 29.744319915771484, 68.92237854003906, 194.51637268066406, 80.11601257324219, 41.97346878051758, 18.085952758789062, 22.161293029785156, 6.466339111328125, 32.0179328918457, 41.96220397949219, 89.91743469238281, 107.0305404663086, 8.028709411621094, 13.46038818359375, 71.11167907714844, -5.841270446777344, 0.18380165100097656, 34.92658996582031, 96.57254028320312, 32.6057014465332, 12.13033676147461, 62.28187561035156, -12.210342407226562, 44.53529739379883, 15.47410774230957, 71.82392883300781, 53.41859436035156, 23.01789093017578, 49.58866882324219, 4.681468963623047, -25.084976196289062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000099.npy"}
{"epoch": 0.14537444933920704, "step": 100, "batch_size": 64, "mean": 20.08426856994629, "std": 38.81177520751953, "min": -84.12049865722656, "p10": -21.2537519454956, "median": 19.305377960205078, "p90": 67.93270492553711, "max": 155.3404541015625, "pos_frac": 0.765625, "sample": [56.384483337402344, 33.03289031982422, -6.872718811035156, 15.72842025756836, 24.408977508544922, -10.739997863769531, 34.53826904296875, 28.155181884765625, 7.97027587890625, 51.53312683105469, 10.674880981445312, 0.9889545440673828, 41.597412109375, 26.658935546875, 4.660673141479492, 83.031494140625, 19.58869171142578, -46.90472412109375, -37.05195617675781, -59.802459716796875, -15.544244766235352, -12.311119079589844, 0.34920692443847656, -67.875244140625, 59.00148391723633, 19.022064208984375, 51.959327697753906, 24.881668090820312, 1.2319908142089844, 6.354394912719727, 46.329315185546875, -2.9914779663085938, 45.86326599121094, 5.712207794189453, 70.82443237304688, 32.79876708984375, 9.551164627075195, 12.076347351074219, 67.61817169189453, 27.450820922851562, 42.96134567260742, 16.851295471191406, 7.44610595703125, 0.29856300354003906, 32.959190368652344, 35.848175048828125, -10.8533935546875, -27.785545349121094, -7.886932373046875, 71.7114028930664, 39.460906982421875, 74.29886627197266, 26.079214096069336, -84.12049865722656, 155.3404541015625, -12.85035514831543, 18.387786865234375, 85.21116638183594, -23.70068359375, 68.0675048828125, 49.1949462890625, 7.404335021972656, 39.68120574951172, 21.50473403930664], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000100.npy"}
{"epoch": 0.14684287812041116, "step": 101, "batch_size": 64, "mean": 24.696561813354492, "std": 43.25938415527344, "min": -56.904457092285156, "p10": -22.171418571472163, "median": 19.606212615966797, "p90": 91.21191711425789, "max": 134.88845825195312, "pos_frac": 0.671875, "sample": [105.65388488769531, 134.88845825195312, 31.408241271972656, 31.836837768554688, -16.56511688232422, 112.51238250732422, 114.338134765625, 11.354804992675781, 98.95991516113281, 117.23884582519531, -39.197174072265625, 31.162643432617188, -40.47852325439453, 51.12248992919922, 14.020477294921875, -9.944129943847656, -24.311676025390625, 15.131322860717773, 67.53751373291016, 28.553550720214844, -16.220558166503906, 27.478729248046875, 55.72023010253906, 67.5534896850586, 19.858932495117188, 19.353492736816406, 9.79427719116211, 68.0777816772461, 126.6334228515625, 28.157241821289062, -7.489013671875, -6.6545867919921875, -14.341249465942383, 54.54005813598633, 19.267763137817383, 40.48750305175781, 73.13325500488281, 49.12657165527344, -6.2093505859375, 6.391117095947266, -2.6573829650878906, -16.769756317138672, 21.321205139160156, -7.868782043457031, -0.907318115234375, -17.1774845123291, 38.392860412597656, 34.46506118774414, 9.08620834350586, 26.92731285095215, -56.904457092285156, 35.504432678222656, 0.8105621337890625, -33.76046371459961, 26.783527374267578, 18.61690902709961, -16.98529052734375, 41.17561721801758, 7.411338806152344, -2.1711463928222656, 53.514732360839844, -30.004215240478516, 41.50639343261719, -39.61189270019531], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000101.npy"}
{"epoch": 0.14831130690161526, "step": 102, "batch_size": 64, "mean": 18.196969985961914, "std": 40.89890670776367, "min": -84.14306640625, "p10": -23.600476074218747, "median": 11.528858184814453, "p90": 60.53031311035157, "max": 146.51211547851562, "pos_frac": 0.65625, "sample": [3.259420394897461, 40.37550354003906, 61.09978485107422, 12.447303771972656, 2.58209228515625, 6.663726806640625, -13.435089111328125, -46.48736572265625, 23.606748580932617, -84.14306640625, 29.782012939453125, 92.41644287109375, -13.249397277832031, 39.508583068847656, 30.987804412841797, -9.125797271728516, 10.61041259765625, 35.04963684082031, 52.954444885253906, 1.6663322448730469, -18.06103515625, -4.8190460205078125, 50.836334228515625, -6.060033798217773, 62.10429382324219, 24.324337005615234, 40.43256378173828, 146.51211547851562, -4.085609436035156, -0.29106903076171875, 115.28553009033203, -0.5192584991455078, 39.15400314331055, 44.091331481933594, 4.101673126220703, -17.788330078125, 34.6007080078125, -84.11947631835938, 2.760040283203125, -30.054969787597656, 0.7585334777832031, -47.904815673828125, -24.443145751953125, -4.253082275390625, 59.20154571533203, 34.72882080078125, 46.321815490722656, 85.27516174316406, 46.657127380371094, 6.46832275390625, 73.8158187866211, -29.149505615234375, -7.311790466308594, 5.703521728515625, 55.19512176513672, -21.634246826171875, 43.521942138671875, 36.91368865966797, 26.74867820739746, 38.19139099121094, -12.841773986816406, -16.727455139160156, 43.91649627685547, 50.480281829833984], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000102.npy"}
{"epoch": 0.14977973568281938, "step": 103, "batch_size": 64, "mean": 36.2520751953125, "std": 45.051151275634766, "min": -41.428009033203125, "p10": -11.638389587402344, "median": 28.791893005371094, "p90": 87.49871673583985, "max": 193.0517578125, "pos_frac": 0.796875, "sample": [-5.941593170166016, 8.744537353515625, 14.22584342956543, 35.533531188964844, -23.377090454101562, -27.34638214111328, 35.01869583129883, 40.8385009765625, 15.181631088256836, 20.723243713378906, 28.821449279785156, 46.15137481689453, 17.46143341064453, 18.82061004638672, -1.3925857543945312, 69.62255859375, 115.56558227539062, 37.107826232910156, 53.0743408203125, 66.2604751586914, -38.61279296875, -11.964111328125, 51.80946350097656, 25.910186767578125, -41.428009033203125, 78.72953033447266, 37.140628814697266, -17.93555450439453, 15.795783996582031, 18.08935546875, 41.348716735839844, 176.39669799804688, 28.76233673095703, 18.958911895751953, 74.99700927734375, 73.84854888916016, -6.3455657958984375, -10.878372192382812, 21.3214111328125, 17.5960636138916, -6.0106964111328125, 42.37016296386719, 69.05827331542969, -33.446990966796875, 24.010604858398438, 102.9045181274414, 85.11813354492188, 27.351181030273438, 193.0517578125, 105.06190490722656, 108.6321029663086, 88.51896667480469, 13.084190368652344, 42.625396728515625, 70.83716583251953, -1.6802864074707031, 32.90801239013672, 3.2486515045166016, 3.355388641357422, 51.152252197265625, 82.95421600341797, 16.906618118286133, 32.775054931640625, 46.71197509765625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000103.npy"}
{"epoch": 0.1512481644640235, "step": 104, "batch_size": 64, "mean": 34.55239486694336, "std": 50.890193939208984, "min": -67.96299743652344, "p10": -12.59286823272705, "median": 19.760730743408203, "p90": 98.19711608886719, "max": 201.8740234375, "pos_frac": 0.75, "sample": [99.16339111328125, -8.486968994140625, 14.332412719726562, 16.16651153564453, 65.41044616699219, -6.652137756347656, -30.770225524902344, 3.2723636627197266, -4.7403106689453125, 40.305152893066406, -13.391168594360352, 37.86220932006836, 156.98190307617188, 54.25273132324219, 90.85489654541016, -3.3317642211914062, 95.94247436523438, 33.00328826904297, -19.558189392089844, 121.5286865234375, 42.07466125488281, 3.7435760498046875, 19.94680404663086, 7.7200927734375, 14.709104537963867, 2.256143569946289, 201.8740234375, -46.94939422607422, 59.50214385986328, 141.0787353515625, 142.8096923828125, 32.64134216308594, 13.969522476196289, 68.87818908691406, 12.444995880126953, 15.04482650756836, -10.730167388916016, -23.35485076904297, 123.23391723632812, 16.49877166748047, 53.32251739501953, -8.57489013671875, 73.27130126953125, -5.695037841796875, 88.02351379394531, 38.72679901123047, 33.366798400878906, 52.8955192565918, 41.502105712890625, 34.68196105957031, 4.673854827880859, 94.95497131347656, 7.737770080566406, 30.122650146484375, 64.40379333496094, 70.9460220336914, -8.39698600769043, 28.9174747467041, -0.47677040100097656, 19.574657440185547, -24.585357666015625, 1.7298259735107422, -67.96299743652344, 8.655960083007812], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000104.npy"}
{"epoch": 0.1527165932452276, "step": 105, "batch_size": 64, "mean": 49.002593994140625, "std": 44.50178146362305, "min": -75.71536254882812, "p10": 2.4654232025146516, "median": 47.29097366333008, "p90": 107.16758117675785, "max": 184.5460968017578, "pos_frac": 0.90625, "sample": [51.196197509765625, 34.25559997558594, 47.512908935546875, 29.922439575195312, 42.46739196777344, 47.762657165527344, 75.49072265625, 114.9111557006836, 8.61958122253418, -75.71536254882812, -7.7498779296875, 84.6580810546875, 86.925048828125, 13.650136947631836, 34.34480285644531, 52.57973861694336, -33.935546875, 1.1453781127929688, -6.463531494140625, 117.04976654052734, 34.66078186035156, 113.14778137207031, 71.54167175292969, 20.803733825683594, 12.008834838867188, 84.62618255615234, 66.55174255371094, 47.06903839111328, 9.673194885253906, 39.976356506347656, 71.51588439941406, 178.72445678710938, 21.738426208496094, 54.31724548339844, 18.79888916015625, 20.366439819335938, 74.79234313964844, 96.57958984375, 49.43147277832031, 15.68710708618164, 52.75584411621094, 97.35183715820312, 38.96582794189453, 40.78620147705078, 46.884124755859375, 54.08636474609375, -0.42075538635253906, 67.52494812011719, 73.51091003417969, 111.37432861328125, 38.223167419433594, 23.3497314453125, -21.057846069335938, 9.563995361328125, 69.40370178222656, 64.19126892089844, 58.60797119140625, 117.01571655273438, 64.0513916015625, 184.5460968017578, 27.057464599609375, 5.545528411865234, 13.662464141845703, 78.54725646972656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000105.npy"}
{"epoch": 0.15418502202643172, "step": 106, "batch_size": 64, "mean": 30.627548217773438, "std": 38.239112854003906, "min": -56.945640563964844, "p10": -12.966500091552728, "median": 30.867530822753906, "p90": 81.41138305664063, "max": 140.21588134765625, "pos_frac": 0.828125, "sample": [29.057235717773438, 5.228519439697266, 9.687765121459961, -4.516633987426758, 18.553142547607422, 2.3771419525146484, 50.99681854248047, -6.670051574707031, 32.92005920410156, 69.32966613769531, 37.93979263305664, 37.587303161621094, 2.3739795684814453, 78.51068115234375, 29.423179626464844, 98.63847351074219, 8.37055778503418, 12.146146774291992, 82.654541015625, 1.6649093627929688, 66.83889770507812, 44.6589469909668, 1.7529983520507812, 89.88137817382812, 25.4635066986084, 45.961769104003906, 6.842033386230469, -15.66497802734375, 57.344566345214844, -40.48820495605469, 49.899139404296875, 15.525741577148438, 43.59376907348633, 42.51033401489258, -3.069398880004883, 4.658660888671875, 24.589603424072266, 0.32804107666015625, 97.32173156738281, 65.44097137451172, 66.51661682128906, 25.59848403930664, -56.945640563964844, 4.746437072753906, 140.21588134765625, 40.27581787109375, 29.714134216308594, 35.18585205078125, 36.379150390625, 9.596343994140625, -0.12016868591308594, 61.471656799316406, 111.19432830810547, 35.992759704589844, 58.322174072265625, 60.13716125488281, -25.559249877929688, 35.247406005859375, -24.80963134765625, 32.02092742919922, 86.36906433105469, -43.92089080810547, -26.754791259765625, 49.626426696777344], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000106.npy"}
{"epoch": 0.15565345080763582, "step": 107, "batch_size": 64, "mean": 34.50071334838867, "std": 45.42250061035156, "min": -97.26492309570312, "p10": -13.207867813110342, "median": 23.786653518676758, "p90": 89.52843170166017, "max": 149.7453155517578, "pos_frac": 0.828125, "sample": [90.18618774414062, 117.58740234375, 49.21437072753906, 71.05782318115234, 25.922996520996094, 31.361412048339844, 54.63221740722656, -4.174526214599609, -97.26492309570312, 9.848302841186523, 17.319494247436523, 81.21923828125, -41.528053283691406, 27.586429595947266, 10.093877792358398, 7.077079772949219, 16.244945526123047, 87.99366760253906, 1.9976043701171875, -29.0184326171875, 111.61090850830078, 70.73657989501953, 82.58779907226562, 6.830141067504883, -18.159835815429688, 2.8361740112304688, -2.3890914916992188, 4.977119445800781, 149.7453155517578, 78.25179290771484, 19.889568328857422, 18.44805145263672, 103.29129791259766, 47.21299743652344, 68.9803466796875, 91.17378997802734, 16.93720817565918, 28.50841522216797, 16.862791061401367, -17.079299926757812, 65.44038391113281, -50.33201599121094, 67.93787384033203, 78.31584167480469, 11.509506225585938, 85.44108581542969, -0.21219635009765625, 86.32138061523438, 52.10856628417969, 57.57268524169922, 27.588462829589844, 22.714502334594727, 2.5491790771484375, -30.183570861816406, -0.6181411743164062, 5.406890869140625, 13.163505554199219, 80.12129211425781, 7.826751708984375, 24.85880470275879, 70.78082275390625, 107.0355224609375, 3.4488677978515625, 10.640243530273438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000107.npy"}
{"epoch": 0.15712187958883994, "step": 108, "batch_size": 64, "mean": 41.77477264404297, "std": 60.98763656616211, "min": -66.04376220703125, "p10": -25.579362297058104, "median": 28.847671508789062, "p90": 130.05593109130862, "max": 223.266357421875, "pos_frac": 0.78125, "sample": [37.7933349609375, 142.4087371826172, 41.33203125, -48.09002685546875, -66.04376220703125, 113.66021728515625, 31.854978561401367, 10.695182800292969, 223.266357421875, 0.00797271728515625, -12.544921875, 135.69198608398438, 9.395561218261719, 7.222419738769531, -5.686592102050781, 79.1893539428711, -7.501564025878906, 102.79005432128906, 173.84999084472656, 110.79450225830078, 123.31895446777344, 84.80412292480469, 1.9846477508544922, -24.008310317993164, 49.842437744140625, 10.471149444580078, 27.47071075439453, 15.882705688476562, 82.557373046875, 5.925403594970703, 27.27743911743164, 0.6485576629638672, 72.2856674194336, -26.252670288085938, 181.27828979492188, -41.74046325683594, 83.54135131835938, -31.32082748413086, -15.295379638671875, 4.84483528137207, 26.730979919433594, 8.013565063476562, 53.28948974609375, 165.968017578125, 36.00237274169922, 11.10063362121582, 93.55773162841797, 76.25607299804688, 32.66905212402344, -55.2474365234375, 15.307266235351562, 53.78374481201172, 132.94320678710938, 65.61663818359375, -2.5224990844726562, -1.4406070709228516, 39.54145050048828, 17.95699691772461, 75.72555541992188, -51.011104583740234, 30.224632263183594, 42.38407897949219, 69.44921875, 23.68462371826172], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000108.npy"}
{"epoch": 0.15859030837004406, "step": 109, "batch_size": 64, "mean": 44.3798713684082, "std": 53.44099807739258, "min": -93.77218627929688, "p10": -10.261032104492186, "median": 38.047250747680664, "p90": 112.3424758911133, "max": 196.07412719726562, "pos_frac": 0.828125, "sample": [53.35676574707031, 8.139392852783203, 141.05860900878906, 90.1971206665039, 32.232826232910156, 56.04220962524414, -61.619476318359375, 10.965591430664062, 18.807907104492188, 51.81092071533203, 138.19692993164062, 43.67234420776367, 76.62477111816406, 59.49571228027344, 54.505226135253906, 5.730348587036133, 23.066913604736328, 38.10265350341797, -13.84146499633789, -5.200904846191406, -6.354766845703125, 196.07412719726562, 0.8067264556884766, -8.4053955078125, 25.46708869934082, 5.038185119628906, 130.433837890625, 30.971158981323242, 34.192405700683594, -41.18037796020508, 67.95651245117188, 71.92585754394531, 10.051246643066406, 76.38568115234375, 57.297119140625, 40.16661834716797, 106.83735656738281, 76.18000793457031, 113.33673858642578, 110.02252960205078, -1.5342636108398438, 13.054252624511719, -93.77218627929688, 30.583438873291016, 1.22259521484375, 75.45157623291016, 37.558074951171875, 35.05691909790039, 20.876901626586914, 84.4801025390625, -54.039703369140625, 33.9027099609375, 37.99184799194336, 120.41009521484375, 132.29527282714844, 40.19026184082031, 91.55451202392578, 38.44872283935547, 94.0264892578125, 109.93791961669922, -11.056304931640625, -25.894622802734375, 2.6466236114501953, 108.37354278564453], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000109.npy"}
{"epoch": 0.16005873715124816, "step": 110, "batch_size": 64, "mean": 39.914703369140625, "std": 56.72380065917969, "min": -102.58993530273438, "p10": -14.097050476074218, "median": 33.41931629180908, "p90": 115.4513122558594, "max": 191.91729736328125, "pos_frac": 0.8125, "sample": [133.27540588378906, 56.86442565917969, 21.812267303466797, 56.50199890136719, 61.93883514404297, 38.458473205566406, 51.634273529052734, -1.9881515502929688, -102.58993530273438, 47.5303955078125, 69.82795715332031, 15.117454528808594, -70.74940490722656, 19.820178985595703, 12.801826477050781, 1.235006332397461, 31.439546585083008, 176.1815185546875, 107.37576293945312, 28.068435668945312, 4.640399932861328, -8.413707733154297, 191.91729736328125, 144.77764892578125, 106.13102722167969, 79.93873596191406, 18.957061767578125, -21.679061889648438, -14.238243103027344, 62.887428283691406, -45.84919738769531, -13.767601013183594, 10.127891540527344, 19.73190689086914, 5.552022933959961, 165.02981567382812, 55.41618347167969, 87.34642028808594, 118.91226196289062, 4.951030731201172, 80.2194595336914, -8.790179252624512, 41.95282745361328, 35.399085998535156, 15.032123565673828, 11.617103576660156, 85.56829071044922, -60.2587890625, 79.17475891113281, -28.065383911132812, 40.304054260253906, 1.7375202178955078, 59.29474639892578, 38.780513763427734, 53.4582633972168, 48.106475830078125, 19.00946044921875, 19.46410369873047, 19.036039352416992, 47.91974639892578, 83.56318664550781, 2.19293212890625, -5.003177642822266, 147.90213012695312], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000110.npy"}
{"epoch": 0.16152716593245228, "step": 111, "batch_size": 64, "mean": 45.586219787597656, "std": 63.17963409423828, "min": -93.1787109375, "p10": -35.48296661376953, "median": 35.5034065246582, "p90": 127.89487457275392, "max": 200.8795166015625, "pos_frac": 0.765625, "sample": [200.8795166015625, -56.39111328125, -86.94973754882812, 130.01983642578125, -7.348217010498047, 57.62339782714844, 15.351028442382812, 36.60350799560547, 28.41840362548828, 155.38075256347656, 34.40330505371094, 80.9912109375, 43.717872619628906, 59.81951904296875, 29.553115844726562, 121.51483154296875, 116.9455795288086, -39.52030944824219, 104.41700744628906, 7.3382415771484375, 83.8936996459961, -30.577102661132812, -23.98696517944336, -10.728668212890625, 70.68867492675781, 121.84648132324219, -8.418647766113281, 119.83049011230469, -37.585479736328125, 14.082283020019531, -25.590972900390625, 34.381248474121094, 143.03387451171875, 24.099014282226562, 159.74740600585938, 58.601409912109375, 32.64347839355469, 57.44684982299805, 69.28369903564453, 133.53036499023438, 24.697967529296875, -44.51982116699219, 19.182987213134766, 31.915454864501953, 141.84710693359375, 74.13768005371094, 29.52234649658203, 66.70083618164062, 122.93663024902344, 2.5762901306152344, 119.01420593261719, 40.249786376953125, 100.98616790771484, 73.63052368164062, 53.90454864501953, 14.32965087890625, -93.1787109375, 33.730621337890625, 30.018896102905273, 42.989013671875, -5.46318244934082, 100.10569763183594, -69.46693420410156, -11.318702697753906], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000111.npy"}
{"epoch": 0.16299559471365638, "step": 112, "batch_size": 64, "mean": 28.681774139404297, "std": 43.01564025878906, "min": -62.052406311035156, "p10": -25.994490814208984, "median": 27.693531036376953, "p90": 96.49846954345705, "max": 124.28042602539062, "pos_frac": 0.71875, "sample": [48.303897857666016, 124.28042602539062, 92.49195861816406, -6.629255294799805, -44.960933685302734, 35.89655303955078, 38.45997619628906, 33.46171569824219, 30.245079040527344, -5.070526123046875, 105.48959350585938, 83.0113525390625, -1.3812217712402344, -2.5904388427734375, 98.21554565429688, 62.86855697631836, 40.668575286865234, 57.51336669921875, 40.2112922668457, 39.65312194824219, 46.321807861328125, 28.70061492919922, 75.22956848144531, -17.22222900390625, 25.396469116210938, 23.680002212524414, 33.980064392089844, 34.25655746459961, -25.607528686523438, 18.095455169677734, 2.0219974517822266, -37.41933059692383, 32.551422119140625, 113.55229949951172, 117.14273834228516, -33.95198059082031, 42.6922607421875, 16.354528427124023, 74.50146484375, -0.7475109100341797, -62.052406311035156, 100.42321014404297, 27.15948486328125, 13.211090087890625, -14.806144714355469, 21.14516830444336, 66.4647216796875, -0.6760311126708984, 14.783103942871094, 102.77189636230469, 28.227577209472656, 9.551204681396484, 47.53997039794922, -45.7215576171875, 13.775964736938477, -26.16033172607422, -29.148239135742188, -12.819831848144531, 80.37745666503906, 0.8726673126220703, 25.319364547729492, 5.173622131347656, -10.36572265625, 40.91987609863281], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000112.npy"}
{"epoch": 0.1644640234948605, "step": 113, "batch_size": 64, "mean": 30.759035110473633, "std": 40.77458572387695, "min": -43.19252014160156, "p10": -17.43873519897461, "median": 31.213282585144043, "p90": 86.79063339233399, "max": 122.36632537841797, "pos_frac": 0.734375, "sample": [-34.77787780761719, 79.26805114746094, -18.027183532714844, 95.09625244140625, 7.67657470703125, 39.75459289550781, 13.653724670410156, 37.4127197265625, 2.817140579223633, -11.748916625976562, 9.709373474121094, 91.69624328613281, 56.51866912841797, 118.03506469726562, 39.64923095703125, -20.271169662475586, -8.748516082763672, -11.185935974121094, 7.825775146484375, 67.19609069824219, -20.596633911132812, 51.757659912109375, 84.62351989746094, -25.7967529296875, 56.17621612548828, -8.876014709472656, 2.336019515991211, 122.36632537841797, 50.96088409423828, -43.19252014160156, 20.147293090820312, 77.38921356201172, 35.75432205200195, -9.645477294921875, 38.222076416015625, 54.818359375, 15.260887145996094, 87.26619720458984, 19.113998413085938, -16.065689086914062, 54.599884033203125, 73.56504821777344, -13.903032302856445, 2.215669631958008, -23.007797241210938, 43.83053970336914, 38.781654357910156, 42.377685546875, -10.140968322753906, 16.397178649902344, 120.67626953125, -2.736225128173828, 34.83272933959961, 30.670412063598633, 106.5516586303711, 85.68098449707031, 34.94234085083008, 67.3964614868164, -7.290964126586914, 9.64703369140625, 12.173309326171875, 15.00540542602539, 31.756153106689453, 50.986968994140625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000113.npy"}
{"epoch": 0.16593245227606462, "step": 114, "batch_size": 64, "mean": 49.26908874511719, "std": 68.18498229980469, "min": -43.71837615966797, "p10": -30.19742202758789, "median": 43.774845123291016, "p90": 118.16929779052738, "max": 368.138671875, "pos_frac": 0.765625, "sample": [187.74713134765625, 95.63835906982422, 56.293556213378906, 50.52820587158203, 28.184478759765625, 121.98623657226562, 21.560049057006836, 48.54368591308594, -40.055267333984375, 71.08395385742188, -32.598548889160156, 78.32632446289062, 44.91271209716797, 23.39702606201172, -38.77648162841797, 91.65510559082031, 14.707088470458984, -35.39905548095703, 35.44832992553711, 130.59742736816406, 12.028432846069336, 57.53717803955078, -5.739105224609375, 125.44940948486328, 42.87867736816406, 61.58396911621094, 36.503318786621094, 24.04288101196289, 39.276824951171875, 63.01395034790039, 67.31647491455078, -5.709846496582031, -10.474853515625, 57.96360778808594, -29.306907653808594, 16.074392318725586, 51.193511962890625, 74.06793212890625, 40.694488525390625, 109.26310729980469, 62.499454498291016, 89.3515625, 6.48265266418457, 51.59349822998047, 42.22575759887695, -15.1517333984375, 242.41087341308594, 7.388219833374023, 44.67101287841797, 150.25927734375, -11.496221542358398, 76.41704559326172, -43.71837615966797, -30.579071044921875, 368.138671875, 58.35353469848633, -4.030439376831055, 93.967529296875, 36.09355163574219, -34.24032974243164, 28.847869873046875, 92.86622619628906, -19.076473236083984, 78.50953674316406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000114.npy"}
{"epoch": 0.16740088105726872, "step": 115, "batch_size": 64, "mean": 33.492794036865234, "std": 68.63591003417969, "min": -156.69357299804688, "p10": -30.722257995605467, "median": 19.283442497253418, "p90": 129.34036483764655, "max": 244.06204223632812, "pos_frac": 0.71875, "sample": [-1.6909751892089844, -2.3025779724121094, 134.7747039794922, 62.01983642578125, -55.6622314453125, -1.7578582763671875, -49.32771301269531, 10.090156555175781, 12.816898345947266, 17.02739715576172, 26.409217834472656, -6.361015319824219, 45.7226448059082, 58.220726013183594, 9.641275405883789, 177.08555603027344, 21.923295974731445, -8.717880249023438, 244.06204223632812, -12.245603561401367, 77.98381805419922, 7.73016357421875, 146.45114135742188, 73.2918472290039, 179.9283447265625, -31.250091552734375, 23.886474609375, 26.856555938720703, 7.913663864135742, 101.12809753417969, 182.2416534423828, 104.34107971191406, -0.16223907470703125, -128.98367309570312, 10.932491302490234, 35.81156921386719, 61.18254089355469, 16.712844848632812, 8.523197174072266, 46.38445281982422, 27.414077758789062, 54.73277282714844, 26.749610900878906, 31.980445861816406, 148.05514526367188, -29.490646362304688, 22.894615173339844, 18.74645233154297, -8.406013488769531, 12.350227355957031, 19.820432662963867, 82.59730529785156, 4.514808654785156, 11.05001449584961, -21.260339736938477, 116.66024017333984, 66.52397155761719, -18.123165130615234, 65.35377502441406, -156.69357299804688, 113.9337158203125, -51.03180694580078, 3.9561939239501953, -31.421310424804688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000115.npy"}
{"epoch": 0.16886930983847284, "step": 116, "batch_size": 64, "mean": 37.400001525878906, "std": 48.41792297363281, "min": -39.806365966796875, "p10": -17.07525424957275, "median": 31.288333892822266, "p90": 93.14068450927735, "max": 219.86732482910156, "pos_frac": 0.78125, "sample": [7.923151016235352, 5.282527923583984, -35.63438034057617, -32.72484588623047, 25.985559463500977, 44.868614196777344, 69.78778076171875, 23.233478546142578, 50.70619201660156, 60.900482177734375, 30.884963989257812, -11.416849136352539, 45.146568298339844, 27.819297790527344, 10.856307983398438, -6.357084274291992, 56.16862487792969, -39.806365966796875, 87.14888000488281, 92.12013244628906, 40.31584167480469, -17.940093994140625, 24.911972045898438, 17.866294860839844, 52.24701690673828, 25.11833953857422, 84.06069946289062, 109.61016845703125, 32.62977600097656, -31.329193115234375, 130.41693115234375, 35.942020416259766, 23.31523895263672, 44.82417297363281, -15.057294845581055, -28.644813537597656, 3.834766387939453, 7.5968780517578125, -4.197359085083008, 106.6392822265625, 5.568199157714844, 142.81214904785156, 31.69170379638672, 89.91654968261719, 84.62220764160156, 36.637081146240234, 56.899169921875, -35.1326904296875, -9.567197799682617, 85.03223419189453, 22.823436737060547, -6.053607940673828, 219.86732482910156, 69.68831634521484, 11.567451477050781, 53.29542541503906, 70.56126403808594, 67.47169494628906, -11.377656936645508, 15.9862060546875, 0.6285514831542969, 40.757568359375, 101.27284240722656, 93.57806396484375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000116.npy"}
{"epoch": 0.17033773861967694, "step": 117, "batch_size": 64, "mean": 30.609127044677734, "std": 53.2126350402832, "min": -104.55789184570312, "p10": -27.363672256469716, "median": 24.973280906677246, "p90": 90.57948150634769, "max": 195.61016845703125, "pos_frac": 0.765625, "sample": [13.184894561767578, 54.671409606933594, 50.8719482421875, -16.826915740966797, 23.240942001342773, -4.282844543457031, 19.74540901184082, 40.30571746826172, 21.860260009765625, 41.87831115722656, 44.59291458129883, 70.27410888671875, 140.63766479492188, 120.08888244628906, 15.689079284667969, 48.043975830078125, -58.97831726074219, 93.65797424316406, 2.002044677734375, -97.00155639648438, -31.879425048828125, 8.359207153320312, 17.01193618774414, 67.17605590820312, 45.73149490356445, -14.968414306640625, -38.84076690673828, 3.5769786834716797, 26.70561981201172, -15.948295593261719, 41.28446960449219, 123.321533203125, -0.5750503540039062, 4.976806640625, 156.93963623046875, 13.824665069580078, 195.61016845703125, 17.540664672851562, 30.607460021972656, 68.3283462524414, 16.339710235595703, 10.043819427490234, 8.475788116455078, -9.148918151855469, -104.55789184570312, 30.98804473876953, 83.39633178710938, 31.73798370361328, 95.9862289428711, 38.991455078125, -45.294403076171875, -10.687833786010742, 8.646503448486328, 57.30582809448242, 82.27156066894531, 81.47637939453125, 5.754291534423828, 51.36114501953125, 75.51668548583984, 28.966644287109375, 44.080589294433594, -2.761280059814453, -37.98878479003906, 75.6452865600586], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000117.npy"}
{"epoch": 0.17180616740088106, "step": 118, "batch_size": 64, "mean": 37.19419860839844, "std": 48.372371673583984, "min": -56.12825012207031, "p10": -13.783021926879877, "median": 25.310985565185547, "p90": 113.19612274169926, "max": 162.22763061523438, "pos_frac": 0.75, "sample": [-7.123279571533203, 20.84759521484375, -0.6472225189208984, -26.86859130859375, 20.593225479125977, 16.83621597290039, 62.51169967651367, 44.176475524902344, 49.16949462890625, 117.55867004394531, 90.77597045898438, 24.165298461914062, 130.15560913085938, 31.79090118408203, -9.10311508178711, 57.532981872558594, -19.562206268310547, -15.7886962890625, 36.15903091430664, 162.22763061523438, 120.46064758300781, -8.164854049682617, -56.12825012207031, 2.7696456909179688, -40.10894775390625, 20.51189613342285, 2.7717037200927734, 34.88636016845703, 15.405216217041016, 16.724960327148438, 10.412727355957031, 79.2580337524414, 3.5148468017578125, 79.57914733886719, 25.29114532470703, 51.90879821777344, 23.350187301635742, 43.21058654785156, 155.84567260742188, -4.166202545166016, 31.56043243408203, -17.958091735839844, -2.2333984375, 41.133338928222656, 151.56243896484375, 89.78539276123047, 119.64787292480469, 44.0311279296875, 81.45610809326172, 14.107389450073242, -0.8374404907226562, 71.1386489868164, 103.016845703125, 71.08106231689453, 54.42854309082031, 25.330825805664062, 36.97807312011719, 17.664104461669922, -1.426126480102539, 26.007827758789062, -22.34954833984375, 84.80703735351562, 3.9464340209960938, -5.191066741943359], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000118.npy"}
{"epoch": 0.17327459618208516, "step": 119, "batch_size": 64, "mean": 36.878746032714844, "std": 59.31734848022461, "min": -71.59136962890625, "p10": -32.3461555480957, "median": 38.059017181396484, "p90": 101.31132354736329, "max": 221.33203125, "pos_frac": 0.671875, "sample": [13.943864822387695, 60.85871887207031, 40.717979431152344, 156.37078857421875, 40.042236328125, -33.24079895019531, -6.152244567871094, 52.087554931640625, -5.335048675537109, 52.62825012207031, -10.729148864746094, -12.2138671875, 221.33203125, 51.116416931152344, 19.950855255126953, 127.2408447265625, 21.643741607666016, 12.1695556640625, 188.3007354736328, 48.58287048339844, 98.72299194335938, 78.09423828125, 79.0929946899414, -8.676456451416016, 32.24668884277344, 38.485198974609375, -36.42292785644531, 18.413009643554688, 58.833953857421875, 59.13523864746094, -1.8406524658203125, -3.6125946044921875, 102.42060852050781, 38.2573356628418, 180.54403686523438, -3.55718994140625, 67.5116195678711, 83.27473449707031, -2.3620166778564453, 28.01367950439453, -59.00575256347656, 44.89482879638672, -63.65594482421875, 135.401123046875, 51.48846435546875, 69.37128448486328, -26.512451171875, -33.107025146484375, 90.16997528076172, 37.86069869995117, -30.57079315185547, -71.59136962890625, -1.7075881958007812, 91.26922607421875, 16.323211669921875, 62.11513137817383, 53.94835662841797, 10.533544540405273, 65.02481079101562, -7.323781967163086, -42.78571701049805, 2.09844970703125, 44.16502380371094, -24.053794860839844], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000119.npy"}
{"epoch": 0.17474302496328928, "step": 120, "batch_size": 64, "mean": 51.01385498046875, "std": 62.62308883666992, "min": -89.33999633789062, "p10": -20.80453338623046, "median": 42.23662567138672, "p90": 140.7125411987305, "max": 202.5439453125, "pos_frac": 0.84375, "sample": [160.0157928466797, 94.13957214355469, 34.19349670410156, 114.12210083007812, 37.820743560791016, 9.144941329956055, -61.750640869140625, 104.87245178222656, 21.40349769592285, 71.11177825927734, 53.55486297607422, 172.42422485351562, 83.9892349243164, 6.584449768066406, 17.935544967651367, 27.540237426757812, 143.04953002929688, 63.33885192871094, 71.16161346435547, -41.501853942871094, -41.49607849121094, 135.2595672607422, 46.25172424316406, 33.10992431640625, 44.50080108642578, 38.0858154296875, 49.52191162109375, 6.918708801269531, 173.45257568359375, -89.33999633789062, -23.901084899902344, 43.55445861816406, 26.58135986328125, 3.3913726806640625, 109.65838623046875, 10.68046760559082, 202.5439453125, 55.85419845581055, 73.71185302734375, 3.553346633911133, 11.781978607177734, -7.579862594604492, 56.77702331542969, 194.16973876953125, 35.68809509277344, 43.242034912109375, 131.83377075195312, 178.40237426757812, 41.23121643066406, 49.011619567871094, 0.242950439453125, 91.4015121459961, -46.12825012207031, 19.8453369140625, 65.62303924560547, 36.70762634277344, -2.115001678466797, 118.08462524414062, 80.32098388671875, -13.579246520996094, 78.32674407958984, -24.32632064819336, 37.5313720703125, 3.34967041015625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000120.npy"}
{"epoch": 0.1762114537444934, "step": 121, "batch_size": 64, "mean": 56.53279113769531, "std": 73.29559326171875, "min": -153.65106201171875, "p10": -9.642184448242185, "median": 51.718421936035156, "p90": 158.00164947509768, "max": 228.11082458496094, "pos_frac": 0.765625, "sample": [9.149063110351562, -6.101844787597656, -2.273773193359375, 228.11082458496094, -5.965066909790039, 11.039054870605469, 100.13357543945312, -16.718063354492188, 80.23873138427734, 80.27938079833984, -4.407440185546875, 66.78673553466797, 146.2009735107422, 111.96643829345703, 93.69869995117188, -0.8730850219726562, 94.054443359375, 68.99749755859375, 53.824134826660156, -43.32494354248047, 40.30146789550781, 115.56175231933594, -62.58441925048828, 159.9301300048828, 176.56126403808594, 66.34638977050781, 37.16364288330078, 16.771251678466797, 118.63287353515625, 3.0164413452148438, 79.36282348632812, -21.451133728027344, 50.2691650390625, 89.84736633300781, 2.959932327270508, 206.62109375, 5.421623229980469, 191.46511840820312, 46.25636291503906, 214.47430419921875, 9.125404357910156, 133.284423828125, -10.871879577636719, 17.318588256835938, -23.042236328125, 9.959648132324219, 78.83039855957031, 25.29541778564453, -6.772895812988281, 57.36346435546875, 6.019243240356445, 19.78217315673828, 62.14945983886719, 138.1270751953125, 153.50186157226562, 74.12338256835938, -153.65106201171875, 58.64105224609375, -4.5940093994140625, 82.30442810058594, 219.23495483398438, 19.315872192382812, -2.256641387939453, 53.16767883300781], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000121.npy"}
{"epoch": 0.1776798825256975, "step": 122, "batch_size": 64, "mean": 69.67489624023438, "std": 69.28430938720703, "min": -96.42488098144531, "p10": 0.5527069091796877, "median": 51.703006744384766, "p90": 168.49698028564453, "max": 241.78985595703125, "pos_frac": 0.90625, "sample": [158.66285705566406, 23.099349975585938, 78.31763458251953, 86.68623352050781, 20.443819046020508, 26.351089477539062, 50.037139892578125, 101.37059783935547, 23.965103149414062, 120.53742980957031, 14.984420776367188, 135.3209228515625, 128.25772094726562, 241.78985595703125, 105.73079681396484, 37.209228515625, 17.760154724121094, 53.368873596191406, 168.5218048095703, 49.578765869140625, 21.452974319458008, 144.09280395507812, 38.12211990356445, -8.558639526367188, 88.22940063476562, 12.954269409179688, 67.5943374633789, 141.80459594726562, 73.94652557373047, 67.41260528564453, 79.60448455810547, -28.29570770263672, 178.64715576171875, -7.905660629272461, 53.889495849609375, 170.84014892578125, 64.58087921142578, 26.62957000732422, 188.88702392578125, 163.22642517089844, 10.031316757202148, 49.438507080078125, -34.098297119140625, 240.2978973388672, 17.874244689941406, 40.208438873291016, 33.24365997314453, 43.782596588134766, 23.68944549560547, -96.42488098144531, -9.179885864257812, 27.949432373046875, 35.635406494140625, 66.47686004638672, 209.52737426757812, 0.7025794982910156, 121.50956726074219, 55.464881896972656, 0.4884757995605469, 15.427314758300781, 168.43905639648438, 78.21073913574219, 19.982213973999023, 161.36761474609375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000122.npy"}
{"epoch": 0.17914831130690162, "step": 123, "batch_size": 64, "mean": 45.67673110961914, "std": 76.27165985107422, "min": -153.89529418945312, "p10": -28.39151782989502, "median": 22.69313335418701, "p90": 149.00979156494148, "max": 246.4869384765625, "pos_frac": 0.65625, "sample": [85.22441864013672, 83.4055404663086, -2.7174606323242188, 9.79220199584961, 92.08937072753906, 80.5224609375, 131.18048095703125, 118.53387451171875, 83.94879913330078, 50.355140686035156, -0.5021953582763672, -20.010330200195312, 8.785751342773438, -43.65941619873047, 156.6509246826172, 91.98081970214844, 17.323505401611328, -13.244598388671875, -2.3116607666015625, 69.08102416992188, -5.9058380126953125, 125.55099487304688, 68.07734680175781, -78.90963745117188, 21.23493194580078, 99.46151733398438, 214.96624755859375, 23.175350189208984, 159.43106079101562, 169.01809692382812, -28.85634422302246, 64.89982604980469, 21.802143096923828, 7.514015197753906, 86.86370086669922, 126.24647521972656, 17.76622772216797, 47.803924560546875, 182.2113494873047, -20.462677001953125, 95.28439331054688, 89.99844360351562, -43.57639694213867, -71.82999420166016, 78.0477294921875, -7.657806396484375, -43.66780471801758, -5.483247756958008, -18.91181182861328, 246.4869384765625, -27.306922912597656, 69.41513061523438, 56.74187469482422, 218.7371826171875, -11.63702392578125, 22.21091651916504, 126.32414245605469, 34.25121307373047, -153.89529418945312, -21.287208557128906, -0.42118072509765625, -9.372196197509766, 0.49814796447753906, 2.044179916381836], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000123.npy"}
{"epoch": 0.18061674008810572, "step": 124, "batch_size": 64, "mean": 67.91138458251953, "std": 98.88216400146484, "min": -127.48863220214844, "p10": -27.238236999511717, "median": 45.596534729003906, "p90": 186.08128356933594, "max": 389.5992126464844, "pos_frac": 0.78125, "sample": [157.39828491210938, -24.606155395507812, 42.77662658691406, 34.00679016113281, 28.894451141357422, 153.41786193847656, -99.82575988769531, -9.490776062011719, 170.6217041015625, 191.72064208984375, -4.34112548828125, 31.85186004638672, 4.510923385620117, 74.44757080078125, 83.24515533447266, 234.76705932617188, 166.4976806640625, 53.983089447021484, -21.217376708984375, 52.497886657714844, -127.48863220214844, -42.93656921386719, 145.55178833007812, -0.2588462829589844, 212.65432739257812, 137.63876342773438, 82.66142272949219, -4.338508605957031, 49.4801025390625, 124.90399169921875, -28.36627197265625, 33.12746810913086, 93.64533996582031, 12.187921524047852, 13.45379638671875, 359.35137939453125, 109.25143432617188, 186.78958129882812, 32.257835388183594, 184.4285888671875, 108.22553253173828, -54.519691467285156, 17.933582305908203, -18.193191528320312, 40.747718811035156, 130.7945556640625, -105.10641479492188, 105.44105529785156, 20.761777877807617, 48.41644287109375, 40.56541061401367, 85.23271942138672, -70.32391357421875, 265.1542053222656, 26.769073486328125, 389.5992126464844, 51.733428955078125, 172.53469848632812, 1.1633110046386719, 32.113304138183594, 51.92552185058594, 84.92879486083984, 14.73074722290039, 10.549354553222656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000124.npy"}
{"epoch": 0.18208516886930984, "step": 125, "batch_size": 64, "mean": 59.915245056152344, "std": 81.57054901123047, "min": -97.84259033203125, "p10": -30.504130554199218, "median": 34.17226028442383, "p90": 173.47404785156255, "max": 258.13812255859375, "pos_frac": 0.765625, "sample": [-3.1064319610595703, 154.54539489746094, 94.10765075683594, 23.285568237304688, -17.457141876220703, 9.207771301269531, 118.44294738769531, -29.255332946777344, 146.29217529296875, 16.393461227416992, 10.941959381103516, -31.039329528808594, 24.72166633605957, -0.00914764404296875, 105.18268585205078, 115.95346069335938, 94.25236511230469, 53.89795684814453, 20.517316818237305, 3.212261199951172, 108.64007568359375, 19.070697784423828, -57.349822998046875, -12.826416015625, 95.0413589477539, 258.13812255859375, -36.161590576171875, 32.055450439453125, 110.54603576660156, 38.572792053222656, -5.0278472900390625, -86.77144622802734, 35.94696044921875, 18.152729034423828, 32.397560119628906, 221.7404022216797, 7.413005828857422, 222.32376098632812, 124.33041381835938, -23.959312438964844, 11.928886413574219, 105.59048461914062, 193.35101318359375, 94.96552276611328, 240.35198974609375, 96.25660705566406, 55.88402557373047, 49.834197998046875, 178.58148193359375, 52.007225036621094, 15.345352172851562, 135.79498291015625, 43.239524841308594, -19.238492965698242, 120.39259338378906, 27.419971466064453, -46.470550537109375, 160.2715301513672, 31.955793380737305, -97.84259033203125, 216.7979278564453, 161.55670166015625, -34.7904052734375, 29.03203773498535], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000125.npy"}
{"epoch": 0.18355359765051396, "step": 126, "batch_size": 64, "mean": 55.9173698425293, "std": 75.7547607421875, "min": -80.50038146972656, "p10": -23.466880035400386, "median": 38.380483627319336, "p90": 162.58994598388674, "max": 274.1844482421875, "pos_frac": 0.828125, "sample": [127.37655639648438, 65.94830322265625, 11.108465194702148, 92.93700408935547, 12.855224609375, 54.007041931152344, -39.27360534667969, 2.1747074127197266, 63.92564392089844, 157.68899536132812, -56.478973388671875, 0.4880790710449219, 274.1844482421875, 19.47301483154297, 88.98853302001953, 0.47211456298828125, 137.69659423828125, 44.13732147216797, 135.07046508789062, 70.281982421875, 45.38959503173828, 53.32490921020508, -7.8411865234375, 39.56516647338867, 69.01766204833984, 195.74063110351562, -24.735977172851562, 164.6903533935547, 33.182613372802734, 48.96244812011719, 20.947154998779297, 26.06679916381836, 77.25775146484375, 14.37060546875, 187.648681640625, 37.19580078125, 58.19807434082031, -80.50038146972656, 155.3065185546875, -18.67728042602539, -65.90489196777344, -40.21504211425781, 248.6697998046875, -4.76951789855957, 28.38881492614746, 11.159442901611328, 1.9107627868652344, 19.501384735107422, 130.99668884277344, 123.14619445800781, -78.12578582763672, 195.60743713378906, 89.15042114257812, 24.49664306640625, 36.18815612792969, 99.57279968261719, 34.3067626953125, 64.82273864746094, 17.283679962158203, -20.505653381347656, 15.43841552734375, 94.21826171875, 11.606643676757812, 183.5958251953125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000126.npy"}
{"epoch": 0.18502202643171806, "step": 127, "batch_size": 64, "mean": 40.63016891479492, "std": 65.19891357421875, "min": -130.34555053710938, "p10": -37.054362487792964, "median": 42.36045455932617, "p90": 121.82489013671876, "max": 191.51512145996094, "pos_frac": 0.78125, "sample": [57.55083465576172, 84.03129577636719, -10.674449920654297, 39.58875274658203, 1.8771514892578125, -100.82737731933594, 155.62391662597656, 10.217514038085938, 100.11209869384766, -130.34555053710938, 7.638799667358398, 64.04340362548828, 77.18803405761719, -34.317466735839844, 18.298778533935547, 37.64894104003906, 96.8796157836914, 11.153684616088867, -128.65162658691406, 79.04286193847656, 26.27997589111328, 73.99908447265625, 132.71112060546875, 21.232513427734375, 69.37289428710938, 176.11282348632812, 21.747180938720703, 86.5870361328125, -23.364768981933594, 14.39154052734375, 123.37297821044922, 80.2917251586914, 109.09928894042969, 26.471115112304688, 53.802459716796875, 154.65005493164062, -2.139707565307617, -19.095077514648438, 20.113792419433594, -2.271167755126953, 17.767024993896484, 118.21268463134766, 9.654975891113281, 79.56965637207031, 48.581939697265625, 65.22400665283203, 46.7523193359375, 124.28099060058594, 112.3279800415039, 51.2349853515625, -46.214012145996094, 50.290504455566406, 19.681161880493164, 191.51512145996094, -62.022308349609375, 63.22052001953125, 47.952613830566406, 94.94242858886719, 45.13215637207031, 14.702852249145508, 32.487640380859375, -26.023277282714844, -38.227317810058594, -40.15794372558594], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000127.npy"}
{"epoch": 0.18649045521292218, "step": 128, "batch_size": 64, "mean": 42.86372375488281, "std": 76.49256896972656, "min": -176.556396484375, "p10": -48.59395599365234, "median": 38.16551971435547, "p90": 149.4205749511719, "max": 260.71075439453125, "pos_frac": 0.6875, "sample": [47.80910110473633, 44.68354034423828, 139.42813110351562, -14.159355163574219, 93.76202392578125, -64.66416931152344, 103.46298217773438, 147.67245483398438, 37.73234558105469, 45.20380401611328, 170.7503662109375, 53.25929260253906, 50.77601623535156, 35.66136932373047, -92.20227813720703, 103.13279724121094, 158.73184204101562, 4.237211227416992, -1.4320735931396484, 75.46876525878906, -73.57380676269531, -3.0689697265625, 260.71075439453125, -176.556396484375, 153.63491821289062, 50.73009490966797, 163.6964111328125, 149.6881103515625, 73.4101333618164, 49.08180236816406, -9.945850372314453, -59.55096435546875, 107.48076629638672, -5.532831192016602, -18.448871612548828, 18.43714141845703, -81.56929016113281, -2.334056854248047, 34.1890869140625, 47.80070877075195, 91.8130874633789, -39.91429138183594, 55.9803466796875, -26.071414947509766, -10.748918533325195, 151.6588897705078, 33.18960952758789, 38.59869384765625, 23.06622314453125, 52.37571334838867, 22.191505432128906, -10.073568344116211, 137.37014770507812, -1.5378589630126953, 12.022626876831055, 3.74798583984375, -11.269615173339844, 55.596168518066406, 106.63005065917969, 11.310100555419922, -52.313812255859375, 132.25003051757812, 148.79632568359375, 1.01702880859375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000128.npy"}
{"epoch": 0.18795888399412627, "step": 129, "batch_size": 64, "mean": 45.32090377807617, "std": 87.15413665771484, "min": -148.45135498046875, "p10": -44.9681770324707, "median": 32.88085746765137, "p90": 149.98919067382812, "max": 378.62078857421875, "pos_frac": 0.671875, "sample": [3.209596633911133, 33.70414352416992, -46.4543342590332, 115.08808898925781, -51.056365966796875, 124.03451538085938, 8.389396667480469, -43.29750061035156, 175.69338989257812, 26.488037109375, -12.034561157226562, 111.38658142089844, 26.937541961669922, 45.68251037597656, 58.42092514038086, 75.64432525634766, 150.65895080566406, -41.02740478515625, -31.956207275390625, 54.453346252441406, 224.83837890625, 88.84776306152344, 226.96109008789062, -100.20429229736328, 90.42486572265625, -94.40573120117188, -45.684181213378906, 42.937774658203125, 378.62078857421875, 144.30364990234375, 28.2703800201416, 22.92803955078125, 2.097808837890625, 79.7870101928711, -12.06854248046875, -18.233417510986328, 32.05757141113281, 123.74151611328125, 31.56641387939453, -13.889923095703125, -75.10771942138672, -8.045553207397461, 51.41363525390625, 81.74241638183594, 150.02932739257812, -148.45135498046875, 17.553476333618164, 20.383935928344727, -8.136266708374023, -7.130764007568359, 142.52569580078125, 178.62355041503906, 40.28218078613281, 48.64605712890625, 71.98040771484375, -17.830337524414062, 42.631103515625, 54.67365264892578, -39.00922393798828, 89.04259490966797, 149.89553833007812, -29.319801330566406, 77.82140350341797, -0.5380420684814453], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000129.npy"}
{"epoch": 0.1894273127753304, "step": 130, "batch_size": 64, "mean": 68.11644744873047, "std": 87.20930480957031, "min": -111.38019561767578, "p10": -22.048445129394526, "median": 46.84958267211914, "p90": 200.9209167480469, "max": 336.10321044921875, "pos_frac": 0.78125, "sample": [75.079833984375, 38.650821685791016, 32.69856262207031, 192.03863525390625, -15.990715026855469, 12.398574829101562, 208.42758178710938, 82.5575180053711, -56.604591369628906, -7.455476760864258, 16.0799617767334, 60.719112396240234, 33.24686813354492, 193.83444213867188, 210.9949951171875, 23.642433166503906, 95.24632263183594, -7.803400039672852, 38.20399475097656, 73.11717987060547, -12.30954360961914, 33.456329345703125, 8.685279846191406, 89.21244812011719, 92.98405456542969, -4.259864807128906, 33.77793884277344, 94.12981414794922, 59.62031555175781, 184.6063232421875, 143.32223510742188, 219.90524291992188, -34.694664001464844, 29.59564208984375, 151.81088256835938, -106.87217712402344, 33.95502471923828, 30.026016235351562, 34.43403244018555, -24.644615173339844, 60.249107360839844, 52.621280670166016, 75.9277114868164, 249.57435607910156, 159.77639770507812, -7.497123718261719, 336.10321044921875, -41.737510681152344, -4.481840133666992, 90.2322769165039, 32.81646728515625, 41.077884674072266, 203.95797729492188, -33.283226013183594, 149.4010772705078, -111.38019561767578, 133.9321746826172, 1.9907646179199219, 126.18060302734375, 118.10401916503906, 216.31948852539062, 57.89945983886719, 32.08408737182617, 63.76092529296875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000130.npy"}
{"epoch": 0.19089574155653452, "step": 131, "batch_size": 64, "mean": 68.61128997802734, "std": 75.48419952392578, "min": -155.53451538085938, "p10": -0.8779634475708008, "median": 51.62292289733887, "p90": 165.05243225097658, "max": 276.5924072265625, "pos_frac": 0.875, "sample": [50.51689529418945, 49.06587219238281, 53.79393768310547, 31.582321166992188, 276.5924072265625, -155.53451538085938, 59.7446403503418, 42.570457458496094, 33.503990173339844, 108.54102325439453, 112.04489135742188, 21.896629333496094, 62.01173400878906, 13.196044921875, 100.46293640136719, 36.50456237792969, 41.360679626464844, -1.5724353790283203, 145.15538024902344, 75.16450500488281, 45.38578796386719, 161.26931762695312, -12.831893920898438, -12.465925216674805, 52.72895050048828, 43.52490234375, 85.94261169433594, 7.032825469970703, 208.642822265625, 39.89361572265625, 42.55331802368164, -82.38861083984375, 166.67376708984375, 9.72702407836914, 228.44070434570312, 4.419731140136719, 104.05611419677734, 30.560028076171875, 140.38157653808594, 47.12655258178711, 129.20523071289062, 31.60761833190918, 211.48074340820312, 26.806072235107422, -92.080078125, 129.72476196289062, 4.6497955322265625, 108.95377349853516, -0.8724346160888672, 25.740116119384766, -0.8803329467773438, 139.10855102539062, 10.984519958496094, 179.9599151611328, 158.75021362304688, 172.02676391601562, 91.34574890136719, 67.19815063476562, 30.017993927001953, 78.88911437988281, 134.70123291015625, 96.5636978149414, 100.66209411621094, 89.30389404296875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000131.npy"}
{"epoch": 0.19236417033773862, "step": 132, "batch_size": 64, "mean": 55.7919921875, "std": 89.71038818359375, "min": -161.7420654296875, "p10": -17.23605976104736, "median": 42.59998893737793, "p90": 148.63128967285158, "max": 384.03314208984375, "pos_frac": 0.8125, "sample": [15.338172912597656, 10.831951141357422, 150.83436584472656, -63.651512145996094, 69.635498046875, 37.360565185546875, 66.518798828125, 295.3045959472656, 322.0931396484375, 8.369377136230469, 72.86373901367188, 110.50702667236328, 21.651830673217773, 78.86370849609375, -161.7420654296875, 3.4735946655273438, 223.9217529296875, -37.483497619628906, 68.0486068725586, -16.606794357299805, 2.3210811614990234, -13.409431457519531, 129.77490234375, 68.89686584472656, 21.4222412109375, -17.50574493408203, 0.23236656188964844, 115.62895965576172, 8.469009399414062, 2.9842147827148438, -12.343032836914062, -13.259674072265625, 19.17057228088379, 94.14473724365234, 81.7890625, 149.60067749023438, 54.795555114746094, 68.79615020751953, 80.38835144042969, 49.115108489990234, 54.53404235839844, 33.73032760620117, 32.017879486083984, 146.369384765625, -108.69598388671875, 47.839412689208984, 122.53070831298828, -70.0028305053711, 12.470836639404297, 93.57565307617188, -21.466188430786133, 62.15038299560547, 107.31916809082031, 17.882835388183594, 59.800628662109375, -13.26068115234375, 32.21180725097656, 106.56005859375, 27.195770263671875, 12.290962219238281, 183.14471435546875, 27.180217742919922, 54.130653381347656, 384.03314208984375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000132.npy"}
{"epoch": 0.19383259911894274, "step": 133, "batch_size": 64, "mean": 49.503807067871094, "std": 78.23494720458984, "min": -108.14418029785156, "p10": -22.13157196044921, "median": 38.91466522216797, "p90": 149.56125335693366, "max": 260.5428466796875, "pos_frac": 0.765625, "sample": [9.852920532226562, 69.79841613769531, -11.610519409179688, 114.56265258789062, 8.148845672607422, -41.88258361816406, -7.928436279296875, 37.64251708984375, -75.97097778320312, 260.5428466796875, 54.593772888183594, 71.7313232421875, -6.7285308837890625, 105.33061981201172, -83.97572326660156, 20.323471069335938, 95.116943359375, -5.895721435546875, 219.1416473388672, 5.771142959594727, 67.00259399414062, 85.38540649414062, -13.397636413574219, 134.12400817871094, 1.501190185546875, 108.92280578613281, 47.76318359375, 80.23245239257812, 86.82776641845703, 72.39889526367188, 10.527908325195312, 18.717208862304688, 25.289514541625977, 248.77194213867188, -1.4941177368164062, 51.19903564453125, 73.74712371826172, -70.78094482421875, 212.5262451171875, 40.18681335449219, 41.12081527709961, 37.055908203125, 54.205039978027344, 250.04922485351562, -108.14418029785156, -15.642120361328125, 66.85133361816406, 181.31033325195312, 18.01434326171875, 52.839900970458984, 7.497169494628906, 124.51556396484375, -56.71299743652344, 156.17721557617188, 103.7502212524414, -24.912765502929688, 66.99382019042969, 48.91593933105469, -9.450035095214844, 21.01149559020996, 6.454498291015625, 0.2929039001464844, 10.059158325195312, 17.974876403808594], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000133.npy"}
{"epoch": 0.19530102790014683, "step": 134, "batch_size": 64, "mean": 52.878135681152344, "std": 70.008056640625, "min": -44.40624237060547, "p10": -33.04430446624755, "median": 43.68878364562988, "p90": 136.33736877441407, "max": 346.2628479003906, "pos_frac": 0.8125, "sample": [16.69171142578125, -44.40624237060547, -39.396888732910156, 122.20454406738281, 205.94546508789062, 34.897850036621094, 17.97879409790039, 43.71236801147461, -35.99445343017578, 49.78839874267578, 123.18278503417969, 25.023719787597656, 12.479829788208008, 4.242637634277344, 53.311729431152344, 42.392730712890625, -29.462915420532227, 48.58683776855469, 78.64262390136719, 26.893165588378906, 5.471990585327148, 11.511650085449219, 100.80840301513672, 19.038551330566406, 167.4342803955078, -0.04815673828125, 116.28656005859375, 98.38473510742188, 56.758541107177734, 19.040515899658203, 49.20367431640625, 137.760986328125, 0.6290073394775391, 49.646942138671875, 174.1326904296875, 115.17717742919922, 133.01559448242188, -42.89739227294922, 87.033935546875, -21.312042236328125, 61.229766845703125, -35.9720458984375, 66.45228576660156, 9.368223190307617, 346.2628479003906, 11.304401397705078, 80.82957458496094, 40.58686828613281, 76.89547729492188, -4.795755386352539, 63.580108642578125, -25.85448455810547, -34.579185485839844, 181.4566650390625, 62.29253387451172, 2.9899826049804688, 60.90673828125, 39.56037139892578, 43.665199279785156, 72.65599822998047, -40.713111877441406, 82.36027526855469, 33.216121673583984, 156.70950317382812], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000134.npy"}
{"epoch": 0.19676945668135096, "step": 135, "batch_size": 64, "mean": 76.19829559326172, "std": 97.66053009033203, "min": -90.70086669921875, "p10": -34.96891708374023, "median": 53.97162628173828, "p90": 222.477751159668, "max": 373.25628662109375, "pos_frac": 0.796875, "sample": [28.96247100830078, 140.00347900390625, 33.64664840698242, 246.902587890625, 132.15765380859375, 78.84063720703125, 254.81358337402344, 40.75383377075195, 253.48519897460938, 227.1106719970703, 138.39401245117188, 92.54008483886719, 200.4486083984375, 128.87579345703125, 279.88873291015625, 122.60598754882812, 31.705711364746094, 208.61279296875, 12.62203598022461, -42.35536193847656, 0.6630344390869141, 112.93354797363281, 163.47702026367188, 77.37965393066406, 39.46527099609375, 100.03511047363281, 0.3047637939453125, 373.25628662109375, -24.898548126220703, 117.3890380859375, -62.399505615234375, 208.81338500976562, 211.6676025390625, -75.55192565917969, 42.17989730834961, -16.248233795166016, 27.53263282775879, 57.94879913330078, 132.27346801757812, -2.7370452880859375, -59.81890869140625, 54.93403625488281, 88.57919311523438, 59.81590270996094, -1.1090202331542969, 53.00921630859375, 3.011852264404297, -17.791244506835938, 47.27190399169922, 16.203689575195312, 33.7285041809082, 114.61540222167969, 94.85188293457031, 89.20635223388672, 8.292648315429688, 102.74578857421875, 44.579505920410156, 236.89559936523438, 13.254585266113281, -90.70086669921875, -30.639381408691406, -36.824432373046875, -47.2542724609375, 6.333778381347656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000135.npy"}
{"epoch": 0.19823788546255505, "step": 136, "batch_size": 64, "mean": 50.888790130615234, "std": 95.86885070800781, "min": -80.29052734375, "p10": -47.45306510925293, "median": 28.470012664794922, "p90": 125.3560417175293, "max": 423.3428955078125, "pos_frac": 0.703125, "sample": [67.07284545898438, 83.45618438720703, 104.405029296875, 103.62960815429688, -6.9961090087890625, -49.5265007019043, 308.18145751953125, 135.74803161621094, -19.533309936523438, 66.09870910644531, 106.64466857910156, 67.67268371582031, -26.24334716796875, 15.052310943603516, 76.59992980957031, 57.140838623046875, 82.7581787109375, 35.01337814331055, 116.92079162597656, 337.1744079589844, -1.8685302734375, 72.80078125, 423.3428955078125, 0.1021728515625, 116.48954772949219, 0.37427520751953125, 230.67977905273438, -8.329097747802734, 266.37249755859375, 25.873291015625, 9.804008483886719, 1.962677001953125, 59.76329803466797, -77.8577880859375, 57.04383850097656, -66.30555725097656, 20.197006225585938, -25.96759033203125, 93.45610046386719, 90.68780517578125, 9.715280532836914, -16.084598541259766, -14.427963256835938, 65.86750030517578, 11.095211029052734, 2.6930980682373047, -25.52267837524414, 31.066734313964844, 94.28158569335938, 125.58014678955078, -47.66480255126953, -62.66606903076172, 37.481201171875, 81.48770904541016, -80.29052734375, 0.9980850219726562, 74.40518188476562, -21.963054656982422, -46.95901107788086, 24.770156860351562, -3.6991233825683594, 124.8331298828125, 4.338493347167969, -62.34423065185547], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000136.npy"}
{"epoch": 0.19970631424375918, "step": 137, "batch_size": 64, "mean": 79.86851501464844, "std": 117.19479370117188, "min": -167.083984375, "p10": -27.86679077148437, "median": 53.125247955322266, "p90": 259.1609436035159, "max": 403.05682373046875, "pos_frac": 0.84375, "sample": [-41.94178771972656, 9.564350128173828, 79.41551208496094, 125.72440338134766, 157.62884521484375, 33.157264709472656, 48.57311248779297, 92.37045288085938, -23.13736343383789, 104.677001953125, 4.684089660644531, 96.98028564453125, 396.47998046875, 19.259145736694336, 39.769073486328125, -113.47662353515625, 116.129638671875, 10.647289276123047, 37.31501770019531, 159.77223205566406, 159.58233642578125, 85.45365142822266, 2.8386402130126953, 403.05682373046875, 115.27088928222656, 337.01544189453125, 197.17044067382812, 132.35020446777344, 18.46390724182129, 172.72341918945312, -97.75355529785156, 9.735687255859375, 71.14665985107422, 12.332054138183594, 136.6966552734375, 9.466667175292969, 128.84230041503906, 52.61156463623047, 6.781074523925781, 53.63893127441406, 65.74925231933594, 44.62115478515625, -5.943519592285156, -167.083984375, -45.457908630371094, -73.3313217163086, 374.32489013671875, -29.893688201904297, 38.413185119628906, 309.9812927246094, 56.685813903808594, 132.13986206054688, 3.432065963745117, 95.9583740234375, -20.4635066986084, 27.57090950012207, 54.284446716308594, 8.810558319091797, 1.3915882110595703, 108.45514678955078, 320.59197998046875, 35.98353576660156, 128.62039184570312, 285.7283020019531], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000137.npy"}
{"epoch": 0.2011747430249633, "step": 138, "batch_size": 64, "mean": 60.12742233276367, "std": 74.38688659667969, "min": -117.13275146484375, "p10": -18.106792449951172, "median": 42.988229751586914, "p90": 153.73201599121097, "max": 283.23919677734375, "pos_frac": 0.78125, "sample": [165.62692260742188, -13.24127197265625, 74.27473449707031, 140.67831420898438, 6.1841888427734375, 119.48035430908203, -18.514053344726562, -1.022247314453125, 8.849166870117188, 180.00462341308594, 32.83391571044922, 20.048294067382812, 41.24440002441406, 11.570331573486328, 28.38880729675293, -117.13275146484375, 15.204153060913086, -23.035987854003906, -2.6452560424804688, -11.523674011230469, 194.13241577148438, 182.08856201171875, 16.06375503540039, 10.843002319335938, 120.474609375, 131.6309814453125, 138.63690185546875, -5.056739807128906, 94.75409698486328, 145.79942321777344, 60.4256591796875, 29.11384391784668, 125.67849731445312, 8.873565673828125, 142.33236694335938, 175.00291442871094, 79.21893310546875, 44.732059478759766, -72.27066802978516, 63.11925506591797, 67.66921997070312, 27.71076202392578, 92.84709167480469, 0.7315464019775391, -5.895530700683594, -54.28294372558594, 102.65621185302734, 142.31887817382812, 13.81551742553711, 37.68125915527344, 115.78413391113281, 33.77085876464844, 136.24777221679688, -37.10053253173828, -20.80000114440918, 62.1802864074707, -17.156517028808594, 70.91087341308594, 22.077835083007812, 283.23919677734375, 106.55660247802734, 58.64219665527344, 108.55231475830078, 157.13169860839844], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000138.npy"}
{"epoch": 0.2026431718061674, "step": 139, "batch_size": 64, "mean": 71.82127380371094, "std": 89.51931762695312, "min": -112.05509948730469, "p10": -31.848586273193355, "median": 58.27299880981445, "p90": 169.18340454101565, "max": 386.9401550292969, "pos_frac": 0.796875, "sample": [-21.147186279296875, 159.0792236328125, 101.5018539428711, 176.55015563964844, -33.79865264892578, 67.6703109741211, 121.74108123779297, -12.678192138671875, 85.66830444335938, 149.83444213867188, -112.05509948730469, 57.67118835449219, -47.88999557495117, 280.1092529296875, 144.08993530273438, 49.46974182128906, 90.73115539550781, 58.87480926513672, 35.734764099121094, 63.77524185180664, 148.43295288085938, 72.92145538330078, -72.46512603759766, 126.53328704833984, 50.21699523925781, 0.8599433898925781, 163.4619598388672, 86.15950012207031, 17.420377731323242, 40.26441192626953, 154.6767120361328, 27.540719985961914, 126.99771118164062, 119.37300872802734, 42.181175231933594, 97.55311584472656, -19.908836364746094, -1.028879165649414, 9.910865783691406, 15.914222717285156, 82.78299713134766, -0.312744140625, 17.608375549316406, 152.48040771484375, 23.629432678222656, 275.7154846191406, 97.86798858642578, -27.298431396484375, 78.05107879638672, -46.66106414794922, 85.95340728759766, -49.38727569580078, 33.183433532714844, 87.72460174560547, 386.9401550292969, 39.28584671020508, 251.5312957763672, 39.07032775878906, 51.075355529785156, -40.82630920410156, 193.70944213867188, 30.108734130859375, 40.74540710449219, 171.6354522705078], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000139.npy"}
{"epoch": 0.20411160058737152, "step": 140, "batch_size": 64, "mean": 64.74617004394531, "std": 91.64198303222656, "min": -130.9687957763672, "p10": -50.74279632568359, "median": 63.94810485839844, "p90": 178.06314086914068, "max": 352.21466064453125, "pos_frac": 0.734375, "sample": [-26.083324432373047, 126.10321044921875, 352.21466064453125, -80.50286102294922, -14.527587890625, 126.97911071777344, 85.91194915771484, 150.51210021972656, 59.86542510986328, 167.35411071777344, 121.75900268554688, 74.27638244628906, 116.09843444824219, 43.29846954345703, 27.142913818359375, 86.22348022460938, 278.6068115234375, 12.723934173583984, 2.9390830993652344, 74.8954086303711, 19.239059448242188, 67.1241683959961, 90.48812866210938, 135.74185180664062, -2.2713241577148438, -58.16162109375, 257.275146484375, -51.27214050292969, 37.69709777832031, 83.98475646972656, 25.238142013549805, 62.14441680908203, 8.977075576782227, 118.27751159667969, -32.621490478515625, 63.462738037109375, 94.66940307617188, 63.099403381347656, -9.221786499023438, 160.61434936523438, 134.84722900390625, -34.257164001464844, 182.65272521972656, -130.9687957763672, -2.1195220947265625, 64.4334716796875, 62.098724365234375, 93.39989471435547, 192.05801391601562, 24.412887573242188, -9.863794326782227, 161.25633239746094, -15.213006973266602, 20.7635498046875, -94.09654235839844, -51.67396926879883, -75.89669036865234, 69.95799255371094, 190.77755737304688, 183.96875, -49.507659912109375, 86.09772491455078, 84.0963134765625, 136.25469970703125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000140.npy"}
{"epoch": 0.2055800293685756, "step": 141, "batch_size": 64, "mean": 61.847801208496094, "std": 96.83552551269531, "min": -246.85104370117188, "p10": -45.15029373168944, "median": 62.68219566345215, "p90": 184.6835113525391, "max": 300.7598876953125, "pos_frac": 0.765625, "sample": [68.42794799804688, 41.270965576171875, -13.49123764038086, 104.84317779541016, 91.09614562988281, -31.671165466308594, 110.46797180175781, 65.49284362792969, 1.7173690795898438, -9.589500427246094, 62.89669418334961, 261.97344970703125, -246.85104370117188, 49.608253479003906, -5.710197448730469, 35.77410888671875, 0.5911865234375, -50.92706298828125, 269.94549560546875, -64.63633728027344, -29.412742614746094, 205.93064880371094, 45.2652473449707, 86.352783203125, 150.4039306640625, 82.58828735351562, 108.96334838867188, 4.248573303222656, -56.167633056640625, 71.09759521484375, 111.00013732910156, 78.26229858398438, 33.17174530029297, 168.88589477539062, 259.39019775390625, 30.146142959594727, 115.61540985107422, 24.988252639770508, -20.42133903503418, 20.159481048583984, -26.474754333496094, 50.47944641113281, -52.65581512451172, 191.45391845703125, 6.824087142944336, 142.98556518554688, 199.4642333984375, 122.91511535644531, 11.865266799926758, -162.70819091796875, 18.948814392089844, 97.00186157226562, 62.46769714355469, 92.48577880859375, 115.10867309570312, 132.35650634765625, 300.7598876953125, 127.16703033447266, -0.21909141540527344, 167.2391357421875, 70.10935974121094, 105.02188110351562, -53.773475646972656, 7.739051818847656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000141.npy"}
{"epoch": 0.20704845814977973, "step": 142, "batch_size": 64, "mean": 70.5022964477539, "std": 108.32785034179688, "min": -207.81704711914062, "p10": -41.0870635986328, "median": 38.027442932128906, "p90": 243.64519653320315, "max": 297.4927673339844, "pos_frac": 0.765625, "sample": [-22.592445373535156, 264.0014953613281, 34.119781494140625, 43.392913818359375, -30.495216369628906, 245.0459442138672, -26.966503143310547, -24.006032943725586, -89.70700073242188, 31.457500457763672, 131.93914794921875, 99.90701293945312, 157.85934448242188, 27.914621353149414, 40.39612579345703, 224.66603088378906, 11.031444549560547, -207.81704711914062, -148.6630859375, 73.37821197509766, 25.956375122070312, 35.21125793457031, 17.77334213256836, 297.4927673339844, 171.8452606201172, 37.06085205078125, 146.4038848876953, 24.58034324645996, 145.35061645507812, -23.700225830078125, 108.272705078125, 127.13836669921875, 245.2482147216797, 84.9704360961914, -1.33160400390625, 213.2920379638672, 217.70469665527344, 17.948070526123047, 20.4135799407959, 42.09417724609375, -58.50120544433594, -50.27340316772461, 3.996166229248047, 240.3767852783203, 67.56986236572266, 247.7090606689453, 14.428855895996094, 21.66364288330078, 71.06181335449219, -45.626426696777344, 58.76169204711914, -20.08135223388672, 6.808786392211914, -81.07919311523438, 197.3196563720703, 254.36468505859375, 27.68859100341797, 200.24807739257812, 26.875709533691406, -8.047948837280273, 125.6060562133789, 256.71954345703125, 38.99403381347656, 126.97614288330078], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000142.npy"}
{"epoch": 0.20851688693098386, "step": 143, "batch_size": 64, "mean": 45.315574645996094, "std": 95.56729125976562, "min": -159.3607177734375, "p10": -55.83040008544922, "median": 28.76413059234619, "p90": 147.10604858398438, "max": 341.33990478515625, "pos_frac": 0.65625, "sample": [84.39470672607422, -7.1738739013671875, 53.50725555419922, 26.942367553710938, 285.8425598144531, 92.27332305908203, -56.20952606201172, 108.15555572509766, 111.10560607910156, 23.855745315551758, 145.67990112304688, -86.23977661132812, -55.6580810546875, 137.14114379882812, 84.06848907470703, -129.62384033203125, -50.145660400390625, -2.5251731872558594, 105.41390991210938, -4.2005615234375, 35.055824279785156, 260.1191711425781, -67.85108184814453, 100.301025390625, 19.622833251953125, -18.495532989501953, 44.31665802001953, 147.71725463867188, 34.141456604003906, 17.337444305419922, -10.640342712402344, -48.4024658203125, -55.90425109863281, 4.65388298034668, 136.8449249267578, 6.769073486328125, 45.843589782714844, 65.73392486572266, 26.734375, 341.33990478515625, -31.02558135986328, -11.917783737182617, 36.97993469238281, 85.78656005859375, -2.412992477416992, 57.50042724609375, 157.82162475585938, 143.54531860351562, -5.126285552978516, -46.12261962890625, -90.16728210449219, 290.90802001953125, -32.46839904785156, 82.8254165649414, 67.45619201660156, 30.585893630981445, 111.39842224121094, 24.15766143798828, 184.42913818359375, -159.3607177734375, -10.789810180664062, 41.54682159423828, 5.429588317871094, 17.375823974609375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000143.npy"}
{"epoch": 0.20998531571218795, "step": 144, "batch_size": 64, "mean": 55.21876525878906, "std": 80.65177154541016, "min": -110.69418334960938, "p10": -35.316018676757814, "median": 47.17738342285156, "p90": 174.45550537109375, "max": 249.9200439453125, "pos_frac": 0.734375, "sample": [-36.08573913574219, 112.78209686279297, -6.089521408081055, 174.51043701171875, -105.28095245361328, -20.503379821777344, 69.01504516601562, 120.0544204711914, 46.876426696777344, 13.225452423095703, 66.4928970336914, -37.32742691040039, 75.59661865234375, 106.61300659179688, 59.56724548339844, 174.32733154296875, 53.07621765136719, 210.73785400390625, 106.7385482788086, 144.27127075195312, 180.91854858398438, 191.89129638671875, -25.190095901489258, -110.69418334960938, 106.29415130615234, -24.736083984375, 48.69085693359375, -2.9470748901367188, 68.74749755859375, 22.494911193847656, -56.494842529296875, 73.11298370361328, 69.23178100585938, 23.687286376953125, -51.18109893798828, 142.33604431152344, 47.962379455566406, 28.65401840209961, 24.744285583496094, 93.02299499511719, 31.010284423828125, 93.9804916381836, 249.9200439453125, 28.70796012878418, 28.18853759765625, 75.79400634765625, -33.52000427246094, 24.64665985107422, 20.513931274414062, 104.55610656738281, 123.37169647216797, 237.1345977783203, -2.806854248046875, -33.274383544921875, -17.80467987060547, -22.700157165527344, 225.91761779785156, 14.602333068847656, 22.903257369995117, 1.0141220092773438, 47.47834014892578, 30.9219970703125, 152.04898071289062, -47.74748992919922], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000144.npy"}
{"epoch": 0.21145374449339208, "step": 145, "batch_size": 64, "mean": 59.55731964111328, "std": 83.35301971435547, "min": -124.64822387695312, "p10": -44.97537345886229, "median": 44.547067642211914, "p90": 155.109765625, "max": 271.1991882324219, "pos_frac": 0.75, "sample": [23.654953002929688, -49.87469482421875, -99.17041015625, 271.1991882324219, 153.97747802734375, 121.22005462646484, 250.78689575195312, 3.394824981689453, -124.64822387695312, 57.98165512084961, 4.929412841796875, 75.48945617675781, 148.71697998046875, -53.4918327331543, 11.072929382324219, 122.64575958251953, 56.12946319580078, 42.938758850097656, 122.58653259277344, -49.16393280029297, -14.181303024291992, 42.16900634765625, 43.739112854003906, 139.29037475585938, 92.53912353515625, -67.30814361572266, -9.730010986328125, 31.303470611572266, -7.067779541015625, 189.01840209960938, 22.975072860717773, 14.21197509765625, 118.04118347167969, -35.20206832885742, -53.74298095703125, 121.81922912597656, 19.322647094726562, 21.861787796020508, 82.65103149414062, -12.615737915039062, 68.09202575683594, -6.082279205322266, 121.76615142822266, -17.858348846435547, 118.26191711425781, 198.40708923339844, 105.07496643066406, -22.961437225341797, 49.39950942993164, 128.28781127929688, 29.89120101928711, 43.505775451660156, 145.66310119628906, 162.04287719726562, 60.500545501708984, 18.455806732177734, -3.4200668334960938, 69.3956298828125, 136.99020385742188, 120.78887176513672, 244.00534057617188, 45.35502243041992, 11.04217529296875, 155.59503173828125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000145.npy"}
{"epoch": 0.21292217327459617, "step": 146, "batch_size": 64, "mean": 46.24092102050781, "std": 90.77281951904297, "min": -241.95364379882812, "p10": -70.27976760864257, "median": 57.835771560668945, "p90": 162.65381317138676, "max": 252.55430603027344, "pos_frac": 0.671875, "sample": [34.82682800292969, -36.588226318359375, 111.20654296875, -66.19550323486328, 19.991609573364258, 2.2663192749023438, 77.27558135986328, 66.8695068359375, 77.60209655761719, 67.44760131835938, 213.68408203125, 106.40033721923828, 44.40686798095703, -12.65673828125, 25.054664611816406, -9.223167419433594, 96.3307113647461, -9.854110717773438, 39.561866760253906, -18.559160232543945, -5.13585090637207, 32.07086181640625, 64.54683685302734, 119.71092224121094, -104.80958557128906, 69.5680160522461, 151.2724151611328, -241.95364379882812, -72.03016662597656, 210.628662109375, -81.91243743896484, 75.68260955810547, -6.41285514831543, 75.20404052734375, 195.5082244873047, -79.12523651123047, 49.05247497558594, 64.35958862304688, 123.82683563232422, 252.55430603027344, 6.210367202758789, -5.608917236328125, 57.02064895629883, 82.79048156738281, 62.32740783691406, 104.87055206298828, 105.14550018310547, -101.85530090332031, 58.65089416503906, -75.99034118652344, 85.06074523925781, 142.22866821289062, 167.53155517578125, 74.46878051757812, -4.40081787109375, 105.78917694091797, 38.52849578857422, -38.3182373046875, 89.31977844238281, -55.58782958984375, 233.62037658691406, -35.485389709472656, 192.75210571289062, -52.10362243652344], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000146.npy"}
{"epoch": 0.2143906020558003, "step": 147, "batch_size": 64, "mean": 62.786094665527344, "std": 94.06626892089844, "min": -150.1434326171875, "p10": -32.36549606323242, "median": 48.02958679199219, "p90": 201.97363586425786, "max": 296.28802490234375, "pos_frac": 0.75, "sample": [108.87672424316406, -68.46875, 41.54585266113281, 242.54293823242188, -29.562416076660156, -9.416252136230469, 99.86895751953125, 100.9970474243164, 49.96210479736328, -42.56958770751953, 30.5531063079834, 81.82330322265625, 62.54669189453125, 130.64215087890625, 25.407291412353516, 207.434326171875, 120.05076599121094, 17.479066848754883, -1.2186431884765625, 182.35214233398438, 163.3134002685547, -26.674583435058594, 178.2440948486328, 23.73447608947754, 289.7379455566406, 48.91670227050781, 211.7738037109375, 110.80653381347656, -16.48396110534668, 47.14247131347656, 32.28032302856445, -143.9494171142578, 6.271553039550781, 118.60623168945312, 114.84536743164062, -30.16790771484375, 64.20365905761719, 296.28802490234375, 80.19074249267578, 20.86435317993164, 189.23202514648438, 13.511810302734375, -14.103208541870117, 14.764419555664062, 33.55766296386719, 62.029541015625, -12.625015258789062, 85.88421630859375, 67.42074584960938, -35.053993225097656, 13.41851806640625, 55.132198333740234, 18.325450897216797, 179.5651397705078, -26.033676147460938, 59.23313522338867, 251.127685546875, -33.30731964111328, 35.63409423828125, -150.1434326171875, 220.61923217773438, -61.21453857421875, 64.85889434814453, 45.68608093261719], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000147.npy"}
{"epoch": 0.21585903083700442, "step": 148, "batch_size": 64, "mean": 54.11583709716797, "std": 86.13773345947266, "min": -104.65142822265625, "p10": -31.25388031005859, "median": 37.82162857055664, "p90": 152.42807006835938, "max": 354.078369140625, "pos_frac": 0.796875, "sample": [-6.531726837158203, 53.07164001464844, 6.2491455078125, 86.18215942382812, -65.65819549560547, 110.7676773071289, 79.70193481445312, 40.07247543334961, -41.85993957519531, 209.16188049316406, 22.04303741455078, 149.7259521484375, 20.956281661987305, -14.018089294433594, 158.13888549804688, 258.0711669921875, 4.830657958984375, 15.963531494140625, -41.90306854248047, -72.18658447265625, 138.50315856933594, 133.69625854492188, -12.317083358764648, 93.94284057617188, 98.32994842529297, 43.99531555175781, 65.3314208984375, 50.162967681884766, 111.20658874511719, 0.6028251647949219, 153.58612060546875, 34.768829345703125, 39.13572692871094, 73.14891052246094, -32.16924285888672, 50.6265754699707, 32.11656951904297, 20.602890014648438, 117.07434844970703, 61.49602127075195, 135.33282470703125, 37.326805114746094, -88.61054992675781, 47.095298767089844, 41.965362548828125, 23.7860107421875, 189.11990356445312, 54.030906677246094, -5.844268798828125, -29.11803436279297, -104.65142822265625, 38.31645202636719, 8.75596809387207, 6.398950576782227, 35.24109649658203, 7.009759902954102, 34.514549255371094, 354.078369140625, 328.65203857421875, 8.131235122680664, -2.8141651153564453, 0.9256744384765625, 91.26824951171875, 5.882783889770508], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000148.npy"}
{"epoch": 0.2173274596182085, "step": 149, "batch_size": 64, "mean": 66.43098449707031, "std": 84.82844543457031, "min": -197.9447784423828, "p10": -9.311510276794433, "median": 58.72701072692871, "p90": 157.1783645629883, "max": 314.93377685546875, "pos_frac": 0.828125, "sample": [118.87667083740234, 71.25717163085938, 212.072998046875, 60.73662185668945, 108.01131439208984, 72.92668151855469, 74.82386016845703, -8.988855361938477, -18.806488037109375, -11.70180892944336, 22.347280502319336, -9.449790954589844, 59.4604377746582, 46.70375061035156, 62.276641845703125, -68.19784545898438, 91.17147827148438, 92.89031982421875, 44.430877685546875, 136.98788452148438, 116.7462158203125, 178.2341766357422, 110.31714630126953, 30.40126609802246, 14.814279556274414, 99.861328125, 48.94762420654297, 129.68472290039062, 77.19557189941406, 136.01025390625, 63.25310516357422, -1.1422805786132812, 314.93377685546875, 42.167015075683594, 157.91952514648438, 24.173553466796875, 230.08294677734375, 65.31725311279297, 29.077865600585938, 112.67981719970703, 24.338451385498047, 74.81388854980469, 11.229913711547852, 43.98377227783203, -0.6773147583007812, 57.99358367919922, 14.089916229248047, 57.758750915527344, -15.084503173828125, -197.9447784423828, 93.46328735351562, 9.510971069335938, 4.612030029296875, 155.44898986816406, 11.011716842651367, 288.7073059082031, 15.642656326293945, 41.679443359375, 143.93910217285156, 274.3784484863281, 8.765876770019531, -7.125270843505859, 92.60368347167969, -90.0610580444336], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000149.npy"}
{"epoch": 0.21879588839941264, "step": 150, "batch_size": 64, "mean": 65.54271697998047, "std": 97.49305725097656, "min": -84.86701965332031, "p10": -19.633739471435543, "median": 40.78498458862305, "p90": 177.31770477294924, "max": 432.5999755859375, "pos_frac": 0.8125, "sample": [303.62359619140625, 326.626220703125, 21.527198791503906, 29.62908935546875, 60.993690490722656, -40.938697814941406, 192.33062744140625, 39.816856384277344, -81.6226577758789, 74.9897232055664, 432.5999755859375, -9.382829666137695, -14.9052734375, 19.938011169433594, -9.726974487304688, 136.1461944580078, 20.39080810546875, 52.44652557373047, 49.49833679199219, 76.41936492919922, 41.75311279296875, 63.044151306152344, 145.90737915039062, 28.007980346679688, 112.52821350097656, 90.96634674072266, 199.9247589111328, 135.14659118652344, 14.166351318359375, 68.46389770507812, 14.783714294433594, -50.51068115234375, 84.66849517822266, 181.10658264160156, 71.3614501953125, 7.954395294189453, 168.47698974609375, 22.542402267456055, 148.18792724609375, -21.66022491455078, 96.79254150390625, -41.553382873535156, -43.652671813964844, 6.421562194824219, 125.24462127685547, 7.417961120605469, 29.90418243408203, 65.22103881835938, 16.086036682128906, 41.86116027832031, 62.28163146972656, 71.82672119140625, 65.917236328125, 19.78421401977539, 28.666112899780273, 357.68206787109375, 11.819211959838867, 18.084930419921875, 13.24748420715332, 30.24224853515625, 101.94855499267578, -11.953994750976562, -84.86701965332031, -0.9081687927246094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000150.npy"}
{"epoch": 0.22026431718061673, "step": 151, "batch_size": 64, "mean": 65.77288055419922, "std": 81.95947265625, "min": -145.7357177734375, "p10": -28.7441104888916, "median": 57.43894958496094, "p90": 187.25802001953127, "max": 255.29400634765625, "pos_frac": 0.828125, "sample": [-145.7357177734375, 225.46629333496094, 51.04338073730469, -0.30657196044921875, 43.707061767578125, 60.47074890136719, 69.52430725097656, 115.51882934570312, 71.37809753417969, 30.155654907226562, 182.44415283203125, -45.627418518066406, -29.523582458496094, 107.31515502929688, 56.99681091308594, -47.41484451293945, -5.001560211181641, 42.41796112060547, 58.22380065917969, 49.98517608642578, 234.76336669921875, 56.162601470947266, 149.64187622070312, 57.88108825683594, -55.21497344970703, 179.708251953125, 26.418228149414062, 9.033050537109375, -26.925342559814453, 22.77838134765625, 35.47529602050781, 11.891073226928711, 19.546863555908203, 8.318971633911133, 114.31338500976562, 182.2822265625, -65.24809265136719, 63.48644256591797, 109.18999481201172, 101.26974487304688, 7.995122909545898, 60.19566345214844, 101.50527954101562, 81.31776428222656, 11.474319458007812, 201.44613647460938, 107.26553344726562, 58.542930603027344, 124.68865966796875, -7.254859924316406, 216.7132110595703, 6.456275939941406, 2.9739246368408203, 146.7149658203125, 14.526199340820312, 15.593376159667969, 209.82919311523438, 115.69811248779297, 255.29400634765625, 189.32110595703125, -50.87702941894531, 79.87207794189453, 21.762451171875, 112.59983825683594], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000151.npy"}
{"epoch": 0.22173274596182085, "step": 152, "batch_size": 64, "mean": 80.4985580444336, "std": 98.08757781982422, "min": -88.37982177734375, "p10": -40.38363647460937, "median": 61.3629207611084, "p90": 222.47957305908204, "max": 330.5556640625, "pos_frac": 0.796875, "sample": [30.324493408203125, 58.68214416503906, 167.9054718017578, 179.6111297607422, 76.22479248046875, 209.90350341796875, 62.25248336791992, 147.43218994140625, 330.5556640625, 37.042816162109375, 160.4688720703125, 133.92430114746094, 38.11652755737305, -66.17176055908203, -4.142539978027344, 233.72940063476562, 143.737060546875, 52.35203552246094, 70.05513000488281, 63.55439758300781, 57.82897186279297, 52.873252868652344, -49.53947448730469, 77.41850280761719, 1.9329872131347656, 326.53155517578125, 131.6732177734375, 232.5188751220703, 43.48997497558594, 26.249061584472656, 66.78779602050781, -50.6075439453125, 20.214473724365234, -7.26922607421875, 1.0292186737060547, 60.473358154296875, 19.30278778076172, 205.01019287109375, 31.840545654296875, 47.030006408691406, 293.4748840332031, -24.666160583496094, 55.42149353027344, 216.76773071289062, 68.25950622558594, -33.56425476074219, 244.16783142089844, 34.550880432128906, -77.52245330810547, 109.94915771484375, 3.707895278930664, 148.05726623535156, 94.729736328125, 77.54271697998047, 224.92750549316406, -44.00608825683594, 167.1357879638672, -88.37982177734375, -9.343223571777344, 121.96340942382812, -43.30622863769531, -31.70037078857422, 139.79751586914062, 83.59654998779297], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000152.npy"}
{"epoch": 0.22320117474302498, "step": 153, "batch_size": 64, "mean": 68.51066589355469, "std": 105.1429214477539, "min": -184.85183715820312, "p10": -36.745690155029294, "median": 56.248504638671875, "p90": 205.71592102050786, "max": 428.1483154296875, "pos_frac": 0.75, "sample": [88.60888671875, -2.2952499389648438, 10.9100341796875, -8.321052551269531, 17.846324920654297, -8.214393615722656, 28.56069564819336, 161.256591796875, -40.74616241455078, 193.6151580810547, 99.94807434082031, 7.645057678222656, -44.98824691772461, 84.73985290527344, 249.40342712402344, -68.98389434814453, 77.07530212402344, 17.104631423950195, 56.83750915527344, 104.43480682373047, 55.65950012207031, 428.1483154296875, 41.042823791503906, 66.11803436279297, 83.76524353027344, 184.23580932617188, -27.64832305908203, -45.62267303466797, 210.90196228027344, 215.114501953125, 145.03526306152344, 43.235923767089844, 257.06298828125, 5.958950042724609, 2.131387710571289, 1.2936210632324219, 169.9556884765625, 99.2478256225586, -35.47718048095703, -37.289337158203125, 132.8029022216797, 12.8756103515625, 11.820648193359375, 279.0801086425781, 25.433792114257812, -16.7564697265625, -64.75576782226562, 66.49234008789062, 5.928016662597656, -184.85183715820312, 91.64921569824219, 368.56005859375, 79.48291015625, -32.03998565673828, 66.96163177490234, -4.133861541748047, 100.20662689208984, 45.6778564453125, 136.03652954101562, 141.38400268554688, 82.60549926757812, -21.062255859375, 66.7280502319336, 107.24964904785156], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000153.npy"}
{"epoch": 0.22466960352422907, "step": 154, "batch_size": 64, "mean": 65.37098693847656, "std": 109.29109191894531, "min": -210.20223999023438, "p10": -56.98156814575195, "median": 57.67643356323242, "p90": 198.84959869384767, "max": 308.2494812011719, "pos_frac": 0.640625, "sample": [-11.357053756713867, 58.924781799316406, -210.20223999023438, 25.984933853149414, 308.2494812011719, 56.42808532714844, -11.35118293762207, 154.69265747070312, 184.87777709960938, -54.945804595947266, 95.66385650634766, 245.59471130371094, -91.07244873046875, 130.08236694335938, 120.46060180664062, 59.319644927978516, 2.814014434814453, 23.173614501953125, 37.25300979614258, -5.06974983215332, 115.53836059570312, 60.39277267456055, -7.490612030029297, 64.91046142578125, -29.4459228515625, 194.35647583007812, -57.85403823852539, 214.2236328125, 156.46371459960938, -96.3001708984375, -35.92412567138672, 181.28216552734375, -76.64269256591797, 176.68719482421875, -54.17950439453125, 205.53782653808594, 305.62249755859375, 173.17648315429688, -2.9571075439453125, 181.2564239501953, 200.7752227783203, 120.78691101074219, 21.978504180908203, 83.2562255859375, 193.3979949951172, -10.844085693359375, -15.579818725585938, -8.709939956665039, -94.16625213623047, 119.70526123046875, 160.26173400878906, -24.0859375, 4.533926010131836, 25.97564697265625, 10.5733642578125, -82.04852294921875, 110.95036315917969, 143.19674682617188, 120.78949737548828, -53.632240295410156, -8.446685791015625, -0.6246128082275391, 86.29817199707031, 291.2269287109375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000154.npy"}
{"epoch": 0.2261380323054332, "step": 155, "batch_size": 64, "mean": 66.2703857421875, "std": 98.12701416015625, "min": -216.5780792236328, "p10": -25.977810096740715, "median": 63.93539810180664, "p90": 176.9959472656251, "max": 367.84466552734375, "pos_frac": 0.734375, "sample": [-8.749412536621094, 98.86222076416016, 62.136329650878906, -7.071258544921875, 100.6343994140625, 14.081565856933594, 81.06743621826172, -29.317214965820312, 77.10821533203125, 23.311450958251953, -8.263412475585938, 64.64297485351562, 105.15975189208984, -82.44659423828125, 112.95967864990234, 76.17572784423828, -32.11172866821289, 47.233673095703125, 62.1671142578125, -1.482879638671875, 79.23239135742188, 70.85684204101562, 47.493621826171875, 367.84466552734375, 189.15780639648438, 9.23567008972168, 139.765869140625, 105.49097442626953, 8.590713500976562, -127.59443664550781, 24.241798400878906, -1.7129058837890625, 25.853004455566406, 96.23318481445312, 194.74978637695312, -216.5780792236328, 292.79620361328125, 48.700828552246094, 136.38873291015625, -16.433578491210938, 2.3366432189941406, 97.74192810058594, 156.8738250732422, 45.21632385253906, 283.7483215332031, 99.38739013671875, 295.1438903808594, 157.14254760742188, -81.41015625, 102.9888916015625, -2.8723297119140625, 185.50454711914062, -15.373401641845703, 113.41783905029297, 91.35725402832031, -49.19951629638672, 63.227821350097656, 134.15347290039062, 125.68146514892578, 12.79458999633789, 139.15074157714844, -4.426321029663086, -18.18586540222168, 76.49353790283203], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000155.npy"}
{"epoch": 0.2276064610866373, "step": 156, "batch_size": 64, "mean": 37.68748474121094, "std": 111.34308624267578, "min": -286.2728271484375, "p10": -106.95597610473631, "median": 23.74689483642578, "p90": 140.9448669433594, "max": 297.10888671875, "pos_frac": 0.734375, "sample": [138.1591796875, 61.262672424316406, 297.10888671875, 265.18853759765625, 179.04750061035156, 116.34201049804688, -202.37985229492188, 3.5211944580078125, -91.45639038085938, 20.61095428466797, -204.6597137451172, 135.19833374023438, 102.94857025146484, -23.923233032226562, -65.38594055175781, -140.671142578125, 41.69297409057617, 113.2374267578125, 259.90399169921875, 81.09349060058594, 21.574050903320312, 123.17205810546875, 50.623199462890625, 15.711406707763672, 3.8662662506103516, -10.815071105957031, 15.254112243652344, 31.812721252441406, 261.4110107421875, -17.836753845214844, 247.7752685546875, 142.13873291015625, 123.75023651123047, 72.84981536865234, 14.234840393066406, 8.396123886108398, 78.98072052001953, -129.38595581054688, -286.2728271484375, -80.1227035522461, 83.3268814086914, -56.86865234375, 0.6441192626953125, 130.38185119628906, 14.107650756835938, -116.51873779296875, 85.59378814697266, 122.12389373779297, 80.01451110839844, 8.25096321105957, -4.434503555297852, 79.69035339355469, 60.924095153808594, 54.76360321044922, -113.5986557006836, -53.72547912597656, 10.084793090820312, 20.821121215820312, 3.94598388671875, -4.211082458496094, 25.91973876953125, 79.53972625732422, 4.763622283935547, 122.50277709960938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000156.npy"}
{"epoch": 0.2290748898678414, "step": 157, "batch_size": 64, "mean": 84.25444030761719, "std": 109.0613021850586, "min": -307.2183837890625, "p10": -28.489734268188474, "median": 75.74987030029297, "p90": 208.3482818603516, "max": 401.81646728515625, "pos_frac": 0.828125, "sample": [163.32009887695312, 170.54559326171875, 88.1652603149414, 199.80677795410156, 136.63095092773438, 10.270172119140625, 79.11662292480469, 147.5357666015625, 3.2830276489257812, 158.39401245117188, 58.11861801147461, 221.78277587890625, 52.217628479003906, 93.35873413085938, 235.5858154296875, 212.00892639160156, 151.65673828125, 108.38652038574219, 41.29209899902344, 162.07373046875, 8.118049621582031, 40.55036926269531, 72.38311767578125, -29.305522918701172, 115.61212158203125, 261.41741943359375, -4.3415069580078125, 32.98080825805664, -1.070037841796875, 123.66670227050781, -307.2183837890625, -39.182884216308594, 27.159164428710938, 53.93505859375, 122.48646545410156, 83.28814697265625, 262.6615905761719, 136.525390625, 109.14677429199219, -33.94450378417969, 31.86404037475586, 289.6916198730469, 37.38818359375, 64.17437744140625, 6.359439849853516, 178.73971557617188, 109.73497009277344, -26.586227416992188, 178.37234497070312, 132.63267517089844, 195.10198974609375, 53.77751159667969, -59.496822357177734, 21.193939208984375, 191.60430908203125, 159.97991943359375, 401.81646728515625, 55.599159240722656, 44.303131103515625, -146.02859497070312, 17.301361083984375, -65.11054992675781, 13.860946655273438, -22.40765380859375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000157.npy"}
{"epoch": 0.2305433186490455, "step": 158, "batch_size": 64, "mean": 89.3277587890625, "std": 137.23439025878906, "min": -192.04898071289062, "p10": -47.35351524353026, "median": 52.415103912353516, "p90": 286.50977478027346, "max": 493.7964782714844, "pos_frac": 0.71875, "sample": [8.633075714111328, 11.496267318725586, 45.485870361328125, 51.59105682373047, 52.34385681152344, -29.514732360839844, 33.09489059448242, 87.54886627197266, 39.03564453125, -54.6438102722168, 217.83209228515625, -116.45233154296875, 260.971435546875, 81.25117492675781, -13.652969360351562, -60.95317077636719, -15.454622268676758, 111.24263000488281, -11.268033981323242, 14.87689208984375, 235.06805419921875, -35.799346923828125, 291.67919921875, -192.04898071289062, 313.33221435546875, 75.29340362548828, 145.7918701171875, 21.125761032104492, 37.09112548828125, 12.465587615966797, 71.06390380859375, 87.98468780517578, -19.111297607421875, 158.838134765625, -56.484127044677734, -4.8315582275390625, 349.9071044921875, 106.6872329711914, -122.71659088134766, 350.44000244140625, 436.25982666015625, -5.2349700927734375, 68.49047088623047, 493.7964782714844, 17.21087646484375, 10.637962341308594, -52.305301666259766, 119.15655517578125, 243.03787231445312, 114.4595718383789, 394.96295166015625, 172.58761596679688, 103.42988586425781, 178.4292449951172, -14.94972038269043, 183.9102783203125, 111.69660949707031, -9.002958297729492, 28.634849548339844, 103.86133575439453, 274.4477844238281, 52.486351013183594, 171.38555908203125, -19.65302085876465], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000158.npy"}
{"epoch": 0.23201174743024963, "step": 159, "batch_size": 64, "mean": 61.21916198730469, "std": 108.86018371582031, "min": -213.44190979003906, "p10": -60.09433517456054, "median": 58.695518493652344, "p90": 198.6082275390625, "max": 325.8755798339844, "pos_frac": 0.75, "sample": [12.176803588867188, 45.31482696533203, 140.6570587158203, -49.685882568359375, 261.9113464355469, 144.34298706054688, -213.44190979003906, -40.42899703979492, 125.5089111328125, 175.03189086914062, 93.29716491699219, -61.04527282714844, 15.300542831420898, -25.4007568359375, 91.16683959960938, -74.58464813232422, -125.1016845703125, 54.662410736083984, 190.35580444335938, 150.833984375, 325.8755798339844, 9.65578842163086, -46.57875061035156, 17.676856994628906, 29.999008178710938, -76.13478088378906, 21.939422607421875, -50.3607177734375, 90.14424133300781, 58.92431640625, -21.33002471923828, 200.71871948242188, 88.56058502197266, 15.396610260009766, 124.37345123291016, -22.27724266052246, 58.46672058105469, -104.65142822265625, -4.036155700683594, 231.55795288085938, 149.86221313476562, 227.67971801757812, 64.27372741699219, 71.57115173339844, 0.9459762573242188, 283.8207092285156, 13.094396591186523, 1.5151290893554688, 74.72232818603516, 46.092742919921875, 138.821533203125, 29.41521453857422, -57.87548065185547, -169.44354248046875, 72.96882629394531, 105.99079895019531, 97.53813934326172, 94.61918640136719, 31.997453689575195, 312.946044921875, 193.68374633789062, 70.20330810546875, 74.4693603515625, 130.32183837890625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000159.npy"}
{"epoch": 0.23348017621145375, "step": 160, "batch_size": 64, "mean": 81.287353515625, "std": 132.5327606201172, "min": -151.7614288330078, "p10": -72.02797622680664, "median": 66.47000122070312, "p90": 249.08752746582033, "max": 475.508544921875, "pos_frac": 0.671875, "sample": [107.79865264892578, -11.206487655639648, 184.04656982421875, 156.99354553222656, -14.05801773071289, -73.73497009277344, 197.42129516601562, 185.58860778808594, 170.20977783203125, 12.5531005859375, -82.81321716308594, -9.06982421875, 145.9324951171875, 242.57260131835938, 122.86054229736328, -77.11955261230469, 41.30281448364258, -37.69721984863281, -148.3969268798828, 232.33290100097656, 284.9697265625, -121.97587585449219, 251.879638671875, -65.25275421142578, 146.02012634277344, 47.048431396484375, 4.533836364746094, -13.853500366210938, 475.508544921875, 67.73442077636719, 15.144668579101562, 65.20558166503906, -26.576171875, 405.18572998046875, -1.5673065185546875, 74.52127075195312, 109.9244613647461, 166.5179443359375, 378.64825439453125, 308.527587890625, 31.79839324951172, 222.28466796875, -151.7614288330078, 208.728759765625, 95.10086059570312, 15.8280029296875, -28.57328987121582, 19.67354393005371, 116.55261993408203, 116.94895935058594, -38.47368621826172, 34.11577606201172, 118.60758972167969, -68.04499053955078, 93.45211791992188, 81.06704711914062, -9.037790298461914, -0.5820388793945312, -99.5921401977539, 162.69773864746094, 32.677879333496094, 256.5471496582031, -40.902557373046875, 115.61647033691406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000160.npy"}
{"epoch": 0.23494860499265785, "step": 161, "batch_size": 64, "mean": 71.23098754882812, "std": 106.10053253173828, "min": -139.59942626953125, "p10": -49.559331512451166, "median": 50.1171932220459, "p90": 236.21326141357432, "max": 326.52203369140625, "pos_frac": 0.75, "sample": [76.9653549194336, 200.9473876953125, -50.79435729980469, 131.92263793945312, 305.8172912597656, -139.59942626953125, -54.217437744140625, -69.45211791992188, 98.29142761230469, -42.656394958496094, 45.025970458984375, 99.8153076171875, 81.37033081054688, -70.3553237915039, 3.7113380432128906, 122.33863830566406, 326.52203369140625, 85.58421325683594, 144.31094360351562, 271.39898681640625, 17.00677490234375, 55.20841598510742, -6.8231658935546875, -37.533016204833984, 35.70756530761719, -10.124645233154297, 133.69009399414062, 42.95384979248047, 281.6697692871094, 304.19970703125, 130.72853088378906, -34.596710205078125, -0.3874320983886719, -26.859516143798828, 245.0894775390625, 110.71837615966797, 33.86109924316406, -46.67760467529297, 17.692794799804688, 85.70912170410156, -13.989883422851562, 24.31127166748047, 10.055213928222656, 28.942893981933594, 16.22663116455078, 26.34161376953125, 212.636962890625, 77.76404571533203, 15.047637939453125, 147.68807983398438, 215.50209045410156, -106.18281555175781, 73.88389587402344, 178.23062133789062, 39.30060577392578, 41.612518310546875, 33.822391510009766, 127.53901672363281, 91.70024108886719, -89.59535217285156, 56.54360580444336, 263.0516052246094, 118.32432556152344, 71.8455581665039], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000161.npy"}
{"epoch": 0.23641703377386197, "step": 162, "batch_size": 64, "mean": 70.75540924072266, "std": 112.76998901367188, "min": -180.9485321044922, "p10": -44.68764877319335, "median": 61.04191970825195, "p90": 203.7344665527344, "max": 458.0642395019531, "pos_frac": 0.765625, "sample": [1.0352706909179688, 50.35133743286133, 106.6751708984375, 247.79283142089844, 59.60169982910156, 221.87997436523438, 30.504484176635742, 140.53700256347656, 46.046669006347656, -6.322847366333008, -21.994062423706055, 197.97967529296875, 69.94633483886719, -106.77791595458984, -40.83738708496094, -19.341690063476562, 164.8050537109375, 170.58474731445312, -4.8788909912109375, 126.683837890625, 80.42028045654297, 27.879562377929688, 60.24247741699219, 8.5430908203125, 20.143102645874023, 206.2008056640625, 132.76031494140625, -37.28028869628906, -180.9485321044922, 147.1316375732422, 103.58602142333984, -112.24479675292969, -15.771852493286133, 17.586137771606445, 153.84930419921875, 70.86941528320312, 236.87066650390625, 458.0642395019531, 64.60783386230469, 119.01110076904297, 85.0439224243164, 95.7845687866211, 158.80921936035156, 171.91708374023438, 152.8107147216797, 3.39215087890625, -46.33776092529297, 1.0560989379882812, 95.64447021484375, 43.73866271972656, -84.38584899902344, 172.90664672851562, 1.097412109375, -95.7063980102539, 391.29254150390625, 21.85607147216797, -113.77599334716797, 30.354385375976562, 80.37702941894531, 84.08478546142578, 61.84136199951172, -36.10472106933594, 39.40728759765625, 217.45065307617188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000162.npy"}
{"epoch": 0.23788546255506607, "step": 163, "batch_size": 64, "mean": 69.22859191894531, "std": 110.85074615478516, "min": -161.1259765625, "p10": -48.64385375976562, "median": 61.543121337890625, "p90": 259.52621459960943, "max": 353.49517822265625, "pos_frac": 0.671875, "sample": [2.2412033081054688, -9.9346923828125, 266.6434326171875, 353.49517822265625, -27.991653442382812, -123.48968505859375, 97.17871856689453, -10.061798095703125, -9.945091247558594, 78.066162109375, 133.30270385742188, -34.422080993652344, 106.38224029541016, 324.1036376953125, 108.94457244873047, 42.585601806640625, 99.96480560302734, -13.979034423828125, 15.817131042480469, -60.88379669189453, 57.786521911621094, 156.39419555664062, 106.45626831054688, -34.256507873535156, 103.07679748535156, -0.3513984680175781, 123.5509033203125, -161.1259765625, -68.8311996459961, 109.9094009399414, 265.624267578125, 13.534923553466797, 142.4461669921875, 13.000295639038086, 70.86344909667969, -10.751470565795898, -52.5648193359375, 302.80902099609375, 130.19110107421875, 72.27339172363281, -69.69564819335938, 35.36119842529297, -39.49493408203125, 245.29742431640625, 42.40000534057617, 106.7972183227539, -33.901893615722656, 291.8238525390625, 144.0098876953125, 30.94844627380371, 62.54341125488281, 110.21420288085938, 60.54283142089844, 188.23846435546875, 80.8554916381836, -22.138641357421875, 82.17064666748047, 14.293533325195312, -98.46163177490234, -1.3952274322509766, 277.0601501464844, 91.78421020507812, 164.14599609375, -10.822263717651367], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000163.npy"}
{"epoch": 0.2393538913362702, "step": 164, "batch_size": 64, "mean": 84.82752990722656, "std": 106.386962890625, "min": -155.25225830078125, "p10": -21.11440906524658, "median": 72.1297492980957, "p90": 207.7270568847657, "max": 437.939453125, "pos_frac": 0.828125, "sample": [-144.2556610107422, 167.70745849609375, 166.80999755859375, 169.98309326171875, 226.6632080078125, 54.69495391845703, 78.54643249511719, 130.11830139160156, 74.69686889648438, 77.56993865966797, 244.75888061523438, 48.855018615722656, 21.500871658325195, 51.475276947021484, -67.143798828125, 131.56082153320312, 65.33167266845703, 65.90701293945312, -21.436330795288086, 72.45162200927734, 6.834440231323242, 133.51187133789062, 74.19965362548828, 190.70672607421875, 29.257539749145508, 154.67701721191406, -28.568180084228516, -90.66626739501953, 151.15896606445312, 16.50904655456543, 42.48899841308594, -18.28301239013672, 154.30686950683594, 180.01693725585938, 145.1993408203125, 115.29731750488281, 48.66417694091797, 104.97061157226562, 24.76525115966797, 171.61448669433594, 38.677452087402344, 88.69776916503906, 66.5210189819336, 33.189697265625, 47.04029846191406, 4.7087249755859375, 71.80787658691406, 34.051902770996094, -155.25225830078125, 143.47238159179688, 77.96063232421875, 44.126708984375, 340.0191955566406, 6.7574462890625, 348.0130615234375, 247.17477416992188, -0.0356597900390625, 215.021484375, 87.46315002441406, -57.16893005371094, -20.363258361816406, 111.16351318359375, 437.939453125, -4.4820709228515625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000164.npy"}
{"epoch": 0.24082232011747431, "step": 165, "batch_size": 64, "mean": 55.120628356933594, "std": 101.55455017089844, "min": -154.37139892578125, "p10": -81.00681762695312, "median": 39.78447341918945, "p90": 206.0189025878907, "max": 282.71087646484375, "pos_frac": 0.703125, "sample": [96.37425994873047, 91.88074493408203, -84.78958129882812, 19.24866485595703, -87.4949951171875, -1.7376899719238281, 78.43595123291016, 114.78005981445312, -12.871612548828125, -81.46988677978516, 133.77200317382812, -154.37139892578125, -82.13497924804688, 34.86973571777344, 34.087249755859375, -93.44994354248047, -11.296817779541016, -50.728302001953125, 163.0079345703125, 152.85682678222656, 83.57420349121094, 26.248809814453125, 17.112796783447266, 282.71087646484375, -147.78627014160156, 37.351593017578125, -17.834707260131836, -49.876976013183594, -5.6090545654296875, 142.6603240966797, -79.92632293701172, 20.72230339050293, 218.1500244140625, 188.62289428710938, 46.085968017578125, -23.16999053955078, 83.58243560791016, 65.308837890625, 78.7422103881836, 4.858068466186523, 42.21735382080078, 65.38751983642578, 158.32630920410156, 77.5037612915039, 213.47433471679688, -17.987060546875, 97.13728332519531, 61.612579345703125, 4.614570617675781, -73.1412124633789, 34.86627197265625, 228.8123779296875, 28.82463836669922, -41.40845489501953, 63.3866081237793, 274.711181640625, 28.06340789794922, 272.43988037109375, 224.60195922851562, 123.53919982910156, 115.42816162109375, 16.049781799316406, 113.28175354003906, 185.48178100585938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000165.npy"}
{"epoch": 0.2422907488986784, "step": 166, "batch_size": 64, "mean": 101.58251953125, "std": 142.63392639160156, "min": -92.73457336425781, "p10": -34.90358619689941, "median": 67.75369262695312, "p90": 279.8194061279298, "max": 571.8861694335938, "pos_frac": 0.796875, "sample": [461.83026123046875, 323.69879150390625, 32.582645416259766, 501.7177734375, -37.33051300048828, 5.476276397705078, 21.03890609741211, 70.41128540039062, 177.13729858398438, 48.519622802734375, 129.43153381347656, -10.64239501953125, 11.664440155029297, 164.08944702148438, -68.71073913574219, 82.0640869140625, 160.959228515625, 19.449462890625, -6.120143890380859, 198.36026000976562, 2.909799575805664, 571.8861694335938, 45.54140090942383, -50.17928695678711, -55.55986022949219, -52.48611068725586, 43.23314666748047, 151.40103149414062, -92.73457336425781, 246.89157104492188, 293.93133544921875, 477.8447570800781, 118.5206069946289, 70.45408630371094, 28.043746948242188, 380.9342041015625, 15.567100524902344, 207.17880249023438, 32.195526123046875, 159.8867645263672, -15.337146759033203, 88.28938293457031, 19.899681091308594, 11.892433166503906, 178.94046020507812, 127.25056457519531, 223.34165954589844, -4.431140899658203, 71.22224426269531, 173.33575439453125, 32.67439651489258, 87.83983612060547, -63.70886993408203, 65.09609985351562, -6.530616760253906, -29.24075698852539, 116.90836334228516, 4.1134185791015625, 124.05233764648438, 5.9788970947265625, 92.29612731933594, 18.397985458374023, 206.55316162109375, 91.35881805419922], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000166.npy"}
{"epoch": 0.24375917767988253, "step": 167, "batch_size": 64, "mean": 107.80416870117188, "std": 119.12782287597656, "min": -135.10128784179688, "p10": -16.22882080078125, "median": 91.52064514160156, "p90": 267.00833740234384, "max": 440.4036865234375, "pos_frac": 0.828125, "sample": [103.86357116699219, 64.92172241210938, 146.83200073242188, 214.7903594970703, 115.59945678710938, 120.30300903320312, 96.60321044921875, 70.0675048828125, 100.18202209472656, 53.63020706176758, 440.4036865234375, 190.98944091796875, 74.06057739257812, -15.419570922851562, 29.521026611328125, 69.01548767089844, 86.43323516845703, -59.175254821777344, 134.712890625, 190.02142333984375, 23.96555519104004, 42.203948974609375, 84.48670959472656, 61.5364990234375, 107.34428405761719, -135.10128784179688, 278.1241455078125, 251.70318603515625, -16.43572998046875, 179.35726928710938, -11.796684265136719, 107.14199829101562, 79.70649719238281, 62.048423767089844, 383.01715087890625, 193.3677978515625, 277.9971008300781, 86.43807983398438, 213.64039611816406, -128.12330627441406, 237.7959747314453, 273.56768798828125, 333.6285705566406, 151.322998046875, 414.5592041015625, -15.74603271484375, 180.64920043945312, 146.02700805664062, 223.21316528320312, 180.5226593017578, 104.34407043457031, 10.440109252929688, -5.601552963256836, 34.69749069213867, 69.34049987792969, 137.166015625, 105.35352325439453, 58.219093322753906, 22.318557739257812, -100.5899887084961, 24.37630844116211, 13.591287612915039, -46.25914001464844, -21.447402954101562], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000167.npy"}
{"epoch": 0.24522760646108663, "step": 168, "batch_size": 64, "mean": 78.75118255615234, "std": 108.04371643066406, "min": -88.80908203125, "p10": -50.58443031311035, "median": 56.382524490356445, "p90": 203.0782928466797, "max": 435.19122314453125, "pos_frac": 0.796875, "sample": [21.657440185546875, 122.26006317138672, 25.447154998779297, 19.06804656982422, 3.864023208618164, 33.19378662109375, 142.451171875, -50.6694450378418, 22.337604522705078, 10.587013244628906, -88.80908203125, 195.4917755126953, 435.19122314453125, 96.81271362304688, -50.38606262207031, 201.20364379882812, 181.47821044921875, 243.1357879638672, 203.8817138671875, 49.281455993652344, 226.38058471679688, 282.9406433105469, 182.75790405273438, 182.99160766601562, -45.19683074951172, -51.65357971191406, -20.44145965576172, 133.1930694580078, 151.00344848632812, 138.49859619140625, 94.52287292480469, 30.67110824584961, -63.279449462890625, 98.91555786132812, 35.427520751953125, 116.36482238769531, -73.76594543457031, -11.798091888427734, -32.40829849243164, 35.981109619140625, 28.97947883605957, 92.91706085205078, 66.63079833984375, 96.98490905761719, 114.47270202636719, 69.98567199707031, 28.556474685668945, -63.89087677001953, 277.7694396972656, 60.824134826660156, 169.61656188964844, 20.289533615112305, 27.585268020629883, 27.502288818359375, 132.294677734375, 61.878143310546875, 29.456449508666992, -84.79888153076172, 51.940914154052734, -37.87190246582031, 386.3755798339844, 18.57550811767578, 95.38298034667969, 140.03530883789062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000168.npy"}
{"epoch": 0.24669603524229075, "step": 169, "batch_size": 64, "mean": 69.27072143554688, "std": 133.41368103027344, "min": -286.6116943359375, "p10": -63.793114471435544, "median": 42.665090560913086, "p90": 229.5330917358399, "max": 573.9267578125, "pos_frac": 0.75, "sample": [79.79566192626953, 92.10398864746094, 8.136837005615234, 17.813255310058594, 219.3890838623047, 116.94068145751953, 19.892166137695312, 160.9485626220703, 92.6612319946289, 36.744991302490234, -2.256824493408203, 42.68773651123047, -286.6116943359375, 3.670408248901367, 57.69810104370117, -24.695663452148438, 25.500213623046875, -83.07762145996094, 88.72653198242188, 174.00845336914062, 158.9617919921875, -59.5732421875, 67.12908935546875, -92.67646789550781, 217.2200164794922, 233.88052368164062, 66.71656799316406, 5.731964111328125, -17.008468627929688, 32.90245056152344, 215.3341064453125, -12.432670593261719, 14.497797012329102, 74.88362884521484, 95.68681335449219, 131.00279235839844, 257.8663635253906, 27.907503128051758, 69.00053405761719, 244.26145935058594, 8.669614791870117, 396.19586181640625, -126.3962631225586, 94.09286499023438, 4.862815856933594, 61.57026672363281, 27.77679443359375, -65.60163116455078, 351.0574035644531, -14.816421508789062, 573.9267578125, -2.327728271484375, 97.33179473876953, 15.534858703613281, 300.24407958984375, -11.040313720703125, 84.89144897460938, 19.184555053710938, 42.6424446105957, -30.798439025878906, -125.93550109863281, 96.06875610351562, 202.98899841308594, -138.16558837890625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000169.npy"}
{"epoch": 0.24816446402349487, "step": 170, "batch_size": 64, "mean": 81.49375915527344, "std": 110.62023162841797, "min": -230.27438354492188, "p10": -32.93245162963866, "median": 66.55783462524414, "p90": 233.10254821777346, "max": 340.5000915527344, "pos_frac": 0.828125, "sample": [195.0494384765625, -140.00588989257812, 143.6328125, 84.55310821533203, 234.24136352539062, 83.90998840332031, 128.22781372070312, 17.173927307128906, 147.2445068359375, 1.3136138916015625, 91.81183624267578, -6.782024383544922, 78.8380355834961, -34.99494934082031, 10.515802383422852, -52.211151123046875, 115.89556121826172, 246.608154296875, 15.2100830078125, 323.8155212402344, 4.705055236816406, -20.80182456970215, 177.0098419189453, 49.81081771850586, 2.4893417358398438, 21.31414031982422, -230.27438354492188, -28.119956970214844, 71.55001831054688, 203.49786376953125, 68.67716979980469, 130.47216796875, 58.27683639526367, 73.46159362792969, 184.9000244140625, -52.38520812988281, 319.18048095703125, 6.162788391113281, 9.712076187133789, 340.5000915527344, 76.18772888183594, -111.23123168945312, 63.062931060791016, 116.9417724609375, 205.66497802734375, -35.159996032714844, 26.93778419494629, 4.25407600402832, 41.73278045654297, 64.4384994506836, 93.62613677978516, 45.562278747558594, 239.7906494140625, 44.60482406616211, 137.22430419921875, 57.525421142578125, 188.77992248535156, 230.4453125, -18.865703582763672, 197.1444091796875, 239.27947998046875, 13.927682876586914, 198.8748779296875, 20.665348052978516], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000170.npy"}
{"epoch": 0.24963289280469897, "step": 171, "batch_size": 64, "mean": 64.94615173339844, "std": 138.068603515625, "min": -267.6822509765625, "p10": -95.99852752685545, "median": 55.06147575378418, "p90": 213.75576477050782, "max": 601.2410278320312, "pos_frac": 0.703125, "sample": [-87.77056884765625, 223.9895477294922, 19.10778045654297, 97.1397705078125, -105.23894500732422, -50.03704833984375, 4.856651306152344, 102.36709594726562, -80.10858154296875, 141.97988891601562, 11.52962875366211, 160.74542236328125, 79.10359191894531, 109.90280151367188, 53.87677001953125, -82.24110412597656, 100.44181823730469, -100.25642395019531, 111.47990417480469, -8.180488586425781, 207.67575073242188, 111.41026306152344, 50.68238830566406, 87.67186737060547, 340.3274230957031, 76.72735595703125, -101.54241180419922, 601.2410278320312, -99.52479553222656, 48.21083068847656, 56.24618148803711, -267.6822509765625, 446.22821044921875, 146.29437255859375, 97.71266174316406, -117.1844482421875, 212.24380493164062, 105.68882751464844, 9.278060913085938, 319.6998291015625, 0.493194580078125, 67.7496109008789, 214.40374755859375, -28.013654708862305, -101.00480651855469, 11.97247314453125, 67.62103271484375, 86.20478820800781, -59.06643295288086, 112.68330383300781, -27.40843963623047, -57.18792724609375, 142.67898559570312, 43.780181884765625, 78.39181518554688, -20.19853401184082, 34.798980712890625, 162.99838256835938, -18.687156677246094, -21.960723876953125, 92.72783660888672, 17.837858200073242, 271.24383544921875, 50.4033317565918], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000171.npy"}
{"epoch": 0.2511013215859031, "step": 172, "batch_size": 64, "mean": 70.02001953125, "std": 117.96954345703125, "min": -208.86138916015625, "p10": -72.81891021728515, "median": 76.39971160888672, "p90": 203.46394042968754, "max": 512.1746215820312, "pos_frac": 0.703125, "sample": [-65.15814208984375, 101.11986541748047, -109.212646484375, 91.629638671875, 38.51524353027344, -49.89196014404297, 149.6830291748047, 104.13883972167969, 76.32576751708984, -79.60124969482422, -135.49252319335938, 53.83584213256836, 57.531715393066406, -76.10209655761719, 156.20462036132812, 90.5185546875, 19.53536605834961, 73.74089050292969, -53.43369674682617, 76.4736557006836, -32.52049255371094, 93.87884521484375, 195.49533081054688, 151.2327117919922, 21.8240966796875, 75.64019012451172, 232.19369506835938, 206.87905883789062, -78.33670043945312, 257.3531494140625, -27.81114959716797, 114.48273468017578, 99.24413299560547, 95.01455688476562, 122.88533020019531, 310.9400634765625, 512.1746215820312, 43.31352996826172, 133.7667236328125, 28.633060455322266, -22.151107788085938, -43.675987243652344, 78.61386108398438, 80.58644104003906, 187.1920166015625, -25.019317626953125, 10.83772087097168, 23.270353317260742, -29.86907958984375, 178.33740234375, 164.00205993652344, -2.029022216796875, -58.22869873046875, 102.12298583984375, 110.2682876586914, -208.86138916015625, 72.55497741699219, 105.39248657226562, 234.13177490234375, 102.83447265625, -15.922866821289062, -81.85240936279297, 253.20782470703125, 188.89410400390625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000172.npy"}
{"epoch": 0.2525697503671072, "step": 173, "batch_size": 64, "mean": 69.5058822631836, "std": 100.24758911132812, "min": -195.70428466796875, "p10": -28.76169624328613, "median": 53.690866470336914, "p90": 216.55238647460945, "max": 355.63568115234375, "pos_frac": 0.796875, "sample": [42.597686767578125, 78.46659088134766, 18.71271324157715, -4.862358093261719, 8.596691131591797, 54.15470886230469, 0.8823280334472656, 145.59742736816406, 121.58927917480469, 8.160123825073242, -25.358551025390625, -44.50550842285156, 16.2435245513916, 0.01320648193359375, 119.26248168945312, 272.340087890625, 71.35198211669922, 60.2303352355957, 72.13333892822266, 197.406982421875, 118.4615478515625, 89.92236328125, 87.15454864501953, -54.357765197753906, 54.35765838623047, 27.180335998535156, 355.63568115234375, 28.357872009277344, 165.29150390625, 223.7938232421875, 67.4185562133789, 18.318418502807617, -59.992401123046875, 116.57764434814453, 104.46929931640625, -26.905197143554688, 124.56696319580078, 2.83380126953125, -1.881805419921875, 233.35797119140625, 58.5637092590332, 199.65570068359375, -29.55733871459961, 40.93498992919922, -8.938758850097656, 14.538097381591797, 242.43826293945312, 121.8214111328125, -43.398284912109375, -0.6380157470703125, 96.64002227783203, 50.82435989379883, 133.7353057861328, -69.33282470703125, -195.70428466796875, 348.07763671875, 35.80668640136719, 5.335086822509766, 100.18975830078125, 19.535808563232422, 283.20574951171875, 61.39408874511719, 53.22702407836914, 42.44854736328125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000173.npy"}
{"epoch": 0.2540381791483113, "step": 174, "batch_size": 64, "mean": 81.2957763671875, "std": 121.37774658203125, "min": -157.71910095214844, "p10": -53.29281005859375, "median": 62.63615417480469, "p90": 263.7418777465822, "max": 464.6582946777344, "pos_frac": 0.78125, "sample": [21.673324584960938, -9.701568603515625, -52.236412048339844, -95.76811218261719, -5.176605224609375, 25.04290008544922, -18.37814712524414, 9.038068771362305, 119.00230407714844, 211.57200622558594, 227.90724182128906, 1.1259384155273438, 65.0740966796875, 19.18682098388672, 57.896263122558594, -53.74555206298828, 104.35319519042969, 165.84910583496094, 149.509033203125, 279.0995788574219, 62.61189270019531, 33.982662200927734, 62.66041564941406, -18.06101417541504, 330.6853332519531, -84.31742858886719, 140.85235595703125, 310.6323547363281, 34.30519104003906, 95.94300842285156, 12.794157028198242, -16.078216552734375, 62.93762969970703, -81.693359375, 145.46923828125, -51.352691650390625, 174.96163940429688, 111.28710174560547, -123.68533325195312, 293.86090087890625, 314.59820556640625, 336.6931457519531, 33.966060638427734, 123.18212890625, 19.98479652404785, -157.71910095214844, 3.3236637115478516, 141.84371948242188, 119.97648620605469, 153.189697265625, 83.55183410644531, 49.38703918457031, 23.18010711669922, 141.78302001953125, -96.69776153564453, 182.8703155517578, 155.6698760986328, 94.23643493652344, 47.414146423339844, 13.38125228881836, 464.6582946777344, 133.18576049804688, 70.75511169433594, 31.39601707458496], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000174.npy"}
{"epoch": 0.2555066079295154, "step": 175, "batch_size": 64, "mean": 112.4501953125, "std": 112.17969512939453, "min": -91.09127807617188, "p10": -7.784081077575682, "median": 89.78525924682617, "p90": 249.64591064453128, "max": 435.1322021484375, "pos_frac": 0.828125, "sample": [22.510831832885742, 38.35919189453125, 58.010841369628906, -91.09127807617188, 87.05232238769531, 161.22067260742188, 234.42279052734375, 89.47089385986328, 24.103090286254883, 146.94534301757812, 324.97393798828125, -77.36769104003906, -10.899093627929688, 322.52813720703125, 123.7320556640625, 38.054229736328125, 224.16363525390625, 231.1555938720703, 251.8563232421875, 164.42840576171875, 180.87448120117188, -70.2127914428711, 203.8728485107422, 435.1322021484375, 43.17631530761719, 137.973388671875, 341.19329833984375, 50.80480194091797, 136.62844848632812, 85.48949432373047, -8.210624694824219, 57.182525634765625, 320.6546325683594, 18.479022979736328, 6.182716369628906, 37.84873962402344, -0.7693061828613281, 278.04071044921875, 8.987211227416992, 86.91133117675781, 124.9102783203125, 117.23780822753906, 126.50884246826172, 227.92129516601562, -0.8995685577392578, 227.21246337890625, 244.48828125, 190.6638641357422, -19.05783462524414, 108.9531478881836, 176.93215942382812, 93.97606658935547, 90.09962463378906, 152.41574096679688, 43.450035095214844, 74.0338134765625, 146.13841247558594, -6.788812637329102, 45.49303436279297, 242.36593627929688, -45.343116760253906, 70.08538818359375, -2.8645477294921875, 55.01104736328125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000175.npy"}
{"epoch": 0.25697503671071953, "step": 176, "batch_size": 64, "mean": 93.65642547607422, "std": 125.18771362304688, "min": -68.66285705566406, "p10": -39.68218917846679, "median": 66.05595397949219, "p90": 234.20401763916016, "max": 522.5595703125, "pos_frac": 0.765625, "sample": [4.2080078125, 25.8092041015625, 91.83309936523438, 215.66831970214844, 72.91566467285156, 37.877506256103516, 11.434829711914062, 326.8672180175781, 19.424131393432617, -26.71319580078125, -42.59220886230469, 493.6839904785156, -11.450704574584961, -68.66285705566406, 235.02413940429688, 69.58609008789062, 27.942474365234375, 233.31700134277344, 185.27301025390625, -4.07415771484375, -5.2204437255859375, 73.01432037353516, 149.85581970214844, -47.20161437988281, 234.58416748046875, 3.6386337280273438, 29.5194034576416, 196.3384552001953, 185.9967041015625, 9.074958801269531, 116.99697875976562, -64.04412841796875, 150.16925048828125, -50.433387756347656, 40.73748779296875, 178.32049560546875, -46.539730072021484, 84.29261016845703, 177.92413330078125, 142.30958557128906, 194.39743041992188, 0.19398117065429688, -46.97306823730469, -2.272602081298828, 156.56832885742188, 141.37124633789062, 131.25111389160156, 72.37333679199219, 144.15423583984375, 294.405517578125, 36.551063537597656, -32.89214324951172, 26.3876953125, 14.978919982910156, 138.95755004882812, 62.52581787109375, -15.573974609375, 7.163173675537109, 134.5334014892578, -0.5336875915527344, 178.13186645507812, 522.5595703125, 14.594589233398438, 364.452392578125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000176.npy"}
{"epoch": 0.25844346549192365, "step": 177, "batch_size": 64, "mean": 94.046875, "std": 123.30945587158203, "min": -113.93650817871094, "p10": -27.212549018859857, "median": 68.71881484985352, "p90": 249.3391265869141, "max": 608.3717041015625, "pos_frac": 0.75, "sample": [173.1397705078125, 49.091888427734375, 241.0896453857422, 128.71231079101562, 67.14900207519531, 129.17727661132812, -113.93650817871094, 104.55409240722656, 205.25534057617188, 172.95767211914062, 261.3624267578125, 45.936363220214844, 96.56185913085938, 265.27032470703125, 134.5557403564453, 81.13719177246094, 16.797096252441406, 373.0371398925781, 30.446205139160156, 173.51422119140625, 70.28862762451172, 200.73353576660156, -78.70106506347656, 153.48023986816406, -1.9592514038085938, 46.392364501953125, 198.11517333984375, -21.95025634765625, -4.980224609375, 4.8179931640625, 64.2525405883789, 298.02386474609375, 113.153564453125, -21.263992309570312, 29.407470703125, 82.14775085449219, -19.393814086914062, 252.87461853027344, 1.6190662384033203, -30.37972068786621, -74.76193237304688, 152.600341796875, -4.823390960693359, 54.453895568847656, -84.05876159667969, 159.2827911376953, 132.1429901123047, 17.493213653564453, -9.773307800292969, -4.760414123535156, 210.79417419433594, 22.87285041809082, 84.49189758300781, -16.28697967529297, 184.1942138671875, 35.389339447021484, 38.56659698486328, -29.467817306518555, 171.19729614257812, 44.519805908203125, 608.3717041015625, 303.1412353515625, 84.75643157958984, -33.823486328125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000177.npy"}
{"epoch": 0.2599118942731278, "step": 178, "batch_size": 64, "mean": 117.29739379882812, "std": 145.477294921875, "min": -140.68675231933594, "p10": -32.018553924560536, "median": 104.08330535888672, "p90": 325.4012847900392, "max": 581.9031982421875, "pos_frac": 0.78125, "sample": [60.60451126098633, 12.813789367675781, 162.682861328125, -15.409393310546875, 147.8253173828125, 92.90946197509766, 112.64473724365234, 44.991004943847656, 21.128585815429688, 291.61773681640625, 1.3433303833007812, 117.59255981445312, 292.4617004394531, 108.71746826171875, 51.627296447753906, 163.6171112060547, 412.0166931152344, 122.56033325195312, 581.9031982421875, -58.60841369628906, 221.0731658935547, 61.33362579345703, 14.050445556640625, 66.1437759399414, -2.568056106567383, 177.92970275878906, 387.2687683105469, -50.373680114746094, 50.56353759765625, -4.463035583496094, -74.42630004882812, 102.24014282226562, 220.269287109375, 23.351333618164062, 141.53363037109375, 194.2666778564453, 139.131591796875, -102.31866455078125, 66.10498809814453, 267.4357604980469, 32.56382369995117, 114.17999267578125, -140.68675231933594, 447.3461608886719, -10.557716369628906, 159.43106079101562, 118.98074340820312, 170.24819946289062, -36.96619415283203, 455.9903869628906, 37.86908721923828, 141.8717041015625, 352.0359802246094, -5.367792129516602, 17.9620361328125, 220.73374938964844, 105.92646789550781, -20.47406005859375, 65.91700744628906, 226.41586303710938, 339.51824951171875, -89.958740234375, 191.25839233398438, -10.791007995605469], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000178.npy"}
{"epoch": 0.26138032305433184, "step": 179, "batch_size": 64, "mean": 81.23482513427734, "std": 142.68878173828125, "min": -291.072021484375, "p10": -83.96157989501953, "median": 79.37912368774414, "p90": 304.3692810058594, "max": 396.7841491699219, "pos_frac": 0.6875, "sample": [118.5461654663086, 165.21221923828125, 181.7117919921875, 396.7841491699219, 65.4128189086914, -97.6640625, 184.6760711669922, 21.699825286865234, -9.721580505371094, 13.040887832641602, 86.7357177734375, 76.54415130615234, 45.936195373535156, 100.75321197509766, -13.049568176269531, 91.84320068359375, -85.24017333984375, 312.743408203125, -55.103878021240234, -52.230072021484375, 78.45317840576172, -80.97819519042969, 354.3862609863281, 25.762849807739258, 229.30625915527344, 124.66131591796875, 167.12637329101562, 315.4151611328125, 175.14205932617188, 40.28912353515625, 79.14503479003906, 294.1989440917969, 353.81561279296875, 96.11715698242188, 84.64520263671875, 132.0680389404297, -36.600669860839844, -34.95729446411133, -0.6787185668945312, 95.75650787353516, 82.48413848876953, -36.19891357421875, 79.61321258544922, -102.08055114746094, 125.1181640625, 335.7388610839844, -91.56439971923828, 236.99000549316406, 116.98029327392578, -120.76776885986328, -53.28392791748047, 248.57029724121094, 28.488296508789062, 252.8270263671875, 308.7279968261719, 219.19947814941406, -179.4642333984375, 142.8516845703125, -59.690757751464844, -68.90357971191406, -291.072021484375, 21.7100830078125, 24.063989639282227, -63.013328552246094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000179.npy"}
{"epoch": 0.26284875183553597, "step": 180, "batch_size": 64, "mean": 106.05424499511719, "std": 128.54905700683594, "min": -206.5408477783203, "p10": -25.49804935455322, "median": 88.55989837646484, "p90": 271.52906494140626, "max": 521.8787841796875, "pos_frac": 0.828125, "sample": [113.9263687133789, 168.97259521484375, -78.73887634277344, 3.1842308044433594, -70.3562240600586, 521.8787841796875, -22.67733383178711, 73.95285034179688, 105.69075012207031, 64.45515441894531, 4.512981414794922, 245.1737060546875, -206.5408477783203, 141.4424285888672, 179.97450256347656, 126.74566650390625, 85.45976257324219, 88.6456298828125, 106.43706512451172, 66.90669250488281, 170.24017333984375, -153.17184448242188, 70.36489868164062, 304.4981994628906, 74.7557144165039, 221.19265747070312, 169.88232421875, 4.669364929199219, 40.551513671875, 54.00827407836914, 286.9720458984375, 1.6231670379638672, 246.19552612304688, 164.41949462890625, 81.43929290771484, -4.96870231628418, 68.62235260009766, 214.33053588867188, -17.76019287109375, 188.1312255859375, 91.11112976074219, 20.92139434814453, 246.42391967773438, 59.2877197265625, -128.59954833984375, 88.47416687011719, 1.82635498046875, 277.9931640625, 250.2080535888672, 92.61499786376953, 150.11358642578125, 266.9468078613281, -26.6103572845459, 338.68707275390625, 80.6949234008789, -22.902664184570312, 203.0123291015625, 165.69459533691406, 29.34107208251953, -76.34867095947266, 273.4928894042969, 278.95208740234375, 28.995567321777344, 192.0994873046875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000180.npy"}
{"epoch": 0.2643171806167401, "step": 181, "batch_size": 64, "mean": 65.62828063964844, "std": 118.03181457519531, "min": -215.48529052734375, "p10": -59.546380615234376, "median": 53.186079025268555, "p90": 210.15459442138683, "max": 396.52215576171875, "pos_frac": 0.765625, "sample": [136.4520721435547, -59.908233642578125, 16.57288932800293, 57.93767166137695, 120.0165023803711, -3.1478748321533203, 32.54636001586914, 35.91847229003906, 88.2480697631836, 120.34358215332031, -2.4928016662597656, 18.40349006652832, 4.206449508666992, 9.403556823730469, -156.73501586914062, 232.06707763671875, 119.09361267089844, -28.13243865966797, 82.82781982421875, -14.850170135498047, 62.49483871459961, -58.702056884765625, -16.275909423828125, 223.1116180419922, -72.71768188476562, 88.97791290283203, 137.47171020507812, 289.35968017578125, 48.434486389160156, 59.528038024902344, 130.97402954101562, 81.03331756591797, -206.6938018798828, 46.15898132324219, 22.260116577148438, 87.31889343261719, 12.747127532958984, 178.3914337158203, 139.51412963867188, -70.21630859375, 14.480644226074219, -215.48529052734375, 231.5534210205078, -18.231441497802734, 357.5819091796875, 170.98939514160156, -169.4733428955078, 62.54692840576172, -33.31697082519531, 137.18890380859375, 176.00294494628906, 396.52215576171875, 151.18621826171875, 37.66144943237305, 134.5555419921875, 7.117191314697266, 4.703474044799805, 3.3965225219726562, 294.65655517578125, 62.60725784301758, 29.63860321044922, 179.92153930664062, 150.23159790039062, 42.23297882080078], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000181.npy"}
{"epoch": 0.2657856093979442, "step": 182, "batch_size": 64, "mean": 115.38099670410156, "std": 138.86439514160156, "min": -252.74520874023438, "p10": -56.10882148742674, "median": 105.48082733154297, "p90": 282.7886383056641, "max": 391.7887878417969, "pos_frac": 0.84375, "sample": [299.95843505859375, 76.83087158203125, 332.05322265625, 198.15866088867188, 91.50164794921875, 230.9412078857422, 239.28781127929688, 270.50653076171875, 358.84222412109375, 77.97503662109375, 131.07276916503906, 46.61249923706055, 31.15669059753418, -19.09427261352539, 272.18841552734375, -252.74520874023438, 256.2695007324219, 64.50794982910156, 4.983419418334961, 52.35997009277344, 388.89459228515625, 1.953338623046875, 264.9808349609375, 119.46000671386719, 155.83786010742188, -63.65264892578125, 254.51971435546875, 90.06008911132812, 14.23532485961914, 69.87381744384766, 152.769775390625, 273.5181884765625, -112.66829681396484, -127.7582015991211, 169.75430297851562, 249.99063110351562, 47.46165466308594, 141.09884643554688, 213.9219970703125, 228.6370849609375, -70.84612274169922, 77.65852355957031, 32.35142517089844, 132.37815856933594, 166.4259490966797, 182.27969360351562, 190.7009735107422, 135.07217407226562, -129.7160186767578, 4.073528289794922, 391.7887878417969, 150.95477294921875, 54.33758544921875, 20.621599197387695, -30.910179138183594, 9.843391418457031, -127.95618438720703, 68.2065200805664, 286.7616882324219, 19.98052215576172, 355.2628173828125, -38.50655746459961, 47.92369842529297, 159.4404296875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000182.npy"}
{"epoch": 0.26725403817914833, "step": 183, "batch_size": 64, "mean": 62.40069580078125, "std": 111.22071838378906, "min": -161.07322692871094, "p10": -87.49653091430663, "median": 60.18310737609863, "p90": 198.53525848388676, "max": 342.3900146484375, "pos_frac": 0.765625, "sample": [25.890853881835938, 76.58699798583984, -12.536115646362305, 67.7770004272461, 13.78927993774414, 171.8160858154297, 142.04432678222656, 146.80511474609375, 44.72627258300781, 59.10938262939453, -83.49900817871094, 133.76931762695312, 69.74308776855469, 17.9216251373291, 28.3033447265625, 93.27123260498047, 342.3900146484375, 235.94204711914062, 142.31517028808594, 33.156944274902344, 62.055458068847656, 118.27601623535156, 304.9510498046875, 102.67124938964844, -10.803031921386719, 6.298336029052734, 78.39639282226562, 60.72208023071289, 7.44580078125, -161.07322692871094, 189.75299072265625, 87.2549057006836, -54.47767639160156, 13.160747528076172, 177.08026123046875, 59.644134521484375, 43.38127136230469, 116.69710540771484, 141.1075439453125, 239.7645263671875, -141.61537170410156, -32.419532775878906, 340.401123046875, -77.55872344970703, 28.936302185058594, -35.37400817871094, 84.29521179199219, -98.24346923828125, -110.80449676513672, -111.21963500976562, 165.89788818359375, 121.0743408203125, 202.29908752441406, 203.47137451171875, 87.82963562011719, 51.054298400878906, 27.113182067871094, -89.20975494384766, 148.1502685546875, 86.36006164550781, -150.3192138671875, -65.87800598144531, 23.568634033203125, 4.20654296875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000183.npy"}
{"epoch": 0.2687224669603524, "step": 184, "batch_size": 64, "mean": 66.34819030761719, "std": 115.09642791748047, "min": -162.876708984375, "p10": -45.81667556762695, "median": 34.039926528930664, "p90": 215.9319549560547, "max": 349.9361877441406, "pos_frac": 0.75, "sample": [24.602081298828125, -104.42426300048828, 68.52701568603516, 77.24272918701172, 196.75477600097656, 194.7861328125, 1.9205474853515625, 14.031585693359375, 165.23321533203125, -16.71080780029297, 325.58349609375, 349.9361877441406, 65.59632873535156, 222.23623657226562, -45.60076141357422, 204.11688232421875, 10.952896118164062, 217.02581787109375, 104.50154876708984, 291.50457763671875, -17.951663970947266, 310.0296630859375, -85.95980834960938, 19.569732666015625, 40.526859283447266, 39.05250549316406, 42.17472839355469, 1.7989883422851562, 91.7265625, 3.3036632537841797, 40.00184631347656, 99.74105834960938, 154.3477020263672, 202.0749053955078, -94.36433410644531, 213.37960815429688, 14.896148681640625, -19.11156463623047, -38.378231048583984, 22.411540985107422, 72.99059295654297, -21.115615844726562, 34.376216888427734, 6.632516860961914, 21.131568908691406, 207.2610626220703, 71.9748764038086, 55.836238861083984, -45.909210205078125, -129.62057495117188, -53.56942367553711, 22.790937423706055, -6.846408843994141, 0.6340560913085938, -162.876708984375, 31.045669555664062, -39.01310729980469, 179.96646118164062, 33.703636169433594, 108.65372467041016, 11.203102111816406, -2.8396987915039062, 99.76118469238281, 343.0271911621094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000184.npy"}
{"epoch": 0.2701908957415565, "step": 185, "batch_size": 64, "mean": 51.844139099121094, "std": 122.27976989746094, "min": -194.59713745117188, "p10": -63.06753158569335, "median": 26.014623641967773, "p90": 208.57815399169928, "max": 437.0908508300781, "pos_frac": 0.609375, "sample": [-166.1329345703125, -66.09855651855469, -19.49443244934082, 84.29592895507812, 79.12001037597656, 16.44089698791504, -4.362176895141602, 79.06501770019531, 402.6507263183594, -1.3254318237304688, 25.947261810302734, 109.45308685302734, 106.47512817382812, -6.6669921875, 265.0206604003906, 45.966064453125, -10.026674270629883, 41.21778106689453, 138.93275451660156, -107.21841430664062, -10.71774673461914, 38.24531936645508, 25.204669952392578, 83.80287170410156, 73.62246704101562, -7.229454040527344, -110.71551513671875, 25.463973999023438, 81.53235626220703, -132.4757080078125, 33.783599853515625, 23.058666229248047, 96.67557525634766, 100.85551452636719, 26.081985473632812, 362.10198974609375, -55.995140075683594, 32.74195098876953, 196.23333740234375, -45.04367446899414, 35.42866134643555, 3.138143539428711, -21.87671661376953, -13.440719604492188, -21.93575096130371, -6.60411262512207, 321.6872253417969, 140.33285522460938, 144.853515625, -1.0018978118896484, 437.0908508300781, -30.46851348876953, 25.0545654296875, 70.27232360839844, -12.808441162109375, 213.86878967285156, 39.16734313964844, 125.11332702636719, -194.59713745117188, 271.1769104003906, -78.89266204833984, 89.42481231689453, -12.321365356445312, -55.12385940551758], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000185.npy"}
{"epoch": 0.27165932452276065, "step": 186, "batch_size": 64, "mean": 98.52206420898438, "std": 121.43245697021484, "min": -202.66583251953125, "p10": -22.7747444152832, "median": 81.44288635253906, "p90": 242.18587951660163, "max": 465.56317138671875, "pos_frac": 0.8125, "sample": [-202.66583251953125, 130.81971740722656, 312.9986877441406, 156.9069366455078, 158.38717651367188, -5.423065185546875, 81.61529541015625, 48.043270111083984, 214.59616088867188, 143.85415649414062, 64.9324951171875, 222.72360229492188, 1.1656265258789062, 87.40084838867188, 40.3567008972168, -18.111103057861328, 96.18348693847656, -120.86325073242188, 43.5818977355957, 53.594993591308594, 312.81536865234375, 51.088871002197266, -8.837516784667969, 178.77481079101562, 40.59130096435547, 81.27047729492188, 98.06907653808594, 146.44503784179688, 336.0484619140625, 102.05775451660156, 179.4642333984375, -33.865394592285156, 156.53733825683594, 76.09037780761719, 30.77351951599121, 127.6253662109375, 404.88775634765625, 250.52685546875, 49.459232330322266, 6.010700225830078, 165.83590698242188, 34.73362731933594, 74.80806732177734, 65.0633544921875, -80.7646484375, 413.7471923828125, -21.464706420898438, 69.98807525634766, -23.962295532226562, -39.72834777832031, 465.56317138671875, 160.98526000976562, 72.96450805664062, 148.76144409179688, 0.662750244140625, 149.72268676757812, 94.77024841308594, -23.33618927001953, 5.062934875488281, 122.21990966796875, 145.62266540527344, 128.3345489501953, -7.080596923828125, 86.970703125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000186.npy"}
{"epoch": 0.27312775330396477, "step": 187, "batch_size": 64, "mean": 83.17179107666016, "std": 109.75000762939453, "min": -152.88888549804688, "p10": -38.695597076416014, "median": 66.0941162109375, "p90": 222.37105712890627, "max": 404.1824951171875, "pos_frac": 0.765625, "sample": [-152.88888549804688, 207.9837646484375, -24.990293502807617, 60.24049758911133, -73.61859130859375, 236.26702880859375, 85.641357421875, 162.20166015625, 54.63695526123047, 37.06061935424805, 73.93630981445312, 65.80690002441406, 349.9906005859375, 84.4856185913086, 123.05976867675781, 213.123291015625, 164.9678497314453, 66.38133239746094, 116.13546752929688, -40.4279899597168, 61.07029724121094, 41.73878479003906, 404.1824951171875, -35.431793212890625, 169.2366485595703, 55.92718505859375, 235.3291015625, -37.5365104675293, 197.1328582763672, 44.83692932128906, -44.62420654296875, 81.78862762451172, 22.420494079589844, -59.028076171875, -39.19234848022461, 167.32687377929688, -30.169044494628906, 9.184080123901367, -8.738899230957031, 224.10684204101562, 153.62876892089844, 354.42254638671875, 95.90283966064453, 43.68102264404297, 67.60491943359375, 34.63833236694336, 72.36399841308594, 44.15827178955078, -125.09131622314453, 46.44811248779297, 111.70398712158203, -15.253273010253906, 169.2928924560547, 93.79512023925781, 78.19669342041016, 200.49302673339844, 87.6552734375, -2.5834732055664062, 56.785125732421875, 234.64549255371094, 218.32089233398438, 38.744346618652344, -25.441152572631836, 19.328487396240234], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000187.npy"}
{"epoch": 0.2745961820851689, "step": 188, "batch_size": 64, "mean": 100.59088134765625, "std": 123.19782257080078, "min": -132.08212280273438, "p10": -19.829008293151848, "median": 76.29676818847656, "p90": 293.8066101074219, "max": 508.1845703125, "pos_frac": 0.828125, "sample": [312.64031982421875, 126.11576080322266, 253.72157287597656, 71.31515502929688, 83.60973358154297, 0.119964599609375, -27.6990966796875, -121.62049102783203, 295.82049560546875, 79.3312759399414, 55.198638916015625, -27.54497528076172, 116.1124038696289, 79.95532989501953, 289.1075439453125, 253.1687774658203, 93.6511001586914, -23.943809509277344, 58.63151550292969, 9.329265594482422, 7.8221588134765625, 31.169517517089844, 37.975830078125, 142.46133422851562, 73.05826568603516, 26.450401306152344, 9.711227416992188, -8.36600112915039, 338.8362121582031, 49.811279296875, 364.92724609375, -132.08212280273438, 297.6167907714844, 126.67393493652344, 160.7911376953125, 120.90264892578125, 59.29998779296875, 198.20652770996094, -9.571510314941406, 203.197265625, 167.22499084472656, 12.496200561523438, -87.53936767578125, -22.751516342163086, 204.2007598876953, 41.66545867919922, 219.12307739257812, 40.20437240600586, 14.975341796875, 334.6671142578125, 85.55614471435547, 172.16387939453125, 56.73160171508789, -10.769813537597656, 173.6339874267578, 85.95372009277344, 99.59587860107422, 508.1845703125, 95.96426391601562, 74.74301147460938, 77.85052490234375, -13.009822845458984, 13.218544006347656, 17.820646286010742], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000188.npy"}
{"epoch": 0.27606461086637296, "step": 189, "batch_size": 64, "mean": 82.26558685302734, "std": 139.37933349609375, "min": -177.12664794921875, "p10": -71.23078994750976, "median": 62.07916069030762, "p90": 206.6018920898438, "max": 587.256103515625, "pos_frac": 0.78125, "sample": [41.168495178222656, 182.54290771484375, -120.60383605957031, 365.40142822265625, -73.46868896484375, -66.00902557373047, 44.22108459472656, 8.059211730957031, 0.7049465179443359, 37.941986083984375, 156.2005157470703, 145.74139404296875, 147.83193969726562, -0.98443603515625, 169.628662109375, 48.428741455078125, 189.89846801757812, 51.179195404052734, -177.12664794921875, 5.924722671508789, -93.95738983154297, 28.211183547973633, 84.27096557617188, 189.37820434570312, 147.3846893310547, 83.05267333984375, -48.270782470703125, 59.975650787353516, 36.5611457824707, 197.20901489257812, 6.142704010009766, 22.426504135131836, 98.01165771484375, 509.8167724609375, 63.671783447265625, -28.10797119140625, 26.87310791015625, 227.4724884033203, 82.96501159667969, 13.947364807128906, -175.3091278076172, 125.82816314697266, -64.16766357421875, 129.83175659179688, 60.48653793334961, 350.705078125, 169.44496154785156, 587.256103515625, -20.023849487304688, 125.59622192382812, 176.0374755859375, 31.628135681152344, 82.29843139648438, 162.12855529785156, 114.44261169433594, 210.62741088867188, 252.12103271484375, 93.36216735839844, -7.573944091796875, -174.94741821289062, 66.00788879394531, -86.5483169555664, 25.577301025390625, 166.47210693359375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000189.npy"}
{"epoch": 0.2775330396475771, "step": 190, "batch_size": 64, "mean": 97.31863403320312, "std": 144.92327880859375, "min": -220.8607635498047, "p10": -65.72489852905274, "median": 69.00610733032227, "p90": 320.6539611816407, "max": 443.2001647949219, "pos_frac": 0.796875, "sample": [-4.3311767578125, 117.13985443115234, 193.517822265625, 161.57794189453125, 19.844120025634766, -92.16708374023438, 260.7997131347656, 42.983924865722656, 13.263944625854492, 371.97900390625, 84.05422973632812, 71.9518051147461, 138.32858276367188, 21.136192321777344, 386.171142578125, 145.57852172851562, 107.34382629394531, 79.01011657714844, 199.63043212890625, 19.711044311523438, 2.8430938720703125, 273.8742980957031, -220.8607635498047, 51.95592498779297, 262.9757995605469, -23.244033813476562, 62.024444580078125, 335.99322509765625, 150.1031951904297, 52.32685852050781, -19.191970825195312, 323.6851806640625, 20.958755493164062, -5.215066909790039, 137.31649780273438, 75.17921447753906, -96.8790512084961, -63.0302734375, -158.25820922851562, 16.425338745117188, 113.17133331298828, -66.8797378540039, 111.6086196899414, 43.86329650878906, -29.927209854125977, -94.19561767578125, 112.25897216796875, 313.58111572265625, 48.432525634765625, 58.553367614746094, -200.2547149658203, 393.3525390625, 123.39630126953125, 22.50699806213379, 210.12254333496094, 104.02606964111328, 66.06040954589844, 396.9456787109375, 5.1716156005859375, 443.2001647949219, 226.22982788085938, 227.82215881347656, 38.88215255737305, 43.957611083984375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000190.npy"}
{"epoch": 0.2790014684287812, "step": 191, "batch_size": 64, "mean": 141.7076873779297, "std": 156.3033447265625, "min": -143.52676391601562, "p10": -10.856475448608393, "median": 118.5624885559082, "p90": 334.0201110839844, "max": 587.9791259765625, "pos_frac": 0.859375, "sample": [481.7724914550781, 112.36091613769531, 587.9791259765625, 34.10614013671875, 305.690185546875, 73.85237121582031, 153.8502960205078, 282.314453125, 16.003110885620117, 370.9564208984375, 106.21470642089844, 169.85775756835938, -13.090278625488281, 279.2528991699219, 298.0040283203125, 44.35401153564453, 36.4178466796875, 84.85321807861328, 37.48789978027344, 74.27980041503906, 270.2305908203125, 53.382781982421875, 45.425567626953125, 298.85455322265625, 126.51795196533203, 8.038986206054688, 8.40256118774414, 561.0509643554688, 189.4258575439453, -20.546424865722656, 232.59603881835938, -80.14797973632812, 17.654808044433594, -56.0739631652832, 142.7926025390625, 174.4859619140625, 173.13937377929688, 145.6880645751953, 336.75689697265625, 32.38915252685547, 8.036783218383789, 253.66510009765625, 230.2574462890625, -143.52676391601562, 78.51850891113281, 533.756591796875, 36.090126037597656, -71.23399353027344, 125.05484008789062, 107.65144348144531, 194.5252227783203, 327.63427734375, 29.780311584472656, -5.644268035888672, 127.39247131347656, 177.03829956054688, -72.76798248291016, 129.5385284423828, -1.0941619873046875, 124.7640609741211, 374.9761962890625, 31.73412322998047, 95.39369201660156, 181.16937255859375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000191.npy"}
{"epoch": 0.28046989720998533, "step": 192, "batch_size": 64, "mean": 92.88227081298828, "std": 150.2056884765625, "min": -265.2543640136719, "p10": -55.91381683349609, "median": 71.10405731201172, "p90": 319.67036743164067, "max": 541.76416015625, "pos_frac": 0.734375, "sample": [-52.447227478027344, -146.3865509033203, 43.96141815185547, 171.65234375, 139.10629272460938, -33.62061309814453, -57.399497985839844, -2.970243453979492, 31.376510620117188, -265.2543640136719, 24.942230224609375, 101.21162414550781, -37.64640808105469, 228.31581115722656, 17.269027709960938, 43.44862365722656, -64.68484497070312, 98.60826110839844, -38.659088134765625, 181.48716735839844, -124.4054183959961, 427.1971740722656, -51.514095306396484, 77.85599517822266, -21.709213256835938, 326.33428955078125, 364.24298095703125, 382.5165100097656, 156.8847198486328, 361.78302001953125, 95.49246215820312, 77.56028747558594, 541.76416015625, 115.58456420898438, 191.41433715820312, 29.999292373657227, 118.44731140136719, 219.6710662841797, 271.46044921875, 66.80610656738281, 43.02998352050781, 304.1212158203125, 9.689111709594727, 42.73236083984375, 135.2674560546875, 57.81731414794922, 59.281349182128906, -178.45643615722656, 212.83306884765625, 81.18115997314453, 88.34732818603516, -9.75634765625, -37.54674530029297, -75.17615509033203, 15.211284637451172, 235.456298828125, 73.96273803710938, -29.656112670898438, 35.36639404296875, 215.97219848632812, 136.77044677734375, 107.47627258300781, 68.24537658691406, 342.5993957519531], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000192.npy"}
{"epoch": 0.28193832599118945, "step": 193, "batch_size": 64, "mean": 54.66958236694336, "std": 145.4112548828125, "min": -437.07354736328125, "p10": -76.72494659423828, "median": 56.493690490722656, "p90": 242.06634979248054, "max": 373.429443359375, "pos_frac": 0.65625, "sample": [117.87458801269531, -17.693206787109375, 29.72576141357422, 182.65078735351562, -437.07354736328125, 94.47541809082031, 70.80619049072266, 59.70769500732422, -419.95947265625, 313.0666198730469, 64.52840423583984, -14.514516830444336, 188.8408203125, -34.56144332885742, -25.458709716796875, 105.23892211914062, 102.13726806640625, 129.39381408691406, -3.5314865112304688, 251.4525146484375, 2.316577911376953, -18.610313415527344, 12.867610931396484, 157.4930419921875, 29.942678451538086, 135.89903259277344, -58.712215423583984, -41.16699981689453, 4.4547119140625, -69.24501037597656, -17.155725479125977, 223.9528350830078, -79.93063354492188, -90.05501556396484, 249.82928466796875, -129.15802001953125, 188.12332153320312, -98.76924896240234, 2.5603275299072266, 53.279685974121094, 10.827354431152344, 146.04745483398438, 279.6391906738281, 175.23175048828125, 118.13066101074219, -137.81719970703125, 329.14154052734375, 90.35985565185547, 129.86819458007812, 128.48182678222656, -67.8150863647461, 106.15093994140625, -19.010051727294922, -64.82560729980469, -54.4490966796875, 141.61270141601562, 373.429443359375, 134.74440002441406, -59.84878158569336, 20.37603759765625, 292.5423278808594, 89.60260772705078, 16.509490966796875, 104.90106964111328], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000193.npy"}
{"epoch": 0.2834067547723935, "step": 194, "batch_size": 64, "mean": 87.92755889892578, "std": 102.08650207519531, "min": -103.00077819824219, "p10": -47.9197052001953, "median": 68.8138313293457, "p90": 247.90613708496093, "max": 304.64105224609375, "pos_frac": 0.8125, "sample": [142.31533813476562, 20.218177795410156, 33.89942932128906, 6.921051025390625, 120.07511901855469, 259.42169189453125, 116.97763061523438, 167.1623077392578, -28.620819091796875, 20.39532470703125, 30.851335525512695, 253.23464965820312, -72.36000061035156, -52.435333251953125, 44.348350524902344, 66.59405517578125, 209.54745483398438, 57.525909423828125, 66.73233032226562, 131.5584716796875, 31.19613265991211, 77.35070037841797, 15.713626861572266, 7.236598968505859, 120.36714935302734, -82.16036987304688, -2.440366744995117, 222.6060791015625, 66.88579559326172, 82.27718353271484, 248.56480407714844, 136.62669372558594, 249.0478057861328, 177.76315307617188, 28.64196014404297, 291.8165283203125, 64.89982604980469, 28.503767013549805, 144.35057067871094, 53.09258270263672, 195.21836853027344, -88.65286254882812, 43.394989013671875, -25.25391387939453, 106.5887451171875, 171.27313232421875, -9.478565216064453, -103.00077819824219, 40.545677185058594, -54.4730224609375, 119.83372497558594, -63.77941131591797, 65.17739868164062, 246.36924743652344, 77.58255004882812, 304.64105224609375, 137.5723419189453, 70.74186706542969, 295.66912841796875, 94.22862243652344, 179.68505859375, 183.7308349609375, 120.43003845214844, -37.38323974609375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000194.npy"}
{"epoch": 0.28487518355359764, "step": 195, "batch_size": 64, "mean": 95.9958267211914, "std": 117.92337036132812, "min": -106.27706909179688, "p10": -51.05067481994628, "median": 81.49372100830078, "p90": 246.97707672119142, "max": 400.4366760253906, "pos_frac": 0.796875, "sample": [86.84425354003906, -22.09088134765625, 224.47325134277344, 58.35207748413086, 69.5823974609375, 400.4366760253906, 71.11652374267578, 324.8739013671875, 198.73455810546875, -12.108814239501953, -7.476593017578125, 46.74934387207031, 4.998313903808594, -94.22564697265625, 248.973388671875, 117.08067321777344, 51.34303283691406, 242.3190155029297, -38.177574157714844, 332.2548828125, -44.96110153198242, 108.92994689941406, -106.27706909179688, 37.038299560546875, 107.91117858886719, 115.23499298095703, 229.16204833984375, 315.88201904296875, 151.29544067382812, 123.04730987548828, 162.13748168945312, -78.42247772216797, 80.73905181884766, 188.03378295898438, 25.555999755859375, 158.64834594726562, -29.775976181030273, 4.137107849121094, 174.510986328125, 210.94029235839844, -89.1739730834961, 252.6678466796875, 41.40401077270508, 82.2483901977539, 396.1680908203125, 143.64553833007812, 78.42757415771484, 63.47297668457031, -55.33446502685547, 15.782583236694336, 160.35791015625, -75.62261962890625, 224.49969482421875, 110.78929138183594, 26.598220825195312, 89.78421020507812, 32.75028991699219, 43.86463165283203, 20.82282066345215, 83.76140594482422, -53.660491943359375, 52.91947555541992, 146.78514099121094, 112.953857421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000195.npy"}
{"epoch": 0.28634361233480177, "step": 196, "batch_size": 64, "mean": 105.875, "std": 144.44163513183594, "min": -221.65155029296875, "p10": -52.899340820312496, "median": 80.00452041625977, "p90": 308.99691772460943, "max": 443.9705810546875, "pos_frac": 0.765625, "sample": [200.88034057617188, 202.0613250732422, 314.6725158691406, 217.10121154785156, -46.658973693847656, 296.541015625, 299.1512451171875, 442.7122802734375, 232.47552490234375, 71.99546813964844, 10.50589370727539, 49.67625427246094, 283.47119140625, 187.1368865966797, 51.304443359375, 14.116884231567383, -16.685123443603516, 87.09544372558594, -7.57623291015625, 38.49882507324219, 432.6905517578125, -29.44241714477539, 51.77500915527344, 31.472454071044922, 128.52737426757812, 133.4669189453125, 124.32745361328125, 144.0361328125, 52.236083984375, -9.815685272216797, 108.3118896484375, 76.20169830322266, 438.39111328125, 443.9705810546875, -23.233444213867188, 182.35951232910156, 87.14315795898438, 65.43046569824219, 111.95924377441406, 86.46954345703125, 66.9074478149414, 57.81243133544922, 222.2257080078125, -3.0880279541015625, 129.90103149414062, 30.588520050048828, 83.80734252929688, -160.43157958984375, 65.15830993652344, -221.65155029296875, -72.2374267578125, 150.68975830078125, 92.36181640625, -59.8755989074707, 313.21649169921875, -55.57378387451172, -141.4700927734375, 69.50121307373047, -0.53802490234375, 132.31378173828125, 265.2203063964844, 318.85400390625, -76.837890625, 4.391794204711914], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000196.npy"}
{"epoch": 0.2878120411160059, "step": 197, "batch_size": 64, "mean": 82.80601501464844, "std": 114.75785827636719, "min": -125.31022644042969, "p10": -35.432180786132804, "median": 54.67739486694336, "p90": 229.21602630615234, "max": 426.8374938964844, "pos_frac": 0.8125, "sample": [58.88384246826172, 331.385986328125, 143.16368103027344, -21.92247772216797, 163.2467498779297, 11.65423583984375, 12.5465087890625, -68.16213989257812, 67.32088470458984, 426.8374938964844, 361.9925231933594, 228.9877166748047, 66.90052795410156, 125.2047119140625, -99.42249298095703, 3.5490036010742188, 120.49359893798828, 285.3553466796875, 30.66913604736328, 47.76584243774414, 100.48123168945312, 220.0154266357422, 40.149356842041016, -0.98162841796875, 52.80292510986328, 11.351591110229492, -93.89366912841797, 130.05538940429688, 69.14533233642578, 88.587158203125, 259.65777587890625, 144.38726806640625, 33.96906280517578, 229.31387329101562, 1.307373046875, 190.15562438964844, -38.65960693359375, 1.2946929931640625, -64.96356201171875, -125.31022644042969, 35.58222961425781, 115.7876205444336, -1.1190567016601562, 11.748359680175781, 18.71504020690918, 77.75198364257812, 317.8758544921875, 204.19692993164062, 25.961715698242188, 196.7507781982422, -11.210298538208008, 51.3371467590332, 81.45243072509766, 26.84040069580078, 16.272911071777344, 119.10472869873047, 173.52281188964844, 82.0139389038086, 204.33688354492188, -27.901519775390625, -63.74517822265625, 21.882728576660156, 20.559112548828125, 56.55186462402344], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000197.npy"}
{"epoch": 0.28928046989721, "step": 198, "batch_size": 64, "mean": 105.88169860839844, "std": 143.87200927734375, "min": -242.02691650390625, "p10": -27.288115119934073, "median": 94.18695831298828, "p90": 236.9903701782227, "max": 610.2509765625, "pos_frac": 0.84375, "sample": [-129.96731567382812, 102.27848815917969, -38.953147888183594, 498.5658264160156, 124.67501831054688, 202.83416748046875, 339.8301086425781, 59.31727981567383, 157.1884307861328, 40.97600555419922, 109.86090087890625, 167.34893798828125, -141.60256958007812, 17.83679962158203, 3.031890869140625, 240.23777770996094, 152.37680053710938, 120.71188354492188, 93.77752685546875, 87.7207260131836, -19.134511947631836, -30.782516479492188, 34.42729187011719, 161.92884826660156, 13.54145622253418, 75.07386016845703, -73.4674301147461, 206.04315185546875, 94.59638977050781, 89.88880920410156, 86.38593292236328, 113.34654998779297, 127.80307006835938, 66.64247131347656, 120.20018005371094, 5.934452056884766, 82.647216796875, 160.18502807617188, 152.87435913085938, 229.4130859375, -242.02691650390625, 531.039306640625, 95.71135711669922, 63.904029846191406, 5.920101165771484, 30.380630493164062, 248.07354736328125, -46.99079895019531, 3.1656436920166016, 70.68067932128906, 43.2389030456543, 96.76937866210938, 99.79025268554688, 140.30975341796875, -9.930299758911133, 137.91519165039062, 163.07064819335938, 200.50262451171875, 18.592815399169922, 425.8878173828125, -6.650360107421875, 610.2509765625, 50.79169464111328, 140.4384765625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000198.npy"}
{"epoch": 0.2907488986784141, "step": 199, "batch_size": 64, "mean": 105.29148864746094, "std": 142.03341674804688, "min": -167.84136962890625, "p10": -51.084406661987295, "median": 86.93523788452148, "p90": 285.5454467773438, "max": 475.93585205078125, "pos_frac": 0.78125, "sample": [220.644287109375, 167.80665588378906, -140.35238647460938, -8.531661987304688, 274.7998962402344, 87.20794677734375, -55.28680419921875, 200.9378204345703, -167.50588989257812, 120.29239654541016, -0.6810073852539062, -79.56866455078125, -167.84136962890625, 205.5272979736328, 210.50125122070312, 62.32141876220703, 16.036788940429688, 138.31509399414062, 398.41802978515625, 41.50757598876953, 75.1248779296875, 172.54730224609375, 365.03497314453125, 308.94024658203125, -89.19522094726562, 28.721725463867188, -38.83142852783203, 125.69168853759766, 214.61434936523438, 22.005905151367188, 33.989227294921875, 258.55810546875, 475.93585205078125, 81.1881103515625, 75.9157485961914, 196.966796875, -26.764236450195312, 116.12022399902344, 30.905548095703125, 98.16915893554688, 65.07015991210938, 281.7146301269531, 126.33814239501953, 34.10670852661133, 287.1872253417969, 124.8770980834961, 71.5096435546875, -31.405832290649414, -41.278812408447266, -164.5836181640625, 273.1614074707031, 85.4996109008789, 105.34629821777344, 91.33380889892578, 354.6171875, 181.23927307128906, 3.3359317779541016, 86.66252899169922, 56.468902587890625, 88.40568542480469, 41.27946472167969, 160.53872680664062, 427.61968994140625, -20.576065063476562], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000199.npy"}
{"epoch": 0.2922173274596182, "step": 200, "batch_size": 64, "mean": 107.39921569824219, "std": 159.03164672851562, "min": -286.91937255859375, "p10": -49.0805881500244, "median": 87.18176651000977, "p90": 345.7196685791016, "max": 557.3422241210938, "pos_frac": 0.796875, "sample": [237.63027954101562, 107.93435668945312, 59.47822952270508, 416.7602844238281, 47.1564826965332, 153.71524047851562, 15.355766296386719, 2.7697296142578125, 421.481689453125, -67.15786743164062, 335.46099853515625, 248.16256713867188, 101.41930389404297, 9.94900894165039, 94.55685424804688, 332.5242004394531, 221.58511352539062, 389.5810546875, 3.0611419677734375, 350.1162414550781, 70.10427856445312, -2.8211612701416016, 160.57574462890625, 224.57908630371094, -22.062610626220703, 71.52565002441406, 181.635498046875, -199.97113037109375, 557.3422241210938, 94.9305648803711, 114.51116943359375, -32.15336227416992, -6.5897979736328125, 67.93814086914062, 53.143463134765625, -25.806211471557617, -286.91937255859375, 26.174713134765625, 99.71620178222656, 362.2391662597656, 0.7219295501708984, 89.53601837158203, 62.88934326171875, 1.0327987670898438, 3.2922019958496094, -145.15744018554688, 84.8275146484375, -56.335113525390625, 296.8683166503906, -27.24965476989746, 13.843063354492188, 419.09381103515625, 48.86476135253906, 23.01532745361328, 103.05928039550781, 262.25506591796875, 174.18910217285156, 161.76461791992188, 110.21653747558594, 178.5958251953125, -123.43568420410156, 147.7406768798828, -73.839599609375, 128.12777709960938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000200.npy"}
{"epoch": 0.2936857562408223, "step": 201, "batch_size": 64, "mean": 104.22012329101562, "std": 146.99517822265625, "min": -288.2005920410156, "p10": -52.851589965820295, "median": 82.37252044677734, "p90": 283.7857543945313, "max": 490.3997802734375, "pos_frac": 0.8125, "sample": [260.8199462890625, 14.587972640991211, 226.65968322753906, -214.810791015625, 48.424232482910156, -98.77607727050781, 10.46694564819336, 14.972402572631836, -30.914749145507812, 88.76370239257812, 2.4810791015625, 25.092498779296875, 37.103668212890625, -18.822021484375, 476.8898010253906, 105.7134017944336, 54.764068603515625, 93.3353271484375, 28.740074157714844, 243.6180419921875, 353.6651306152344, 135.779052734375, 11.281707763671875, 93.02961730957031, 208.74020385742188, 288.6437072753906, -76.76811218261719, 223.7781524658203, 205.33499145507812, -10.217565536499023, 74.68348693847656, 129.06674194335938, 137.9366455078125, 127.69566345214844, -35.55662536621094, -72.23509216308594, 179.61367797851562, 78.59918975830078, 53.062713623046875, 490.3997802734375, 111.014404296875, -26.66014862060547, 173.301025390625, 77.03472900390625, 59.05931854248047, 171.63719177246094, 158.46121215820312, 252.9526824951172, 399.1418151855469, 67.83110046386719, 86.1458511352539, -67.7891845703125, 159.5792999267578, 74.58284759521484, -288.2005920410156, 301.42059326171875, 427.3314514160156, 148.32290649414062, 12.237176895141602, -60.26371765136719, 272.4505310058594, 47.57238006591797, 13.25469970703125, 134.02847290039062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000201.npy"}
{"epoch": 0.29515418502202645, "step": 202, "batch_size": 64, "mean": 73.8116683959961, "std": 144.14883422851562, "min": -217.91433715820312, "p10": -99.30036392211912, "median": 62.44511413574219, "p90": 264.43345947265624, "max": 529.2406005859375, "pos_frac": 0.703125, "sample": [529.2406005859375, -49.05918502807617, 115.8126449584961, 329.7115478515625, 4.332761764526367, -57.72112274169922, 174.1903076171875, -53.588584899902344, 124.96841430664062, 98.00052642822266, 57.81890106201172, -198.8842315673828, 177.96624755859375, 247.91461181640625, 227.5140380859375, 49.9769287109375, -203.75579833984375, 47.94732666015625, 109.57891845703125, 348.030029296875, 119.03910064697266, 290.4124755859375, -135.51739501953125, 197.42529296875, 26.45685577392578, 29.110923767089844, 384.15423583984375, -18.0842342376709, 112.57818603515625, 138.607421875, 101.10835266113281, 125.40562438964844, 263.75653076171875, -28.863906860351562, -36.900390625, -128.4939727783203, 79.57084655761719, 62.6484375, 13.194969177246094, 229.33441162109375, 105.63758850097656, -5.8011627197265625, 20.439197540283203, 265.4770812988281, 171.07522583007812, 119.8258056640625, 85.51953887939453, -217.91433715820312, 62.241790771484375, 93.49658203125, 19.416057586669922, 44.77046203613281, -32.64430236816406, -70.70677947998047, 45.743255615234375, -110.56755065917969, 82.47914123535156, -73.01026153564453, -44.935001373291016, 55.246551513671875, 264.72357177734375, -133.0411376953125, 94.98037719726562, -23.442943572998047], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000202.npy"}
{"epoch": 0.2966226138032305, "step": 203, "batch_size": 64, "mean": 91.34191131591797, "std": 123.81547546386719, "min": -173.36431884765625, "p10": -64.34622344970703, "median": 78.02753448486328, "p90": 265.84788818359385, "max": 449.72369384765625, "pos_frac": 0.78125, "sample": [24.012409210205078, 139.58599853515625, 167.12118530273438, 66.01288604736328, -28.92141342163086, 285.7425231933594, -76.857177734375, 144.42803955078125, 28.128009796142578, 121.10977172851562, -2.4583053588867188, 109.57831573486328, 108.20572662353516, 164.9022216796875, -107.79203796386719, 180.1459503173828, 135.32473754882812, -32.763397216796875, 196.9595947265625, 137.52098083496094, 274.7275695800781, 422.62115478515625, -72.43836975097656, 119.81764221191406, 141.06210327148438, 37.035797119140625, 449.72369384765625, 12.500923156738281, -105.90853881835938, 143.9153289794922, 2.4275741577148438, 169.48193359375, 175.61529541015625, 277.7737731933594, 41.423187255859375, 182.90283203125, 93.30731201171875, 114.9333267211914, 76.94784545898438, 133.72048950195312, 36.80352783203125, 65.20673370361328, 45.04595184326172, 245.12863159179688, -40.051368713378906, -173.36431884765625, -66.82704162597656, 67.38528442382812, -23.920358657836914, 7.867870330810547, 178.00125122070312, 18.134788513183594, 27.265697479248047, 48.27571105957031, -58.557647705078125, -79.73770141601562, 148.481689453125, 79.10722351074219, 277.3785095214844, 326.7271423339844, -4.006433486938477, 228.66725158691406, 17.842065811157227, 23.451210021972656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000203.npy"}
{"epoch": 0.29809104258443464, "step": 204, "batch_size": 64, "mean": 82.447265625, "std": 143.89468383789062, "min": -216.28277587890625, "p10": -72.62644119262693, "median": 62.12775421142578, "p90": 284.1597061157228, "max": 489.710205078125, "pos_frac": 0.703125, "sample": [209.32664489746094, 73.9068603515625, 40.137264251708984, 42.38414001464844, 120.63436126708984, -3.8265304565429688, 210.70074462890625, -104.88545989990234, 95.53156280517578, 45.41938018798828, -39.04843521118164, 20.77802848815918, 13.755279541015625, 90.7479248046875, 57.45167541503906, -34.50005340576172, -216.28277587890625, 196.93753051757812, 52.285186767578125, 15.976175308227539, 241.91868591308594, -133.71224975585938, -23.497604370117188, -11.323894500732422, 66.8038330078125, -13.74966049194336, 76.27667236328125, 153.4942169189453, 155.201171875, -83.90486145019531, 393.520263671875, -14.740211486816406, 53.9508056640625, 36.498558044433594, -42.573516845703125, 155.22097778320312, 52.36286926269531, 142.7010498046875, 450.35284423828125, -209.6175994873047, 302.26300048828125, 174.57264709472656, 167.20262145996094, 68.39897155761719, 73.19876098632812, -10.474641799926758, 43.64683532714844, 222.4306182861328, 335.0057373046875, -3.93804931640625, 79.98243713378906, 165.71786499023438, -125.44735717773438, 68.37400817871094, 489.710205078125, 349.53350830078125, -35.51924133300781, 154.84072875976562, 26.202089309692383, 358.0852966308594, -104.39947509765625, 81.94003295898438, 108.99664306640625, -46.31012725830078], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000204.npy"}
{"epoch": 0.29955947136563876, "step": 205, "batch_size": 64, "mean": 107.62008666992188, "std": 137.759765625, "min": -228.45083618164062, "p10": -60.29521751403807, "median": 108.7945671081543, "p90": 262.7959350585938, "max": 401.71270751953125, "pos_frac": 0.78125, "sample": [33.995697021484375, 115.18377685546875, 212.02157592773438, 363.231689453125, 109.48619842529297, 26.7576847076416, -9.828531265258789, 28.021453857421875, 12.677505493164062, 222.929931640625, 258.93585205078125, -114.59423828125, -85.98002624511719, 97.27281188964844, -16.094886779785156, 79.16560363769531, 16.94013786315918, 372.2302551269531, 249.3053741455078, 30.326854705810547, 116.96204376220703, 95.24166870117188, -109.53164672851562, 177.0460968017578, 301.77008056640625, 249.54702758789062, 205.02120971679688, -13.61667251586914, -41.48828887939453, 181.76101684570312, -46.84669876098633, 401.71270751953125, 110.82255554199219, 71.66205596923828, 33.889854431152344, 121.09896850585938, 383.1617126464844, 238.20672607421875, 108.10293579101562, 130.07431030273438, -220.5460205078125, 149.61598205566406, 122.68151092529297, 228.3452606201172, 146.07980346679688, -66.05886840820312, 149.01918029785156, 87.0431900024414, 264.45025634765625, 90.2582015991211, 243.54762268066406, 192.84571838378906, 175.84133911132812, 73.30098724365234, 155.4464874267578, 357.6230773925781, 105.72223663330078, -9.770835876464844, 6.3022003173828125, -228.45083618164062, -68.51820373535156, 207.250244140625, -30.0777587890625, 39.15257263183594], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000205.npy"}
{"epoch": 0.3010279001468429, "step": 206, "batch_size": 64, "mean": 116.11079406738281, "std": 123.67452239990234, "min": -213.15240478515625, "p10": -39.04083538055418, "median": 106.5759162902832, "p90": 267.2243194580078, "max": 404.7598876953125, "pos_frac": 0.84375, "sample": [66.74362182617188, 169.54312133789062, 88.43836975097656, 123.42276000976562, 404.7598876953125, 303.5903625488281, 76.51654052734375, 59.26268768310547, 202.85728454589844, 264.524658203125, 55.1942138671875, 131.27220153808594, 30.942426681518555, 93.3346939086914, -213.15240478515625, 51.95999526977539, 82.41970825195312, 153.0413055419922, 54.899383544921875, -17.738431930541992, 224.46664428710938, -54.42743682861328, 238.28857421875, 346.2708435058594, 99.40028381347656, 181.22958374023438, 21.959869384765625, 209.01939392089844, -68.04058837890625, -109.14481353759766, 170.75917053222656, 300.18218994140625, 220.83499145507812, 241.4351348876953, 15.274538040161133, 253.80020141601562, 55.28736114501953, -48.17043685913086, 161.42822265625, -137.31544494628906, 43.881004333496094, 191.64645385742188, 7.010990142822266, 149.97247314453125, 84.75247192382812, -13.158599853515625, 133.29293823242188, 84.52388000488281, 193.86073303222656, 94.20004272460938, -57.48828887939453, 14.444145202636719, 113.75154876708984, 164.4691162109375, 356.4954833984375, 261.9224853515625, -4.038087844848633, 120.7744140625, 118.45143127441406, 268.3813171386719, 62.556541442871094, 19.272098541259766, 306.51434326171875, 211.2314453125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000206.npy"}
{"epoch": 0.302496328928047, "step": 207, "batch_size": 64, "mean": 133.4520263671875, "std": 150.95108032226562, "min": -138.156982421875, "p10": -25.36036911010742, "median": 98.51725769042969, "p90": 306.697021484375, "max": 565.6060180664062, "pos_frac": 0.84375, "sample": [15.951000213623047, -138.156982421875, 67.85668182373047, 105.55233764648438, 173.26861572265625, 42.77228546142578, 306.652587890625, -25.288543701171875, -12.9912109375, 10.31158447265625, 102.86676025390625, 83.18485260009766, 306.716064453125, -10.423904418945312, 137.43780517578125, 147.3829345703125, 47.74571228027344, 109.69377136230469, 28.82381820678711, 37.72771453857422, 57.2786750793457, 183.66586303710938, 30.86278533935547, 246.4784698486328, 36.57673645019531, -31.863311767578125, 296.2568359375, 65.82778930664062, 28.49329948425293, -48.796504974365234, 285.17816162109375, 267.6041259765625, -47.41693115234375, 0.22201156616210938, 35.141258239746094, 258.70245361328125, 37.89707946777344, 225.2174835205078, 146.2457275390625, 233.8949432373047, 247.10089111328125, 127.85084533691406, 102.29011535644531, 137.73416137695312, 244.83306884765625, 252.0009765625, 89.37484741210938, 377.3526611328125, 514.5421142578125, 255.86021423339844, 47.371734619140625, 437.0751953125, -53.74610900878906, 32.96088409423828, 565.6060180664062, 41.804595947265625, 94.74440002441406, 102.31240844726562, 88.67347717285156, 496.473876953125, -80.0816421508789, 179.46133422851562, 420.1740417480469, -25.391151428222656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000207.npy"}
{"epoch": 0.3039647577092511, "step": 208, "batch_size": 64, "mean": 108.03218841552734, "std": 152.25390625, "min": -325.37030029296875, "p10": -31.479232025146477, "median": 86.1397476196289, "p90": 323.6066009521486, "max": 426.44781494140625, "pos_frac": 0.8125, "sample": [-192.02442932128906, 144.87396240234375, -149.73672485351562, 259.048828125, 22.04486083984375, -18.34830093383789, 378.30120849609375, 31.137632369995117, 124.17931365966797, 264.431640625, 238.2357940673828, 57.999671936035156, 76.1268310546875, 170.18589782714844, 213.59088134765625, 62.292762756347656, 422.720703125, 30.11907958984375, 426.44781494140625, 142.88720703125, 3.094024658203125, 35.996559143066406, 239.69818115234375, 179.64846801757812, 96.15266418457031, 26.438209533691406, 55.876800537109375, 15.944793701171875, -116.39302062988281, 258.8421325683594, -143.2704620361328, 7.046840667724609, 53.929691314697266, 111.01563262939453, 131.63006591796875, 291.372314453125, 268.74774169921875, 41.1515007019043, -0.12401580810546875, 230.2047119140625, -34.359039306640625, 67.06990051269531, 24.077117919921875, 344.590087890625, 144.42156982421875, 357.4707336425781, 98.53102111816406, -325.37030029296875, -24.759681701660156, 150.50518798828125, 15.18182373046875, 125.39732360839844, 337.4212951660156, 100.40960693359375, 55.94285583496094, -130.2728271484375, 22.50994873046875, 375.1542663574219, 169.48550415039062, 287.943359375, 212.62396240234375, 74.9876708984375, -1.76690673828125, -24.651702880859375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000208.npy"}
{"epoch": 0.3054331864904552, "step": 209, "batch_size": 64, "mean": 95.91326904296875, "std": 149.48805236816406, "min": -320.36505126953125, "p10": -80.02110595703125, "median": 87.04819869995117, "p90": 313.67150268554695, "max": 417.29083251953125, "pos_frac": 0.71875, "sample": [336.81488037109375, 85.74497985839844, 172.48326110839844, -40.99144744873047, 375.9071044921875, 411.7674255371094, 320.377685546875, 111.39100646972656, 131.12002563476562, 114.48355102539062, -62.78074264526367, 93.29048919677734, 178.21981811523438, 87.2894515991211, 165.44375610351562, 142.86630249023438, 162.77947998046875, -144.10865783691406, -81.00637817382812, 54.80382537841797, 256.6699523925781, 233.40191650390625, -3.958341598510742, -90.28567504882812, 55.28131103515625, -41.00663757324219, -12.007303237915039, 4.830284118652344, 102.73226928710938, 150.48916625976562, 118.41650390625, 89.01011657714844, -130.4744873046875, -35.975040435791016, 61.73291778564453, 56.75152587890625, -77.72213745117188, -12.53695297241211, 381.2174072265625, 283.396240234375, 50.02447509765625, 417.29083251953125, 11.396102905273438, -38.83959197998047, -30.156600952148438, 49.36292266845703, 71.84294128417969, 78.38424682617188, 244.32069396972656, 142.58323669433594, 140.65032958984375, -36.00421142578125, -109.68695068359375, -320.36505126953125, 170.3052215576172, 86.80694580078125, 44.0786018371582, 50.94506072998047, 139.82476806640625, -144.09825134277344, 337.048828125, 298.02374267578125, 217.54092407226562, 261.31121826171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000209.npy"}
{"epoch": 0.3069016152716593, "step": 210, "batch_size": 64, "mean": 110.97904968261719, "std": 146.30441284179688, "min": -224.18240356445312, "p10": -25.952098274230952, "median": 84.09422302246094, "p90": 317.1271057128907, "max": 461.60308837890625, "pos_frac": 0.75, "sample": [-1.4671630859375, 22.520166397094727, 135.5616912841797, 82.7064208984375, -10.072982788085938, -69.29579162597656, 185.54348754882812, -21.587697982788086, 133.81967163085938, -32.678627014160156, 194.13650512695312, 23.78441619873047, 26.099506378173828, 112.70439910888672, 284.3188171386719, 147.53363037109375, -10.076889038085938, 73.77635192871094, 215.185546875, 29.650501251220703, 303.20550537109375, 92.13037109375, 21.75914192199707, 243.18109130859375, 284.1217346191406, 2.5799312591552734, 14.720909118652344, 323.093505859375, 194.04690551757812, 358.3170471191406, 133.0867462158203, 241.02024841308594, 110.3691177368164, 300.82659912109375, -139.23228454589844, -27.822555541992188, -71.2064208984375, -15.474227905273438, 37.726806640625, 20.707622528076172, 146.65118408203125, 170.45150756835938, 353.70037841796875, 112.16555786132812, -3.0789928436279297, -12.840545654296875, 436.4217224121094, -4.0926513671875, 124.72273254394531, 157.513671875, 295.34002685546875, 461.60308837890625, 291.823974609375, 57.34752655029297, -60.68956756591797, 85.48202514648438, -16.83440399169922, 0.9963340759277344, 12.608922958374023, 354.0723876953125, 6.241218566894531, 15.325950622558594, -224.18240356445312, 392.59002685546875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000210.npy"}
{"epoch": 0.30837004405286345, "step": 211, "batch_size": 64, "mean": 92.56353759765625, "std": 162.31161499023438, "min": -206.50860595703125, "p10": -58.1384666442871, "median": 60.089717864990234, "p90": 263.6896911621094, "max": 562.41357421875, "pos_frac": 0.734375, "sample": [126.98011016845703, -51.6055908203125, 252.96124267578125, 70.70723724365234, 81.45172119140625, 161.24131774902344, -51.09718322753906, 115.59806060791016, 513.2855834960938, -206.50860595703125, -29.13268280029297, 44.329437255859375, 28.674335479736328, 1.588775634765625, 74.33135986328125, 230.4438934326172, 148.2842559814453, 193.570556640625, 167.8876953125, -17.028915405273438, -4.585113525390625, -37.39666748046875, 196.82888793945312, 105.98052978515625, -60.938270568847656, 1.5302581787109375, -48.519309997558594, 486.7792053222656, 43.68813705444336, 562.41357421875, 481.0057373046875, -42.74809265136719, -117.21139526367188, 133.0273895263672, 252.02099609375, 320.774169921875, -139.85821533203125, 139.7810516357422, 49.00898742675781, 0.2156982421875, 124.43272399902344, 418.18402099609375, -33.985939025878906, 69.13824462890625, 12.254453659057617, 54.34296417236328, 57.319923400878906, -88.98020935058594, 15.33804702758789, 203.23483276367188, -163.5493621826172, 62.85951232910156, 210.80348205566406, 46.64532470703125, 160.7049102783203, 268.28759765625, 103.77190399169922, 246.9251708984375, 8.156326293945312, 130.5706787109375, 34.83986282348633, 14.994354248046875, -12.885330200195312, -197.09759521484375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000211.npy"}
{"epoch": 0.30983847283406757, "step": 212, "batch_size": 64, "mean": 132.1768035888672, "std": 169.31903076171875, "min": -253.42335510253906, "p10": -56.51351470947264, "median": 135.42850494384766, "p90": 321.6357147216797, "max": 594.1607666015625, "pos_frac": 0.75, "sample": [-253.42335510253906, -86.71147918701172, 315.3689880371094, 261.20880126953125, 45.31108856201172, 253.5135498046875, 115.84121704101562, -64.15739440917969, -38.267845153808594, -3.5389671325683594, 235.46250915527344, 175.19097900390625, -212.3009796142578, -38.67779541015625, -22.382274627685547, 12.410728454589844, 288.45928955078125, 214.42568969726562, 428.972412109375, 240.5465850830078, 112.92071533203125, -111.10326385498047, 29.77841567993164, 115.31488037109375, 136.12921142578125, 169.71771240234375, 236.21456909179688, -74.84144592285156, 321.8059387207031, -8.566110610961914, 115.81766510009766, 182.45860290527344, 594.1607666015625, 184.13795471191406, 54.924354553222656, -213.32957458496094, 321.238525390625, -26.344833374023438, 129.36679077148438, 553.9169921875, 287.6519775390625, 135.47994995117188, 47.4373779296875, 226.86181640625, 15.387130737304688, 145.59039306640625, -33.204402923583984, 53.732173919677734, 322.0270690917969, 52.941654205322266, 150.81765747070312, 369.46600341796875, 392.37091064453125, 249.39035034179688, 282.671630859375, -11.862957000732422, 224.22410583496094, 68.69058227539062, 158.32778930664062, 135.37705993652344, 61.16881561279297, 295.5084228515625, -34.014190673828125, 172.30401611328125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000212.npy"}
{"epoch": 0.31130690161527164, "step": 213, "batch_size": 64, "mean": 129.48403930664062, "std": 154.94004821777344, "min": -216.19146728515625, "p10": -46.436990356445314, "median": 120.32686996459961, "p90": 344.66795349121094, "max": 444.578125, "pos_frac": 0.8125, "sample": [302.00286865234375, -84.9847412109375, 31.84313201904297, -95.00273895263672, 217.1445770263672, 17.236068725585938, 58.57396697998047, 115.54319763183594, -216.19146728515625, -77.43911743164062, 10.633407592773438, 320.384033203125, 349.9293212890625, -42.824623107910156, 85.81978607177734, 178.00030517578125, 9.79496955871582, 63.762939453125, 165.63262939453125, 324.51171875, 444.578125, 317.27557373046875, 48.52881622314453, 425.95318603515625, 181.62118530273438, 288.9918212890625, 394.70452880859375, -16.085094451904297, -6.340003967285156, 2.421863555908203, 182.40432739257812, 125.11054229736328, 197.44329833984375, 132.1286163330078, -46.34942626953125, 29.258636474609375, 175.58636474609375, 354.1072082519531, 289.8276672363281, 151.84036254882812, -4.100582122802734, 189.4142608642578, 343.8580017089844, -46.474517822265625, 218.7216339111328, 68.84335327148438, 191.64825439453125, 425.5264587402344, -134.16937255859375, 212.14810180664062, 254.52749633789062, 308.8333740234375, 345.01507568359375, 42.71609878540039, 11.098831176757812, 6.460582733154297, -132.748291015625, 74.92562866210938, 94.70220947265625, 66.68070220947266, 26.171852111816406, 14.850738525390625, 172.2738494873047, 128.67713928222656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000213.npy"}
{"epoch": 0.31277533039647576, "step": 214, "batch_size": 64, "mean": 114.10516357421875, "std": 149.00509643554688, "min": -171.79708862304688, "p10": -53.50607528686523, "median": 91.0162353515625, "p90": 298.9115661621094, "max": 526.84130859375, "pos_frac": 0.75, "sample": [70.98516082763672, 408.5975036621094, 12.127132415771484, 184.4256591796875, -11.911436080932617, -9.421445846557617, -76.923583984375, -171.79708862304688, 286.8277282714844, -82.36204528808594, 159.89190673828125, 41.695701599121094, 101.96388244628906, -0.6936874389648438, 32.374359130859375, 100.93220520019531, 229.69473266601562, -22.349489212036133, 37.57801818847656, 198.29031372070312, 65.71389770507812, 66.8106918334961, 103.29830169677734, 268.7222900390625, 255.7725830078125, 81.10026550292969, 21.564796447753906, -2.0542469024658203, 124.72444152832031, -50.77429962158203, 119.55538940429688, 62.47815704345703, -0.7068252563476562, 341.8454284667969, -84.7896499633789, 526.84130859375, 339.82598876953125, 19.506271362304688, 29.553176879882812, 390.9476013183594, 220.3612823486328, 208.5926055908203, 258.88543701171875, 302.1456298828125, -143.9709930419922, 525.1810302734375, 27.656461715698242, 193.83316040039062, 221.93356323242188, 103.85213470458984, -53.878623962402344, 184.8467559814453, -39.954063415527344, 121.03158569335938, -54.47261047363281, 234.55307006835938, 69.68373107910156, 291.36541748046875, 61.115692138671875, 144.45489501953125, 129.32994079589844, 123.66380310058594, -52.63679504394531, 55.29628372192383], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000214.npy"}
{"epoch": 0.3142437591776799, "step": 215, "batch_size": 64, "mean": 119.69615173339844, "std": 135.5141143798828, "min": -147.45973205566406, "p10": -66.05050659179688, "median": 100.44810485839844, "p90": 305.65821228027346, "max": 487.7257080078125, "pos_frac": 0.84375, "sample": [68.73554992675781, 318.3721923828125, 121.97708129882812, 263.3544006347656, -64.51242065429688, -66.70968627929688, 32.36888885498047, 126.09075164794922, 135.38035583496094, 102.11563110351562, 141.6916961669922, 52.329132080078125, 2.233766555786133, -27.036184310913086, 425.56170654296875, 251.84634399414062, 248.11676025390625, 307.211669921875, 487.7257080078125, 178.16232299804688, 98.53282165527344, 391.49212646484375, 109.30792999267578, -88.20475769042969, 78.46110534667969, 194.13735961914062, -67.427734375, 75.67229461669922, 190.11962890625, 249.6671142578125, 23.458120346069336, 81.19436645507812, 59.330108642578125, -3.477010726928711, 175.5249786376953, 98.78057861328125, -114.45682525634766, 155.0736846923828, 198.59768676757812, 246.43515014648438, 339.931884765625, 96.61176300048828, 188.45431518554688, 15.1175537109375, 142.0064239501953, 51.13786315917969, 50.63335418701172, -69.37751770019531, 45.8430290222168, 225.22982788085938, -147.45973205566406, 144.18853759765625, 397.06787109375, 34.506996154785156, 81.67884826660156, 77.34237670898438, 302.0334777832031, 10.620529174804688, 216.0838623046875, 41.75299835205078, 104.29879760742188, 42.26863098144531, -93.12101745605469, 106.46883392333984], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000215.npy"}
{"epoch": 0.315712187958884, "step": 216, "batch_size": 64, "mean": 119.4287338256836, "std": 168.41932678222656, "min": -278.05816650390625, "p10": -66.35827560424804, "median": 94.43572235107422, "p90": 362.34328002929686, "max": 549.188720703125, "pos_frac": 0.765625, "sample": [205.880615234375, 410.89532470703125, -76.40239715576172, 135.91217041015625, 214.71981811523438, 38.97767639160156, -175.079345703125, 158.10650634765625, 222.3225555419922, 18.12561798095703, 549.188720703125, 94.66474914550781, 112.29226684570312, -74.84002685546875, -29.402606964111328, -23.316741943359375, -96.64937591552734, 221.475341796875, 34.01337814331055, -183.91513061523438, 13.857131958007812, 318.84613037109375, -8.475090026855469, -4.81927490234375, 111.22792053222656, 153.89869689941406, 52.455810546875, 187.4195556640625, 127.9505615234375, 3.067045211791992, 94.20669555664062, 192.0816650390625, 13.857070922851562, -52.78819274902344, 399.1295471191406, 362.0301513671875, -53.48149108886719, 314.1748962402344, 46.73150634765625, 69.35140228271484, 58.66857147216797, 258.2369689941406, 8.669061660766602, -278.05816650390625, 62.39186096191406, 190.18113708496094, 244.89889526367188, 393.7291259765625, 48.657196044921875, -58.001014709472656, 279.9879455566406, 48.27299499511719, 117.4637222290039, 165.34576416015625, -59.582435607910156, 335.2540283203125, 253.0347900390625, 343.3760986328125, 403.68072509765625, 384.066650390625, 20.018966674804688, -69.26220703125, 32.239990234375, 362.47747802734375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000216.npy"}
{"epoch": 0.31718061674008813, "step": 217, "batch_size": 64, "mean": 141.24981689453125, "std": 156.08253479003906, "min": -154.84226989746094, "p10": -13.9316291809082, "median": 102.65357208251953, "p90": 369.0518249511719, "max": 548.2091064453125, "pos_frac": 0.84375, "sample": [139.27301025390625, 141.0719451904297, 106.32616424560547, 134.66510009765625, 62.599334716796875, 57.94035339355469, 77.76398468017578, 231.8611297607422, 78.94918060302734, 94.49237060546875, 51.290863037109375, 359.1160888671875, 148.99703979492188, 373.30999755859375, -115.18473052978516, 5.446189880371094, 36.02650451660156, -2.9051361083984375, 199.475830078125, 548.2091064453125, -0.22110366821289062, 53.42015838623047, 285.1378173828125, 434.1371154785156, 98.9809799194336, 223.02288818359375, -25.61847686767578, 67.2508544921875, -154.84226989746094, 124.18710327148438, 497.3621520996094, 196.96261596679688, 6.50324821472168, 69.65967559814453, 261.72381591796875, 10.12933349609375, 251.12997436523438, 297.4326171875, -10.352081298828125, 506.77899169921875, 420.3787841796875, -90.44298553466797, 333.71435546875, 285.83831787109375, 147.2805633544922, 461.6321105957031, 20.533248901367188, 107.07644653320312, 70.25733184814453, 60.76502227783203, 84.24589538574219, 158.63851928710938, 264.8687744140625, 13.171953201293945, 152.32308959960938, 75.73468780517578, -32.75048065185547, -15.465721130371094, 28.815338134765625, -47.56535339355469, 137.8763885498047, 197.05796813964844, 253.1273956298828, 31.366540908813477], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000217.npy"}
{"epoch": 0.3186490455212922, "step": 218, "batch_size": 64, "mean": 107.09344482421875, "std": 162.35406494140625, "min": -257.1349182128906, "p10": -79.3652442932129, "median": 90.87701416015625, "p90": 337.6669128417969, "max": 558.41357421875, "pos_frac": 0.75, "sample": [100.85067749023438, 61.708717346191406, 168.29681396484375, 256.58599853515625, 149.8675537109375, 25.499866485595703, 339.4462585449219, 83.33818054199219, 3.0438499450683594, 64.04364776611328, 161.8476104736328, 41.600982666015625, -15.740097045898438, 134.4832763671875, 457.4583740234375, 179.5457305908203, 127.42674255371094, -19.479080200195312, 152.35928344726562, 189.764404296875, 44.19587707519531, 54.758522033691406, 0.13633346557617188, 243.57373046875, -79.63090515136719, -257.1349182128906, 333.5151062011719, 95.20799255371094, 194.15701293945312, 198.21627807617188, 39.844879150390625, 381.99908447265625, 428.03338623046875, 249.21405029296875, 203.3853759765625, 58.73304748535156, -15.626752853393555, -25.38377571105957, 88.13017272949219, -31.30728530883789, -31.001998901367188, -100.00716400146484, 151.02291870117188, -102.68087768554688, -78.74536895751953, -56.732261657714844, 558.41357421875, 342.39520263671875, 205.2232208251953, -71.22179412841797, -151.05563354492188, 239.1682891845703, 437.741455078125, 13.016122817993164, 41.45347213745117, 93.62385559082031, 152.8744354248047, 93.91937255859375, 36.102027893066406, -178.16738891601562, 76.65946197509766, -121.80979919433594, 329.531494140625, 108.29171752929688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000218.npy"}
{"epoch": 0.3201174743024963, "step": 219, "batch_size": 64, "mean": 93.5392837524414, "std": 173.95372009277344, "min": -429.3218688964844, "p10": -106.14521255493163, "median": 71.47319412231445, "p90": 351.0690521240235, "max": 487.51885986328125, "pos_frac": 0.765625, "sample": [467.0440673828125, 130.0309600830078, 11.404394149780273, 341.48504638671875, 9.50114631652832, 430.82891845703125, -429.3218688964844, 11.821828842163086, 36.978424072265625, 292.6280212402344, -10.846900939941406, 99.71976470947266, 198.79168701171875, 433.12030029296875, 487.51885986328125, 179.25770568847656, 54.365882873535156, 246.56414794921875, 94.351806640625, 123.549560546875, 271.47613525390625, 32.80125427246094, -130.7052764892578, 9.8192138671875, -43.185340881347656, -58.866600036621094, 361.9392395019531, -4.9446868896484375, 355.1764831542969, 392.4491882324219, 29.00362205505371, 276.6568603515625, -2.8636627197265625, 53.38752746582031, 21.446731567382812, 184.68798828125, 117.9475326538086, 20.52600860595703, 116.68684387207031, 21.34478759765625, 294.8265686035156, 135.07403564453125, 101.310546875, -179.80755615234375, 89.61463928222656, 163.6643524169922, -181.82449340820312, 7.89338493347168, 8.442249298095703, 253.72828674316406, 34.514312744140625, -123.49857330322266, 9.327934265136719, -48.1339225769043, -204.97755432128906, 73.03372955322266, 181.71510314941406, 79.542724609375, -102.01235961914062, -29.825180053710938, 141.44760131835938, 69.91265869140625, 86.88487243652344, -107.91643524169922], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000219.npy"}
{"epoch": 0.32158590308370044, "step": 220, "batch_size": 64, "mean": 78.65856170654297, "std": 123.54718017578125, "min": -224.60791015625, "p10": -73.38679275512695, "median": 76.30416107177734, "p90": 205.05901336669922, "max": 452.9305725097656, "pos_frac": 0.75, "sample": [154.39398193359375, 44.818511962890625, 75.33274841308594, -27.548728942871094, 43.936317443847656, 310.58251953125, 342.08416748046875, 8.768741607666016, 154.91957092285156, 145.7686309814453, 108.50443267822266, 259.340087890625, 137.3453369140625, 136.1053466796875, -190.06663513183594, 117.10968017578125, 167.17214965820312, 35.40997314453125, 136.85708618164062, 204.47512817382812, 60.93833923339844, 21.700510025024414, 3.0657196044921875, -224.60791015625, -77.29827117919922, 207.6652374267578, -13.793069839477539, 52.14892578125, 205.3092498779297, 63.15675735473633, -54.350372314453125, 157.7552490234375, 122.60899353027344, -13.387130737304688, 221.00804138183594, 132.6470184326172, 452.9305725097656, 174.79356384277344, -6.873481750488281, 115.46819305419922, 14.176214218139648, 49.83749008178711, 3.551004409790039, 115.11787414550781, -102.06439208984375, 139.09286499023438, 71.825439453125, 136.62596130371094, -125.29995727539062, 77.27557373046875, 66.50288391113281, -1.6840438842773438, 151.102294921875, -64.260009765625, 35.166934967041016, 158.9075469970703, 177.25929260253906, -59.621482849121094, 181.14654541015625, 182.691162109375, -15.88680648803711, 152.96929931640625, -124.77639770507812, -151.70242309570312], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000220.npy"}
{"epoch": 0.32305433186490456, "step": 221, "batch_size": 64, "mean": 173.74200439453125, "std": 161.9979705810547, "min": -190.5802001953125, "p10": -21.709833526611323, "median": 185.96067810058594, "p90": 391.070751953125, "max": 546.478759765625, "pos_frac": 0.84375, "sample": [80.65116119384766, 111.52578735351562, 229.80262756347656, -12.08273696899414, 107.20259857177734, 546.478759765625, 10.492469787597656, 392.6512451171875, 199.69345092773438, 314.19464111328125, 313.4818420410156, 22.277530670166016, 313.14996337890625, 280.3659362792969, 173.25331115722656, 265.19952392578125, -23.493896484375, 433.0923156738281, -1.1036396026611328, 89.49010467529297, 185.80892944335938, 281.324462890625, 407.038330078125, 462.24749755859375, 453.38446044921875, 283.9797058105469, 209.9818878173828, -190.5802001953125, 54.60859680175781, 67.20171356201172, 173.9974365234375, 213.20037841796875, -33.68104553222656, 75.85594940185547, 64.69929504394531, 186.1124267578125, 118.01177215576172, -146.4693145751953, 269.9093017578125, 362.2779541015625, -17.547019958496094, 286.42559814453125, 227.26937866210938, 296.7339172363281, 117.35057067871094, 219.06845092773438, 35.944244384765625, 387.3829345703125, 338.18023681640625, 229.95489501953125, -43.9749870300293, -85.1856460571289, 22.44170379638672, 258.1360168457031, 60.00813293457031, -76.7552490234375, 27.832056045532227, 118.2278060913086, 54.82912063598633, 203.08697509765625, 304.4710998535156, 98.96511840820312, 485.76739501953125, 225.64332580566406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000221.npy"}
{"epoch": 0.3245227606461087, "step": 222, "batch_size": 64, "mean": 131.14132690429688, "std": 164.9459686279297, "min": -240.71627807617188, "p10": -62.12859573364256, "median": 117.2270736694336, "p90": 365.390545654297, "max": 563.9057006835938, "pos_frac": 0.78125, "sample": [-47.94312286376953, 451.5380554199219, -13.197402954101562, -116.34500122070312, -8.399345397949219, -89.44348907470703, -30.473175048828125, 138.2991180419922, 12.039764404296875, 235.892333984375, 106.62871551513672, 116.62525177001953, 441.81585693359375, 96.3502197265625, 117.82889556884766, 226.51910400390625, -240.71627807617188, 169.33609008789062, 259.0748596191406, 177.5336151123047, 161.0509490966797, 123.5244140625, -198.95751953125, 322.9994812011719, -134.47317504882812, 283.20196533203125, 499.5469970703125, 104.21006774902344, 62.073829650878906, 17.82811737060547, 164.54135131835938, 27.08624267578125, 168.58099365234375, -43.809112548828125, 232.73695373535156, 202.43641662597656, 60.0301513671875, 157.3495635986328, -20.559955596923828, 240.4510498046875, -27.47394561767578, 98.29322814941406, 82.97956085205078, 82.79231262207031, 339.34600830078125, 243.4346923828125, 105.47832489013672, 258.00201416015625, 154.5347900390625, 47.043846130371094, -68.20808410644531, 166.7415313720703, 376.552490234375, 70.94656372070312, 402.5824890136719, 187.43557739257812, 139.4557342529297, -100.94613647460938, 234.010986328125, 22.02434539794922, 563.9057006835938, 55.86474609375, 416.777099609375, 108.65898895263672], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000222.npy"}
{"epoch": 0.32599118942731276, "step": 223, "batch_size": 64, "mean": 110.50656127929688, "std": 180.68943786621094, "min": -348.51007080078125, "p10": -112.67254333496093, "median": 115.98501205444336, "p90": 326.2665283203125, "max": 612.6012573242188, "pos_frac": 0.75, "sample": [54.11720275878906, -76.1567153930664, 225.387939453125, 122.55429077148438, 143.59811401367188, 379.3805847167969, 206.27651977539062, 117.6386947631836, 217.8583526611328, 168.03848266601562, -153.22122192382812, 58.3416748046875, 95.69413757324219, 248.09292602539062, 316.66741943359375, 302.23272705078125, 133.2650146484375, 95.20175170898438, 564.7570190429688, -27.05063247680664, 137.4213104248047, -40.58922576904297, 81.82286834716797, -43.84872055053711, -59.08898162841797, 327.49578857421875, -139.8087921142578, 367.2979431152344, 105.0349349975586, -114.22870635986328, 7.3915863037109375, 10.4453125, 272.59722900390625, -348.51007080078125, 122.28388977050781, 145.689208984375, -209.1118621826172, 200.25274658203125, 74.97221374511719, 612.6012573242188, 252.0168914794922, -109.04149627685547, 184.98797607421875, 277.1788635253906, -131.77731323242188, -75.94505310058594, 289.3349609375, 327.6302795410156, 143.85523986816406, -2.3472061157226562, 49.27192687988281, 50.64186096191406, 33.69940185546875, 412.294677734375, 27.26634979248047, 33.817527770996094, -12.790121078491211, 4.583562850952148, 153.9457244873047, 114.33132934570312, 123.96355438232422, 158.9036865234375, -259.59539794921875, 323.39825439453125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000223.npy"}
{"epoch": 0.3274596182085169, "step": 224, "batch_size": 64, "mean": 127.66184997558594, "std": 186.03489685058594, "min": -319.9190368652344, "p10": -71.80010681152343, "median": 77.7532844543457, "p90": 365.2017181396485, "max": 685.304931640625, "pos_frac": 0.796875, "sample": [246.68516540527344, 5.297637939453125, 144.46058654785156, 19.08617401123047, -112.88957977294922, 581.3084716796875, 250.80262756347656, 70.7696762084961, 84.06426239013672, -319.9190368652344, 49.61688995361328, 186.86952209472656, 165.91372680664062, 151.95492553710938, 454.06439208984375, 88.91514587402344, 68.6418685913086, 199.09861755371094, 195.5523681640625, 685.304931640625, 305.5863952636719, 126.56954956054688, 353.2878112792969, 266.6205749511719, -124.34236145019531, -75.389404296875, 281.3416442871094, 71.44230651855469, 250.65115356445312, -28.490341186523438, 32.50962829589844, 41.5528564453125, 42.98829650878906, 624.318359375, 246.10919189453125, 254.51922607421875, 116.29581451416016, -79.5107192993164, 32.29530334472656, 27.424327850341797, 257.881591796875, 16.00372314453125, 48.84022521972656, -31.60455322265625, 109.36550903320312, -75.82250213623047, -59.49281311035156, 20.611160278320312, -118.8890380859375, 9.626270294189453, 194.30990600585938, 129.21060180664062, 370.30767822265625, -1.7796611785888672, 7.991207122802734, 66.37138366699219, 20.418792724609375, 335.10516357421875, 40.3216552734375, -63.425079345703125, 87.77058410644531, 399.61541748046875, 435.7938537597656, -9.55087661743164], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000224.npy"}
{"epoch": 0.328928046989721, "step": 225, "batch_size": 64, "mean": 149.67507934570312, "std": 185.7623291015625, "min": -265.8305969238281, "p10": -82.94855880737303, "median": 156.37598419189453, "p90": 385.3197509765625, "max": 721.7486572265625, "pos_frac": 0.734375, "sample": [245.9024658203125, 168.4517822265625, 425.02398681640625, 361.034912109375, 59.42229461669922, 117.60375213623047, 39.992095947265625, -114.542236328125, -136.03399658203125, 461.962890625, 365.80804443359375, 287.0470275878906, -146.97727966308594, -70.02300262451172, 274.26800537109375, -38.37589645385742, -131.27418518066406, 17.912038803100586, -8.465038299560547, 69.10301208496094, 120.47735595703125, 160.11807250976562, 240.4974365234375, 171.97909545898438, -10.019472122192383, 162.59783935546875, 61.56501770019531, 173.83425903320312, -41.1368522644043, 271.7713928222656, 348.81298828125, 172.97317504882812, 229.50257873535156, 111.31389617919922, -54.32950210571289, 479.0721435546875, 449.5184326171875, -4.884498596191406, 249.6204833984375, 412.2642822265625, 102.90068054199219, 338.117919921875, 214.6693115234375, -24.24065589904785, 152.63389587402344, 142.32736206054688, 721.7486572265625, 288.15087890625, 205.41470336914062, 100.12477111816406, 387.08660888671875, 14.06109619140625, 234.32772827148438, -97.94659423828125, -88.48808288574219, 381.19708251953125, 284.1121826171875, 180.53982543945312, 161.41270446777344, -265.8305969238281, 124.66890716552734, 146.06277465820312, -55.74158477783203, -21.493988037109375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000225.npy"}
{"epoch": 0.3303964757709251, "step": 226, "batch_size": 64, "mean": 123.3972396850586, "std": 177.0221710205078, "min": -208.48182678222656, "p10": -37.687967300415025, "median": 73.20746231079102, "p90": 360.8680328369141, "max": 621.6605224609375, "pos_frac": 0.796875, "sample": [621.6605224609375, -154.27615356445312, -44.23849105834961, 28.219024658203125, -3.5227890014648438, 174.5117645263672, 6.303924560546875, 69.87669372558594, 183.5472412109375, -13.598857879638672, 129.0299072265625, 15.536224365234375, 2.165771484375, 57.60186004638672, -177.97402954101562, 79.77545166015625, -110.85807800292969, 379.54022216796875, 366.6231384277344, 14.175254821777344, 344.958740234375, 347.439453125, 76.5382308959961, 137.86170959472656, 179.23277282714844, 321.62738037109375, 26.195640563964844, -22.403411865234375, 118.1170654296875, 449.0212707519531, 276.91729736328125, -75.81541442871094, 193.5406036376953, 78.86422729492188, 67.05088806152344, 576.3599853515625, 244.80935668945312, 567.1422729492188, 302.3287048339844, -6.477806091308594, -2.3505783081054688, 151.83480834960938, -208.48182678222656, 152.9854736328125, 42.97336196899414, 123.69512939453125, 22.77576446533203, 25.5452880859375, 6.835046768188477, 68.48125457763672, -5.8948974609375, 266.6064453125, 108.74530029296875, 198.5767822265625, 55.51879119873047, 126.81739807128906, -81.228759765625, 5.935007095336914, 511.82061767578125, 54.04322814941406, 45.34253692626953, 127.47450256347656, 50.703399658203125, 221.26171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000226.npy"}
{"epoch": 0.33186490455212925, "step": 227, "batch_size": 64, "mean": 146.48440551757812, "std": 176.49774169921875, "min": -376.98760986328125, "p10": -23.854394721984846, "median": 117.53427124023438, "p90": 373.92160644531253, "max": 603.5208129882812, "pos_frac": 0.859375, "sample": [316.4310607910156, 104.10533142089844, 169.578369140625, 392.1905517578125, 221.83212280273438, 44.696929931640625, 53.92756652832031, 127.29625701904297, 295.8154296875, 18.860198974609375, 92.62342071533203, 110.48056030273438, 192.85671997070312, 33.40320587158203, -30.649677276611328, 28.547386169433594, 176.11575317382812, 97.51725769042969, 216.84263610839844, 603.5208129882812, 153.1540985107422, 67.86616516113281, 348.39501953125, 206.87063598632812, 42.864646911621094, 304.759033203125, 168.56298828125, 252.06890869140625, 98.18740844726562, 95.71192932128906, 200.87649536132812, 90.37028503417969, 377.07684326171875, 6.106098175048828, -82.824462890625, -128.7805633544922, 202.60531616210938, 225.28005981445312, 526.3248291015625, 124.58798217773438, 205.6584014892578, 444.700927734375, -122.04133605957031, 587.4724731445312, 0.3931255340576172, 19.420509338378906, 203.534423828125, 38.898963928222656, 231.48574829101562, 42.769325256347656, 128.4702606201172, 366.55938720703125, 96.730224609375, -2.0366668701171875, 5.397699356079102, -54.40076446533203, 365.7834777832031, -376.98760986328125, -7.998735427856445, 100.484619140625, -74.44701385498047, 15.241067886352539, 127.88677978515625, 485.9710998535156], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000227.npy"}
{"epoch": 0.3333333333333333, "step": 228, "batch_size": 64, "mean": 106.93231201171875, "std": 170.448486328125, "min": -253.28143310546875, "p10": -108.57582702636718, "median": 100.74028396606445, "p90": 326.8691375732422, "max": 506.75433349609375, "pos_frac": 0.734375, "sample": [-193.344970703125, -239.5552978515625, -253.28143310546875, 239.03005981445312, 152.349365234375, 360.42169189453125, -211.15670776367188, 191.5894775390625, 220.51699829101562, -168.16207885742188, 36.805389404296875, 300.6217346191406, 506.75433349609375, 309.02435302734375, 475.256591796875, 349.6685791015625, 353.61590576171875, 186.19964599609375, 126.02494812011719, -39.346961975097656, -50.20527648925781, 328.2321472167969, 120.60013580322266, 323.68878173828125, 193.01373291015625, 171.1381072998047, 136.64051818847656, 20.402141571044922, -110.15070343017578, 101.09336853027344, 54.39976501464844, 155.1749725341797, 11.149703979492188, 226.04806518554688, 210.16815185546875, 112.38124084472656, 125.88709259033203, 91.68896484375, 33.55815124511719, 15.44842529296875, 84.1996841430664, 238.2147216796875, -87.5738525390625, -12.224884033203125, 93.23832702636719, 95.30294036865234, 448.8019714355469, 25.627716064453125, 90.89511108398438, 63.856544494628906, 271.14617919921875, -15.94888687133789, -24.089435577392578, 204.04452514648438, 235.9036407470703, 161.16458129882812, 206.95811462402344, 100.38719940185547, -104.90111541748047, -15.462677001953125, -20.06177520751953, 36.634979248046875, -165.05426025390625, -40.78062438964844], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000228.npy"}
{"epoch": 0.33480176211453744, "step": 229, "batch_size": 64, "mean": 113.62403869628906, "std": 166.81871032714844, "min": -404.2523193359375, "p10": -65.91524124145506, "median": 103.11714553833008, "p90": 337.24053039550785, "max": 559.84130859375, "pos_frac": 0.78125, "sample": [12.698410034179688, 14.722021102905273, -125.2794189453125, 419.30035400390625, 55.00871658325195, -161.39987182617188, 113.2117691040039, 184.62574768066406, 83.28541564941406, 219.56350708007812, 269.1954650878906, 373.06549072265625, 110.55105590820312, 334.0567321777344, 86.66862487792969, 96.56339263916016, 236.59939575195312, 111.81511688232422, 24.483417510986328, 334.5603942871094, -115.51765441894531, 53.610443115234375, 1.0598335266113281, 231.14613342285156, 145.33900451660156, 45.41387176513672, 559.84130859375, -198.1646728515625, 42.79722595214844, 243.62191772460938, 177.94952392578125, 243.61151123046875, 199.43844604492188, 380.21826171875, 368.8878173828125, 44.928794860839844, 338.38916015625, 19.99327850341797, 109.6708984375, -70.97848510742188, 69.8901138305664, 124.5924072265625, 54.710472106933594, 201.7554168701172, -37.777854919433594, 168.04747009277344, 223.015869140625, 228.3735809326172, -33.0072021484375, -123.49913024902344, 139.66392517089844, 444.143310546875, 32.07762145996094, 77.58999633789062, 67.82865142822266, 164.8006134033203, -11.923103332519531, 183.6842041015625, -15.723037719726562, 174.67446899414062, -404.2523193359375, -0.7929668426513672, -16.384933471679688, -54.10100555419922], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000229.npy"}
{"epoch": 0.33627019089574156, "step": 230, "batch_size": 64, "mean": 168.3480224609375, "std": 165.43734741210938, "min": -310.6220703125, "p10": -17.190424346923823, "median": 153.92749786376953, "p90": 413.2921417236328, "max": 516.3578491210938, "pos_frac": 0.84375, "sample": [88.61378479003906, 372.89453125, -19.446104049682617, 303.67047119140625, 215.76266479492188, 83.2682876586914, -11.92717170715332, 231.1378936767578, 110.47805786132812, 145.11709594726562, 213.63941955566406, 172.45413208007812, 4.495262145996094, 235.5465087890625, 30.75396728515625, 113.63424682617188, -42.79026794433594, 79.95281982421875, 85.42617797851562, 161.69606018066406, 31.7742919921875, -101.31554412841797, -7.367393493652344, 218.1400146484375, 148.899658203125, 177.01400756835938, 227.3003692626953, 356.9573974609375, 312.8832092285156, 152.50579833984375, 92.36760711669922, 36.981163024902344, 346.37371826171875, 473.7898254394531, 155.940185546875, -119.7439193725586, 136.92160034179688, 176.51815795898438, 252.19442749023438, 464.50421142578125, -46.455387115478516, -1.4203948974609375, -310.6220703125, -44.731239318847656, 40.53851318359375, 47.37474060058594, 410.5885925292969, 127.65065002441406, 371.18115234375, 448.4876708984375, 155.3491973876953, 122.22956848144531, 126.10289001464844, 24.7611026763916, 414.4508056640625, 289.600341796875, 516.3578491210938, 239.8751678466797, 416.90386962890625, 214.1804962158203, 219.86512756347656, 145.99847412109375, 231.2862091064453, 507.7037353515625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000230.npy"}
{"epoch": 0.3377386196769457, "step": 231, "batch_size": 64, "mean": 115.2940673828125, "std": 153.0771026611328, "min": -142.38638305664062, "p10": -44.70459976196289, "median": 88.94082260131836, "p90": 336.02146606445314, "max": 640.080810546875, "pos_frac": 0.796875, "sample": [466.05535888671875, 107.40956115722656, -42.04430389404297, 291.6224365234375, 14.995643615722656, 76.35929107666016, 334.7377624511719, 15.5653076171875, 65.84675598144531, 111.24124908447266, 103.3004150390625, 337.99505615234375, -77.00160217285156, 73.5557861328125, 495.4093933105469, 143.82440185546875, 194.5262908935547, 68.60562133789062, 137.88961791992188, 208.10772705078125, 394.8499450683594, 53.37381362915039, 111.35062408447266, -3.9037017822265625, 52.41864013671875, -63.12799835205078, 336.5716247558594, -36.31161880493164, 211.83673095703125, 640.080810546875, 434.8002014160156, 238.3828125, 122.755859375, 21.592153549194336, -142.38638305664062, 67.83828735351562, 102.21180725097656, 157.98013305664062, 98.4420166015625, -45.8447265625, -3.14593505859375, 40.55309295654297, 202.37716674804688, 75.0486068725586, -3.2313766479492188, 118.1410140991211, -26.197540283203125, -108.60491180419922, 43.790218353271484, 142.38729858398438, 317.40606689453125, 159.12478637695312, 79.43962860107422, 10.949779510498047, -53.14950942993164, 127.80859375, -98.71807098388672, 125.50633239746094, 26.030059814453125, 18.662994384765625, 10.336822509765625, 147.97689819335938, 119.41323852539062, 26.002700805664062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000231.npy"}
{"epoch": 0.3392070484581498, "step": 232, "batch_size": 64, "mean": 93.408935546875, "std": 152.89102172851562, "min": -158.70159912109375, "p10": -68.62047271728515, "median": 54.66422462463379, "p90": 286.7632629394532, "max": 563.7075805664062, "pos_frac": 0.71875, "sample": [166.70834350585938, 106.46825408935547, 10.325048446655273, 55.61072540283203, 24.31134796142578, -111.94419860839844, 297.8712463378906, 257.8921813964844, -29.56645965576172, 249.27780151367188, 51.371681213378906, 384.28912353515625, 89.3871078491211, -74.6854019165039, -24.850021362304688, 58.125038146972656, -12.381454467773438, 188.24383544921875, -104.08805847167969, 72.60277557373047, 152.6879119873047, 563.7075805664062, 41.609710693359375, 62.5760498046875, 144.17791748046875, -48.68256759643555, 552.6788940429688, -2.5488128662109375, 32.94293212890625, 244.80368041992188, 155.91195678710938, 239.53115844726562, -63.334564208984375, 75.49346923828125, -91.76852416992188, 110.93992614746094, -92.17072296142578, 13.500343322753906, -65.145263671875, 449.0487060546875, 36.754852294921875, -59.436302185058594, 5.875648498535156, -70.10984802246094, 16.092182159423828, 91.6563720703125, 41.1602783203125, -158.70159912109375, 254.58863830566406, 133.980224609375, 333.57208251953125, 25.234554290771484, -4.654876708984375, 311.8037109375, 229.06097412109375, -27.544923782348633, 137.56600952148438, 260.8446350097656, -16.460668563842773, 50.33531951904297, 53.71772384643555, 38.03885269165039, 88.6029052734375, 75.26644134521484], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000232.npy"}
{"epoch": 0.3406754772393539, "step": 233, "batch_size": 64, "mean": 147.81594848632812, "std": 209.2672882080078, "min": -197.78182983398438, "p10": -67.90078887939451, "median": 99.70556259155273, "p90": 464.7687408447266, "max": 681.890625, "pos_frac": 0.734375, "sample": [13.307031631469727, 10.686298370361328, 495.99102783203125, -33.05616760253906, 175.4317626953125, 204.35638427734375, -184.71490478515625, -110.5031967163086, 0.7794036865234375, -190.50306701660156, 115.78132629394531, 62.41007995605469, 162.71800231933594, 267.19219970703125, 273.5887451171875, 681.890625, 126.5589370727539, 87.33511352539062, 507.16650390625, 303.974609375, 66.63761901855469, 180.443603515625, 76.79664611816406, 508.4079284667969, 24.85692596435547, 288.22113037109375, 4.590034484863281, -73.78384399414062, -183.70782470703125, -42.828006744384766, 159.99107360839844, 523.54931640625, 190.80496215820312, 167.95367431640625, 304.4149169921875, 455.9830017089844, 23.705251693725586, 441.316162109375, -54.17366027832031, 51.09071350097656, 112.07601165771484, 399.68060302734375, -5.947032928466797, 302.01397705078125, 64.45639038085938, 374.1571044921875, -39.7769889831543, -197.78182983398438, -133.6774139404297, -17.150009155273438, 48.733787536621094, -27.495140075683594, 351.989013671875, 81.1866455078125, -52.29302978515625, 203.8192596435547, -29.129558563232422, 468.5340576171875, -33.909488677978516, 602.5341186523438, 336.3782958984375, 66.79499816894531, 307.5830078125, 192.78439331054688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000233.npy"}
{"epoch": 0.342143906020558, "step": 234, "batch_size": 64, "mean": 152.16229248046875, "std": 164.7661590576172, "min": -154.03648376464844, "p10": -31.820835876464834, "median": 139.5507583618164, "p90": 373.64287719726565, "max": 622.61572265625, "pos_frac": 0.828125, "sample": [-75.3254165649414, 255.9357147216797, 258.7070007324219, 236.62921142578125, 304.7580261230469, 90.0007553100586, 150.43197631835938, 204.1946563720703, 372.9335632324219, 5.336931228637695, 297.51336669921875, 142.5704803466797, -154.03648376464844, -0.07601547241210938, 74.3840560913086, 252.89886474609375, 79.56646728515625, -64.80005645751953, 64.7054443359375, 327.8050537109375, 622.61572265625, 161.26547241210938, 8.168676376342773, -21.30445098876953, 57.896934509277344, -20.250289916992188, 174.28411865234375, 338.7830505371094, 15.969329833984375, 402.5628662109375, -11.354757308959961, 125.55181884765625, 419.68707275390625, 141.80108642578125, 484.387451171875, 120.40744018554688, 45.91643524169922, 168.81912231445312, 74.58537292480469, 240.3331756591797, 253.00778198242188, 149.41195678710938, 198.76034545898438, 83.38372039794922, 35.406898498535156, 76.94235229492188, -97.93704223632812, 18.40814971923828, 254.66281127929688, 96.97125244140625, 106.01322174072266, 432.22369384765625, 573.922119140625, -36.327857971191406, 307.1825866699219, 161.42469787597656, -63.50543212890625, 13.541143417358398, 178.8041229248047, 50.39582824707031, 137.30043029785156, -117.17578125, 373.9468688964844, 177.36328125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000234.npy"}
{"epoch": 0.3436123348017621, "step": 235, "batch_size": 64, "mean": 98.66987609863281, "std": 164.05435180664062, "min": -345.05682373046875, "p10": -69.6199462890625, "median": 98.59652709960938, "p90": 270.59411926269536, "max": 527.5606079101562, "pos_frac": 0.71875, "sample": [26.325531005859375, 235.77731323242188, -345.05682373046875, 25.461212158203125, 156.70594787597656, 527.5606079101562, 137.24749755859375, 318.4308776855469, -68.45822143554688, 187.87493896484375, -170.60153198242188, -18.951568603515625, 127.23617553710938, 260.0826416015625, 44.918121337890625, 23.600372314453125, -45.59941101074219, -27.850257873535156, 92.28678894042969, -253.63192749023438, 25.97994613647461, 25.25885009765625, -15.6046142578125, 80.84464263916016, 350.25885009765625, 221.5537872314453, -15.490226745605469, -64.45223999023438, 145.33316040039062, 207.21665954589844, 133.46743774414062, 93.90702819824219, 274.8110656738281, -66.6351318359375, 118.75114440917969, 191.06040954589844, -0.6266803741455078, 259.8423767089844, 43.40122985839844, 140.42440795898438, 27.939159393310547, 102.84375, 459.9454040527344, 250.92666625976562, 192.37652587890625, 175.8575897216797, 187.70761108398438, -46.307838439941406, -158.93618774414062, 155.087158203125, 494.81024169921875, 260.75457763671875, -135.00909423828125, 28.888168334960938, 127.75790405273438, -2.8569469451904297, -82.90973663330078, 188.531494140625, 94.34930419921875, 16.081405639648438, 198.14141845703125, 116.86146545410156, 349.4894104003906, -70.11782836914062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000235.npy"}
{"epoch": 0.34508076358296624, "step": 236, "batch_size": 64, "mean": 156.1163330078125, "std": 139.2214813232422, "min": -72.81287384033203, "p10": -16.456497573852523, "median": 133.8141098022461, "p90": 363.448486328125, "max": 457.1396484375, "pos_frac": 0.875, "sample": [366.1063537597656, 65.27660369873047, 6.0576019287109375, 164.68157958984375, 205.57960510253906, 142.9398956298828, 395.4403991699219, 52.37611770629883, 75.75384521484375, 143.03924560546875, 206.0634765625, 357.2467956542969, 104.47386932373047, 241.19338989257812, 302.0267333984375, 309.7246398925781, 80.68917846679688, 246.58665466308594, 10.799430847167969, 370.200927734375, 137.66815185546875, -26.15728759765625, 76.79036712646484, 303.0418395996094, 254.07186889648438, 10.40005874633789, 134.43402099609375, 404.880615234375, 235.14193725585938, 49.729713439941406, 448.8992614746094, 133.19419860839844, -50.92686462402344, 230.85679626464844, 18.820720672607422, 258.7034912109375, 86.76276397705078, 457.1396484375, 107.02860260009766, -23.293018341064453, 125.0193862915039, 97.50232696533203, 33.853919982910156, 47.72447967529297, 9.902027130126953, 236.0867462158203, 102.10063171386719, 71.25369262695312, -72.81287384033203, -57.838356018066406, 123.23336791992188, 432.5769958496094, 351.5206298828125, 124.83689880371094, 211.50848388671875, 136.6212158203125, 294.5423889160156, 97.85514831542969, -0.5046157836914062, 138.8768768310547, -51.76026916503906, 307.4195556640625, -48.08540725708008, 186.56837463378906], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000236.npy"}
{"epoch": 0.3465491923641703, "step": 237, "batch_size": 64, "mean": 119.02336120605469, "std": 176.49856567382812, "min": -354.4436340332031, "p10": -75.06575393676758, "median": 110.33648681640625, "p90": 373.65242919921906, "max": 663.691650390625, "pos_frac": 0.765625, "sample": [165.9381866455078, 199.38858032226562, 182.05441284179688, 178.092529296875, -3.174835205078125, 107.77982330322266, 112.89315032958984, 161.620849609375, 26.467296600341797, 191.43289184570312, 73.16986846923828, 31.525192260742188, 420.96826171875, 167.38287353515625, 231.82614135742188, -64.89115905761719, -144.50392150878906, 144.81307983398438, -70.26166534423828, 101.81185913085938, 162.82168579101562, -88.5130615234375, 405.141845703125, 663.691650390625, -84.28224182128906, 412.544921875, 275.6929931640625, -21.0072021484375, 27.092788696289062, 12.851009368896484, 461.8863525390625, 86.94349670410156, 218.5921173095703, 228.48719787597656, 31.67346954345703, 233.4412384033203, 11.338859558105469, 116.31732940673828, 300.1771240234375, -34.701759338378906, 258.0851745605469, 53.670982360839844, 138.76229858398438, 224.35678100585938, 444.34368896484375, 47.249176025390625, 9.968719482421875, 3.1319198608398438, -354.4436340332031, 294.0581970214844, -54.460350036621094, 66.84626770019531, -168.06898498535156, -44.427734375, 209.39532470703125, 278.59173583984375, -77.12464904785156, 432.29864501953125, 4.819938659667969, -116.84262084960938, -57.8770751953125, 140.53964233398438, 62.748992919921875, 187.34884643554688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000237.npy"}
{"epoch": 0.34801762114537443, "step": 238, "batch_size": 64, "mean": 124.41336059570312, "std": 177.7018280029297, "min": -219.54782104492188, "p10": -36.68266448974609, "median": 92.81853485107422, "p90": 302.9472137451172, "max": 753.2879028320312, "pos_frac": 0.78125, "sample": [-25.961746215820312, -43.753578186035156, 6.1311798095703125, 139.7196807861328, 22.74435806274414, -4.0528717041015625, 144.4666290283203, 277.47491455078125, -128.7389678955078, 406.92987060546875, 74.73651885986328, 167.13011169433594, 64.92344665527344, 58.75777053833008, -9.957239151000977, 165.8577117919922, 301.3523254394531, -80.42274475097656, 60.99418640136719, 50.14383316040039, 16.512540817260742, 708.85595703125, 35.56715393066406, 148.9045867919922, -33.67054748535156, 544.338134765625, 111.8326416015625, -7.423973083496094, 137.9960174560547, -111.95646667480469, 118.51058959960938, 102.20828247070312, 180.04547119140625, 139.43905639648438, 351.648681640625, 16.433712005615234, 78.19662475585938, 133.9624481201172, 753.2879028320312, 175.29962158203125, 0.4483451843261719, 5.06109619140625, 81.21305847167969, -98.57604217529297, -14.762100219726562, 277.26348876953125, 292.6104736328125, 163.00762939453125, 83.42878723144531, 458.6918029785156, 232.1310272216797, 22.621185302734375, 222.23666381835938, 252.803466796875, -219.54782104492188, 303.6307373046875, 207.53713989257812, 61.48603820800781, 10.204048156738281, -27.301925659179688, 163.43118286132812, 166.94970703125, -37.97357177734375, 107.39708709716797], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000238.npy"}
{"epoch": 0.34948604992657856, "step": 239, "batch_size": 64, "mean": 132.85975646972656, "std": 162.14022827148438, "min": -265.1212463378906, "p10": -38.982910537719725, "median": 125.44805145263672, "p90": 355.8350006103516, "max": 462.55511474609375, "pos_frac": 0.8125, "sample": [364.75848388671875, 197.00318908691406, -152.8954620361328, 352.5181884765625, -106.96685791015625, 147.69244384765625, 52.69781494140625, 121.3311767578125, -40.3751335144043, -148.18112182617188, 46.85954666137695, 43.870635986328125, 79.09823608398438, 280.21014404296875, 74.75340270996094, 2.4701156616210938, 449.3603210449219, 129.56492614746094, 182.97775268554688, 139.5975341796875, 210.36351013183594, 194.3724822998047, 239.39112854003906, -62.793949127197266, 29.20848846435547, 38.762542724609375, 196.20033264160156, 382.9915466308594, 219.3038787841797, 111.7098388671875, -16.910558700561523, 25.496917724609375, 354.59515380859375, 420.9925537109375, -182.56370544433594, 41.67670440673828, -2.2796859741210938, 329.65948486328125, 7.1855316162109375, 356.3663635253906, 53.361148834228516, 353.6201171875, 316.24609375, 144.74285888671875, -35.73439025878906, 30.394813537597656, -265.1212463378906, 156.18173217773438, -35.4799919128418, 156.1285858154297, 158.39849853515625, 266.05401611328125, 317.023193359375, 71.71446228027344, 462.55511474609375, 210.50340270996094, 376.3265380859375, 87.78158569335938, 117.52110290527344, 134.783203125, 62.84027099609375, -20.97142791748047, 262.94879150390625, 11.131841659545898], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000239.npy"}
{"epoch": 0.3509544787077827, "step": 240, "batch_size": 64, "mean": 116.85966491699219, "std": 191.6036834716797, "min": -250.8791961669922, "p10": -187.09870910644526, "median": 126.83114242553711, "p90": 382.52255249023443, "max": 580.6845703125, "pos_frac": 0.765625, "sample": [129.4651336669922, 148.29554748535156, 184.08724975585938, 96.1762466430664, 96.04713439941406, 368.9940185546875, 166.86563110351562, 435.616943359375, -222.53494262695312, -61.10784912109375, 185.2178955078125, 131.58689880371094, -230.40890502929688, 213.56927490234375, 388.32049560546875, 104.1561050415039, -206.20040893554688, 68.36947631835938, 34.933372497558594, 257.3557434082031, 145.99005126953125, 427.1978759765625, 156.93380737304688, 167.06399536132812, 68.2269287109375, 12.185840606689453, -223.87167358398438, 125.92833709716797, 82.4202880859375, 497.0020751953125, 245.41249084472656, 101.33779907226562, 97.67231750488281, 21.114585876464844, -210.435546875, 365.1454162597656, 219.83473205566406, 174.75772094726562, -208.99972534179688, 80.52021026611328, -51.64986801147461, 331.6409912109375, -142.528076171875, 47.42838668823242, 580.6845703125, -118.74533081054688, 283.2506103515625, -35.78631591796875, 250.6502685546875, 186.5602569580078, 127.73394775390625, 434.414794921875, -35.971832275390625, 120.1013412475586, -116.28926086425781, 140.28765869140625, 88.4758529663086, -49.030609130859375, 138.41104125976562, 201.18943786621094, 48.779449462890625, 168.4752960205078, 497.5723571777344, -250.8791961669922], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000240.npy"}
{"epoch": 0.3524229074889868, "step": 241, "batch_size": 64, "mean": 185.10760498046875, "std": 210.5679931640625, "min": -210.94956970214844, "p10": -51.17555694580077, "median": 156.93976593017578, "p90": 426.15867004394534, "max": 878.0677490234375, "pos_frac": 0.828125, "sample": [101.26941680908203, 77.48051452636719, -204.21392822265625, -22.176340103149414, 42.9234619140625, 254.26776123046875, 283.1720275878906, 427.3853759765625, 92.97140502929688, 199.96240234375, 118.24041748046875, -57.9786376953125, 8.109996795654297, 537.5697631835938, -42.47267150878906, 157.175048828125, -35.20258712768555, 353.7401428222656, 174.8068084716797, 163.5609130859375, 391.336181640625, 181.13818359375, 91.28571319580078, 128.12815856933594, 878.0677490234375, 423.2963562011719, 212.95925903320312, 150.06600952148438, 594.311767578125, 44.61102294921875, 844.0542602539062, 485.64678955078125, 339.047119140625, 82.2935562133789, 85.99192810058594, 156.70448303222656, 216.08148193359375, 98.28336334228516, -54.905364990234375, 187.2953643798828, 130.79595947265625, 54.57603454589844, 119.81106567382812, -9.758132934570312, 269.15008544921875, 92.84750366210938, 384.831298828125, -65.57958984375, 248.45297241210938, 211.85418701171875, 73.53021240234375, 47.19579315185547, 286.5396728515625, 485.6370544433594, 312.4544677734375, 186.25668334960938, -130.70233154296875, 310.91925048828125, 226.07684326171875, 279.66973876953125, -210.94956970214844, 395.7408447265625, 91.63664245605469, -110.38546752929688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000241.npy"}
{"epoch": 0.35389133627019087, "step": 242, "batch_size": 64, "mean": 129.2273712158203, "std": 177.03660583496094, "min": -258.25152587890625, "p10": -37.164443206787105, "median": 111.30123901367188, "p90": 305.4880340576172, "max": 744.157958984375, "pos_frac": 0.78125, "sample": [-39.54485321044922, 42.07912826538086, 154.44131469726562, 291.24139404296875, 55.630855560302734, -258.25152587890625, 0.490936279296875, -9.697647094726562, 305.711669921875, 130.1017303466797, 30.396886825561523, 261.142578125, 615.8280639648438, 7.573118209838867, -26.783546447753906, 65.92882537841797, 142.95510864257812, 109.69329833984375, 97.026123046875, 17.133512496948242, 172.3615264892578, 116.39736938476562, 249.92172241210938, 195.63116455078125, 247.54461669921875, -23.41295623779297, 121.88335418701172, 110.79705047607422, 111.80542755126953, -62.402984619140625, 177.17520141601562, 206.18594360351562, 88.34888458251953, -135.8643798828125, 278.6883239746094, 104.30582427978516, 304.9662170410156, 433.27349853515625, -4.641777038574219, 82.29971313476562, 124.21694946289062, 189.0430145263672, -21.348251342773438, 744.157958984375, 16.167945861816406, 86.85834503173828, 523.58544921875, 119.86515808105469, -31.610153198242188, 19.192277908325195, -127.28539276123047, 371.18048095703125, 188.28700256347656, -21.43518829345703, 91.52931213378906, 128.659423828125, 165.62921142578125, -162.34568786621094, 238.4470672607422, 219.4952392578125, 13.884973526000977, 492.8183898925781, 194.05999755859375, -60.863014221191406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000242.npy"}
{"epoch": 0.355359765051395, "step": 243, "batch_size": 64, "mean": 160.332763671875, "std": 187.10276794433594, "min": -299.8633117675781, "p10": -38.77735195159911, "median": 127.98522186279297, "p90": 414.69776000976566, "max": 768.7969360351562, "pos_frac": 0.859375, "sample": [358.18634033203125, 26.24685287475586, 81.93497467041016, 467.0438232421875, 190.7900390625, 65.39290618896484, 133.8889617919922, 180.27745056152344, -44.49192428588867, -89.59894561767578, -56.917755126953125, 68.29220581054688, 271.5247497558594, 41.24401092529297, -150.47328186035156, 42.591156005859375, -89.52752685546875, 323.47625732421875, 133.5105438232422, 170.70883178710938, -299.8633117675781, 41.06404113769531, 159.39096069335938, 27.851341247558594, 6.514520645141602, 332.97161865234375, 75.62216186523438, 196.49868774414062, 97.81896209716797, 416.58197021484375, 318.46490478515625, 269.44146728515625, -25.443349838256836, 110.83837890625, -93.82940673828125, 361.67108154296875, 266.4784851074219, 517.1397705078125, 339.60589599609375, 281.1604919433594, 316.2083740234375, 106.98389434814453, 449.63458251953125, 768.7969360351562, 416.950439453125, 41.979644775390625, 142.32733154296875, 59.18764114379883, 410.30126953125, -22.548126220703125, 168.31466674804688, 208.44290161132812, 37.77960205078125, 122.45989990234375, 561.5787353515625, 49.21410369873047, 157.33566284179688, 89.43531799316406, 35.817222595214844, 274.5864562988281, 25.861984252929688, 57.594390869140625, 24.60871696472168, 234.366943359375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000243.npy"}
{"epoch": 0.3568281938325991, "step": 244, "batch_size": 64, "mean": 159.19573974609375, "std": 165.59078979492188, "min": -187.62020874023438, "p10": -12.223080825805662, "median": 136.781982421875, "p90": 401.69468383789075, "max": 593.2811889648438, "pos_frac": 0.859375, "sample": [104.97950744628906, -1.5110397338867188, 107.020263671875, 148.53262329101562, 8.802410125732422, 58.77297592163086, 476.8541564941406, 132.5714874267578, -55.0615234375, 23.3350830078125, 209.5088653564453, 212.46151733398438, 8.2071533203125, 310.93170166015625, 41.390289306640625, 131.7325897216797, -9.947036743164062, 447.1741943359375, 39.807132720947266, 84.8429946899414, -187.62020874023438, 133.26983642578125, 269.21826171875, -71.56976318359375, 213.40884399414062, 33.26039123535156, 171.03140258789062, 255.67236328125, 143.5839080810547, 42.085357666015625, 252.50401306152344, 111.49201202392578, 330.9139404296875, 337.36151123046875, 463.7391357421875, 593.2811889648438, 196.3140411376953, 412.9021301269531, 47.540992736816406, 591.9479370117188, 222.3182373046875, -13.198528289794922, 357.8085021972656, 185.04818725585938, 140.29412841796875, -68.29467010498047, 140.3135528564453, -49.24723815917969, 9.681716918945312, 231.88221740722656, 53.85890197753906, 9.433937072753906, 163.90194702148438, 84.81945037841797, 51.19000244140625, 142.91558837890625, 375.5439758300781, 214.91864013671875, -85.16596984863281, 126.50119018554688, 98.645263671875, 473.77191162109375, 220.9596710205078, 279.88348388671875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000244.npy"}
{"epoch": 0.35829662261380324, "step": 245, "batch_size": 64, "mean": 127.57707214355469, "std": 185.16285705566406, "min": -452.5479431152344, "p10": -44.80689888000486, "median": 102.97591400146484, "p90": 335.1096771240235, "max": 630.4524536132812, "pos_frac": 0.78125, "sample": [120.09369659423828, 433.755126953125, 190.03945922851562, 169.1457061767578, 196.72369384765625, 143.65805053710938, 64.00389099121094, -22.202560424804688, 206.18634033203125, 106.28744506835938, 89.23031616210938, 107.19245910644531, 151.24070739746094, 24.32986831665039, 139.3000030517578, 180.8209228515625, 219.80331420898438, 319.9544677734375, 86.30996704101562, 42.501033782958984, 546.278076171875, 143.6211700439453, -20.641265869140625, 368.31744384765625, 81.47234344482422, -9.374679565429688, 238.01637268066406, -54.49447250366211, -204.29661560058594, -8.278221130371094, 307.5576171875, 152.50442504882812, 630.4524536132812, -171.58778381347656, 186.5391845703125, 25.901390075683594, 13.083457946777344, 99.66438293457031, 312.411376953125, 74.17161560058594, -17.27633285522461, 208.22659301757812, 146.6077423095703, -69.60604858398438, 60.94348907470703, -452.5479431152344, 295.60321044921875, 292.24591064453125, 33.22925567626953, 228.3184356689453, 341.6047668457031, 27.078262329101562, 620.2882690429688, -94.0361328125, 3.686786651611328, 524.2325439453125, 78.55296325683594, 106.9354019165039, -11.498809814453125, -93.35260009765625, 81.3680419921875, 79.14985656738281, 97.3437728881836, -1.8572616577148438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000245.npy"}
{"epoch": 0.35976505139500736, "step": 246, "batch_size": 64, "mean": 130.9169921875, "std": 193.5309295654297, "min": -262.6287841796875, "p10": -84.98112182617186, "median": 92.43461608886719, "p90": 381.2198822021484, "max": 814.3658447265625, "pos_frac": 0.78125, "sample": [172.25625610351562, 382.1212463378906, 144.57708740234375, 96.12281036376953, 424.22613525390625, 195.90597534179688, 814.3658447265625, -158.21774291992188, 46.751949310302734, -262.6287841796875, 57.252891540527344, 56.313133239746094, 68.46656799316406, -0.41354942321777344, 188.95977783203125, 379.11669921875, 332.145263671875, 66.66754913330078, 42.93336486816406, 57.91389465332031, 43.10197448730469, 394.5085754394531, 157.07427978515625, -97.2931137084961, 109.32237243652344, -100.19137573242188, 226.26806640625, -56.481719970703125, 331.7131042480469, -117.17687225341797, 4.591339111328125, -27.102012634277344, 109.20269775390625, 99.51960754394531, 327.347412109375, 31.30583953857422, 15.27928352355957, -3.787456512451172, 7.5104827880859375, 230.06002807617188, 406.7220764160156, 104.73542022705078, 40.932167053222656, 227.55030822753906, 93.27296447753906, 203.32345581054688, 94.8017578125, -67.81144714355469, -0.6565132141113281, 231.30491638183594, -92.33955383300781, 91.59626770019531, 318.0250549316406, 183.5832061767578, 26.343536376953125, 53.29083251953125, 14.61396598815918, 607.1812744140625, 321.8456115722656, 623.6806640625, -34.389076232910156, 220.9169921875, -100.14738464355469, 20.701847076416016], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000246.npy"}
{"epoch": 0.36123348017621143, "step": 247, "batch_size": 64, "mean": 179.29928588867188, "std": 200.50534057617188, "min": -328.31427001953125, "p10": -28.274651336669912, "median": 157.73197174072266, "p90": 434.3281127929689, "max": 701.7332763671875, "pos_frac": 0.875, "sample": [-284.9410400390625, 105.26956176757812, -72.78578186035156, 385.75885009765625, 66.27680969238281, 484.15460205078125, 173.07379150390625, 390.80096435546875, -32.884315490722656, 400.9373779296875, 380.2353820800781, 486.8631286621094, 199.39593505859375, 282.72784423828125, 55.766265869140625, 387.4976806640625, 448.638427734375, 38.714271545410156, 362.2524719238281, 276.0130615234375, 171.72854614257812, 39.60530090332031, 59.33039855957031, 146.78863525390625, 19.481201171875, 278.86077880859375, 29.846420288085938, 76.12165832519531, 84.13440704345703, 110.99203491210938, 244.48931884765625, 349.43463134765625, 47.378456115722656, 152.7863006591797, 32.05088806152344, 395.318115234375, 460.4112243652344, 324.51739501953125, 108.27064514160156, -160.58013916015625, 6.056865692138672, 212.9516143798828, -17.518768310546875, 197.96218872070312, 71.7433853149414, 25.42015838623047, 154.88864135742188, 400.22576904296875, 647.15771484375, 103.07408905029297, 44.40338897705078, -115.3748779296875, 71.02528381347656, 701.7332763671875, 217.9937744140625, -44.751991271972656, 216.26693725585938, 317.36431884765625, 220.1903533935547, -328.31427001953125, 449.9442138671875, 18.303543090820312, 160.57530212402344, 239.10189819335938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000247.npy"}
{"epoch": 0.36270190895741555, "step": 248, "batch_size": 64, "mean": 163.85696411132812, "std": 154.7410430908203, "min": -175.6778564453125, "p10": -3.825963592529292, "median": 140.08058166503906, "p90": 364.9290985107423, "max": 507.7076416015625, "pos_frac": 0.890625, "sample": [314.80877685546875, 373.6405334472656, 116.2874984741211, 182.3489990234375, 205.60910034179688, 0.9117202758789062, 125.0452880859375, 397.90802001953125, 115.64002990722656, -175.6778564453125, 23.635162353515625, 271.29193115234375, 198.12506103515625, 39.865997314453125, 240.13031005859375, 61.006103515625, 24.45829963684082, 39.23683166503906, 176.1852264404297, -26.172473907470703, 145.62515258789062, 282.108154296875, 262.58111572265625, 445.4773254394531, 34.56097412109375, -27.768585205078125, 140.46072387695312, 324.89776611328125, 121.05561828613281, 245.96176147460938, 201.7213897705078, -163.98643493652344, 139.700439453125, 316.77581787109375, 104.15924072265625, 197.75881958007812, 85.08317565917969, 3.5043182373046875, -5.8563995361328125, 62.8037223815918, 230.0004119873047, 312.741943359375, 197.99288940429688, 456.240966796875, 108.26005554199219, 108.12310791015625, 469.87017822265625, 344.6024169921875, 45.516754150390625, 507.7076416015625, 3.8681411743164062, 219.92861938476562, 341.70819091796875, 277.0823974609375, 322.42950439453125, 419.5732727050781, -9.024948120117188, 47.09471130371094, 137.82408142089844, 296.8492431640625, 112.71143341064453, 2.350269317626953, 13.88400650024414, -101.39846801757812], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000248.npy"}
{"epoch": 0.3641703377386197, "step": 249, "batch_size": 64, "mean": 88.60502624511719, "std": 199.90875244140625, "min": -370.2371826171875, "p10": -160.9801986694336, "median": 72.29601287841797, "p90": 337.7037597656251, "max": 702.2343139648438, "pos_frac": 0.671875, "sample": [-3.9134998321533203, 314.18463134765625, -161.74607849121094, -81.76512145996094, 126.3686752319336, -135.08160400390625, 35.76280212402344, 132.8893280029297, 71.01959228515625, 409.65545654296875, 38.935577392578125, 43.612457275390625, 59.81273651123047, 30.940086364746094, 235.33763122558594, -370.2371826171875, -117.80867004394531, 406.3104248046875, 138.49514770507812, -80.41645812988281, 138.17471313476562, 220.01441955566406, -168.8167724609375, 124.4399642944336, -40.66474914550781, 367.1239318847656, -164.43470764160156, 281.49383544921875, 96.1009292602539, 73.57243347167969, -31.223552703857422, -73.65737915039062, 279.0753173828125, 70.40464782714844, 132.8883056640625, -27.429977416992188, 190.22381591796875, -350.72320556640625, 125.94290161132812, 279.43487548828125, 236.65811157226562, -22.656234741210938, 702.2343139648438, -92.48028564453125, -221.2757568359375, -159.19314575195312, 62.015647888183594, 20.325847625732422, 107.49200439453125, 228.99501037597656, 114.01956939697266, 295.7731628417969, -170.43560791015625, 256.31170654296875, 347.78338623046875, 279.33392333984375, 65.21308898925781, -126.76107788085938, -93.6373519897461, 263.59503173828125, 95.0275650024414, 410.0823974609375, 443.76483154296875, 14.239776611328125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000249.npy"}
{"epoch": 0.3656387665198238, "step": 250, "batch_size": 64, "mean": 161.8289031982422, "std": 172.019775390625, "min": -208.91404724121094, "p10": -33.36103439331055, "median": 167.66445922851562, "p90": 396.18367614746103, "max": 518.7883911132812, "pos_frac": 0.796875, "sample": [245.47195434570312, -19.534547805786133, 38.22770690917969, -5.753887176513672, -67.36874389648438, 190.41407775878906, 81.4653091430664, 508.93475341796875, 188.05575561523438, -208.91404724121094, 78.3548812866211, 509.1045227050781, 235.86767578125, 139.0200958251953, 249.2279052734375, 185.7854766845703, 57.32378387451172, 151.677001953125, 348.55841064453125, 375.9947814941406, 198.49313354492188, -82.54512786865234, -30.81775665283203, 116.2540283203125, 100.09126281738281, 241.14779663085938, 82.96990203857422, 16.282114028930664, 89.01876068115234, -31.2208251953125, -34.27826690673828, 14.987184524536133, 118.67599487304688, 293.50189208984375, -56.910850524902344, 518.7883911132812, 157.60687255859375, 203.64755249023438, -149.50384521484375, -21.545982360839844, 218.83665466308594, 404.8360595703125, 355.86578369140625, 202.7220916748047, 177.7220458984375, 45.42987823486328, 282.3359069824219, 141.85733032226562, 275.50592041015625, 150.3810577392578, 61.59115219116211, -13.897689819335938, 180.9093475341797, 278.3711242675781, -172.9765625, 260.909912109375, 411.0986328125, 213.16883850097656, 182.14602661132812, 489.1265563964844, 332.33831787109375, 16.88393211364746, 512.46484375, 322.8675537109375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000250.npy"}
{"epoch": 0.3671071953010279, "step": 251, "batch_size": 64, "mean": 110.7523422241211, "std": 195.4483642578125, "min": -406.5888671875, "p10": -67.83662185668943, "median": 85.24739456176758, "p90": 334.97482299804693, "max": 697.9622802734375, "pos_frac": 0.703125, "sample": [428.7960510253906, 12.451793670654297, 382.33453369140625, -18.355125427246094, 237.3165283203125, -308.0223388671875, -406.5888671875, 697.9622802734375, 2.3205642700195312, 206.63633728027344, 172.16934204101562, -154.65579223632812, 312.447265625, 49.542457580566406, 100.04376220703125, -33.01104736328125, -313.43841552734375, 446.3466796875, -10.254934310913086, 195.38572692871094, 55.86845779418945, 312.17236328125, -11.289093017578125, -50.919273376464844, 37.666473388671875, 165.99452209472656, -130.10459899902344, -75.0869140625, 95.8795166015625, 290.29052734375, 61.1492919921875, -141.29458618164062, 316.2242736816406, -39.713279724121094, 72.6078109741211, 233.0211181640625, 63.05687713623047, 139.85940551757812, 138.9631805419922, 179.38442993164062, -5.0326690673828125, 216.5428924560547, -27.868432998657227, 307.63287353515625, -3.8487548828125, 374.0160827636719, 227.3368682861328, 192.3854522705078, 81.24076080322266, 119.59705352783203, -37.715721130371094, 21.94109344482422, 600.267822265625, 80.92526245117188, 343.0107727050781, 207.38031005859375, -30.30989646911621, 252.96759033203125, 89.2540283203125, -23.556507110595703, 34.24962615966797, 172.92385864257812, 59.215370178222656, 122.43629455566406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000251.npy"}
{"epoch": 0.368575624082232, "step": 252, "batch_size": 64, "mean": 146.83950805664062, "std": 152.2213592529297, "min": -503.88739013671875, "p10": -19.838100242614743, "median": 156.6497573852539, "p90": 313.91908264160156, "max": 484.18115234375, "pos_frac": 0.84375, "sample": [251.32510375976562, 373.0677490234375, 20.921222686767578, 191.83908081054688, 269.97430419921875, 75.57351684570312, 148.73464965820312, 363.2583923339844, 93.00079345703125, 139.03518676757812, 217.73692321777344, 35.272247314453125, 181.41781616210938, 54.78471374511719, 127.21305847167969, 172.9813690185547, 62.286563873291016, 134.72950744628906, 180.43255615234375, -65.98900604248047, 284.0556640625, 164.5648651123047, -31.495014190673828, 431.7039489746094, 264.9219055175781, -110.46318054199219, -20.74309730529785, 118.929931640625, -17.7264404296875, 142.7941436767578, -54.548404693603516, 199.55706787109375, 26.14421272277832, 174.30490112304688, 88.41888427734375, 136.95504760742188, 112.013916015625, 400.1111755371094, 384.5814208984375, 311.7724914550781, 55.277305603027344, 184.6482696533203, 195.5063934326172, 56.24263000488281, 179.82965087890625, 314.83905029296875, 281.3081359863281, 289.7442626953125, 188.38287353515625, 186.9383087158203, -11.633529663085938, -503.88739013671875, 298.92327880859375, 272.165283203125, 126.65496826171875, 107.28907775878906, -8.278043746948242, 193.06198120117188, 484.18115234375, -53.848506927490234, 113.40748596191406, 60.30192565917969, 177.71603393554688, 175.509033203125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000252.npy"}
{"epoch": 0.3700440528634361, "step": 253, "batch_size": 64, "mean": 144.8230743408203, "std": 180.3534393310547, "min": -410.2072448730469, "p10": -65.8384651184082, "median": 143.4568328857422, "p90": 374.55, "max": 572.645263671875, "pos_frac": 0.796875, "sample": [-57.20441436767578, 13.821372985839844, 275.8238830566406, 30.65069580078125, -13.696563720703125, 124.58243560791016, 235.22915649414062, 64.76644134521484, 245.84754943847656, 62.318572998046875, 111.86508178710938, 572.645263671875, 202.39468383789062, 74.23104095458984, 234.827880859375, 129.0040283203125, 382.3465881347656, 98.46403503417969, -89.08172607421875, 104.32859802246094, -10.421552658081055, -410.2072448730469, 234.82672119140625, 375.8465881347656, 357.1052551269531, 9.907129287719727, 444.9071044921875, 22.212562561035156, -21.01519775390625, 38.87424087524414, 18.637901306152344, -57.063629150390625, 163.28067016601562, 288.2505187988281, 307.1676330566406, 211.79443359375, 291.4097595214844, -134.16061401367188, 290.26776123046875, 101.35755157470703, 263.99163818359375, 338.29638671875, 177.4656982421875, 302.13079833984375, 430.798095703125, 151.09461975097656, 135.8190460205078, 371.5246276855469, 163.77764892578125, -268.9885559082031, 410.57891845703125, 186.20541381835938, -90.61224365234375, 78.75115966796875, -100.67474365234375, 109.9046859741211, 156.62496948242188, 243.7866668701172, 270.3928527832031, -32.572357177734375, -69.53877258300781, 65.25936889648438, 464.6976013183594, 183.82135009765625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000253.npy"}
{"epoch": 0.37151248164464024, "step": 254, "batch_size": 64, "mean": 125.81629180908203, "std": 211.8404541015625, "min": -244.54856872558594, "p10": -117.23667602539062, "median": 101.87121963500977, "p90": 425.7121307373049, "max": 878.2420654296875, "pos_frac": 0.734375, "sample": [87.8125, -43.96178436279297, 159.78488159179688, 102.21503448486328, 101.52740478515625, 144.42628479003906, -0.5917510986328125, -107.7953872680664, 149.8035888671875, 17.151390075683594, 176.9710693359375, 236.5841522216797, 246.53567504882812, 50.51612854003906, -159.05459594726562, -244.54856872558594, 221.8648681640625, -42.65351486206055, 130.21749877929688, 209.47860717773438, 6.954654693603516, 198.30072021484375, 546.1641845703125, 313.6116027832031, 373.8392639160156, -190.8770751953125, 460.651611328125, 297.8685302734375, -11.676658630371094, -5.832693099975586, 112.0041732788086, 203.29953002929688, 55.354515075683594, -187.42227172851562, 14.505062103271484, 220.52786254882812, -51.429527282714844, 447.943359375, -116.53317260742188, 25.401233673095703, -27.75939178466797, -195.99627685546875, 65.27230834960938, 455.39678955078125, 35.928443908691406, 228.24209594726562, 6.130241394042969, 355.04541015625, 65.7529067993164, 196.0211181640625, 878.2420654296875, 124.13236999511719, 85.79367065429688, 1.3576812744140625, -167.71630859375, 595.14892578125, 106.72479248046875, 234.62445068359375, 155.2252197265625, 66.11732482910156, -117.53817749023438, 608.4449462890625, 161.671142578125, -12.95751953125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000254.npy"}
{"epoch": 0.37298091042584436, "step": 255, "batch_size": 64, "mean": 132.2594451904297, "std": 191.86410522460938, "min": -434.691650390625, "p10": -69.57223968505858, "median": 116.79199981689453, "p90": 351.9410644531251, "max": 587.6817626953125, "pos_frac": 0.75, "sample": [587.6817626953125, -108.83648681640625, 33.959686279296875, 358.12274169921875, 206.87384033203125, -49.19781494140625, -47.517066955566406, -21.155654907226562, -434.691650390625, 52.4398307800293, 329.854248046875, 124.02528381347656, -9.049041748046875, 19.489131927490234, 272.26666259765625, 327.0669250488281, -161.55316162109375, -247.22308349609375, 127.26611328125, 319.1078796386719, 268.31964111328125, 19.567476272583008, 40.32402801513672, -4.123394012451172, 73.4736328125, -57.88701629638672, -169.46792602539062, 384.24359130859375, -28.261394500732422, 230.68609619140625, 225.8750762939453, 235.7916259765625, 110.82792663574219, 320.81781005859375, 381.18109130859375, 308.46539306640625, 272.83843994140625, 393.27838134765625, 337.51715087890625, 569.4240112304688, 39.48273468017578, -30.474964141845703, 280.1414489746094, 128.77435302734375, 22.407787322998047, 237.5471649169922, 103.14117431640625, 252.8536376953125, -140.43544006347656, 328.9948425292969, 409.662841796875, 64.63041687011719, 186.07955932617188, 51.562782287597656, -74.58019256591797, 122.75607299804688, 24.172218322753906, 70.39208221435547, 42.225669860839844, 148.19711303710938, -43.46460723876953, 58.874576568603516, 291.9138488769531, 297.9273681640625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000255.npy"}
{"epoch": 0.3744493392070485, "step": 256, "batch_size": 64, "mean": 144.20822143554688, "std": 188.68353271484375, "min": -305.77105712890625, "p10": -57.3468376159668, "median": 136.20104598999023, "p90": 370.94339599609384, "max": 644.5574951171875, "pos_frac": 0.75, "sample": [122.98146057128906, 587.8319091796875, 147.20587158203125, 221.30604553222656, 210.462890625, 269.073486328125, 31.403528213500977, 185.85043334960938, -26.2589111328125, 26.411331176757812, 318.7672119140625, 253.32806396484375, 175.05349731445312, 350.43353271484375, 283.4958190917969, -131.02687072753906, 25.47744369506836, 226.39500427246094, -305.77105712890625, 48.71073913574219, 272.470458984375, 117.40211486816406, -204.441162109375, 334.4549865722656, 56.57083511352539, 303.7529602050781, -57.24787139892578, -68.98028564453125, -86.88899230957031, 168.5741424560547, 379.73333740234375, -13.696701049804688, 242.68685913085938, 91.80245971679688, 68.47837829589844, 125.19622039794922, 428.17767333984375, -56.44740295410156, 65.62037658691406, 56.73991394042969, -167.1168670654297, 16.173492431640625, 36.15948486328125, -18.958984375, 167.1375732421875, -28.92081069946289, 227.71697998046875, 154.05462646484375, -57.389251708984375, 255.2178955078125, -40.26719665527344, 111.90852355957031, -8.324317932128906, 192.02532958984375, 322.06396484375, 468.6025390625, 192.75498962402344, 53.53572463989258, 644.5574951171875, 553.1439208984375, 260.8407897949219, 404.7477111816406, -39.64281463623047, 284.2156677246094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000256.npy"}
{"epoch": 0.37591776798825255, "step": 257, "batch_size": 64, "mean": 151.2591552734375, "std": 197.95106506347656, "min": -329.2390441894531, "p10": -55.00740127563476, "median": 124.54714584350586, "p90": 437.4923980712891, "max": 706.81103515625, "pos_frac": 0.78125, "sample": [446.4669189453125, -43.850337982177734, 42.89674377441406, 97.59443664550781, 378.8502502441406, 13.714645385742188, 262.1155700683594, 439.1888427734375, 287.3387145996094, 183.53692626953125, -57.721275329589844, -53.487937927246094, 131.65553283691406, 323.6553955078125, 129.078125, 51.230560302734375, 229.57192993164062, 267.89581298828125, 10.579578399658203, 433.5340270996094, 45.96458435058594, 428.0568542480469, 302.4448547363281, -199.85565185546875, 98.20455932617188, 125.25584411621094, -105.43882751464844, 12.368553161621094, -28.054058074951172, 468.7586364746094, 147.7428741455078, 255.6134796142578, 289.2987365722656, 706.81103515625, 90.9205551147461, -55.658599853515625, 115.696533203125, -3.055316925048828, 137.78005981445312, -329.2390441894531, 123.83844757080078, 478.6653137207031, 51.356204986572266, 249.49972534179688, 166.8091583251953, 219.869140625, 70.90089416503906, -17.99396514892578, 557.8638305664062, 262.3188171386719, 241.97976684570312, 381.4720458984375, 49.71050262451172, 116.88658905029297, 60.811676025390625, 497.4299621582031, -84.48965454101562, -10.316164016723633, 169.7039794921875, -232.10549926757812, 31.560195922851562, 32.74266052246094, -30.357620239257812, 214.969482421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000257.npy"}
{"epoch": 0.37738619676945667, "step": 258, "batch_size": 64, "mean": 150.2278289794922, "std": 214.11106872558594, "min": -523.324951171875, "p10": -84.52714614868162, "median": 148.6353302001953, "p90": 393.50095214843753, "max": 765.7698974609375, "pos_frac": 0.78125, "sample": [325.81463623046875, 33.21516418457031, -111.28620910644531, 297.13336181640625, 89.85028076171875, 142.12832641601562, 266.3910217285156, 255.04922485351562, 254.32583618164062, -21.702484130859375, -38.827911376953125, -8.253700256347656, -65.22046661376953, 89.30653381347656, 120.9898910522461, -52.27961730957031, -3.8653640747070312, 118.55392456054688, 203.1815643310547, -12.019844055175781, 58.51728820800781, 257.1224060058594, 552.9796142578125, 236.29945373535156, 476.5665588378906, 278.99066162109375, 145.948974609375, 17.023094177246094, 162.59768676757812, 378.4251708984375, 187.7820587158203, 301.9827880859375, 329.20184326171875, 177.71278381347656, -244.20941162109375, 109.31953430175781, 308.8600158691406, -403.38995361328125, 0.44306182861328125, 139.59754943847656, 422.335693359375, 24.720890045166016, -110.45811462402344, 406.1888122558594, 33.81943893432617, -523.324951171875, 765.7698974609375, 170.022216796875, -168.0007781982422, 95.53145599365234, 223.56625366210938, -92.80143737792969, 305.22442626953125, 398.36602783203125, 382.14910888671875, 30.522546768188477, 288.66180419921875, 472.04833984375, 160.1608123779297, 291.51953125, 81.83094787597656, 151.32168579101562, 330.33453369140625, 120.8169937133789], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000258.npy"}
{"epoch": 0.3788546255506608, "step": 259, "batch_size": 64, "mean": 125.29249572753906, "std": 189.1765594482422, "min": -241.7370147705078, "p10": -106.75092468261718, "median": 98.13822937011719, "p90": 372.02821350097656, "max": 549.8701171875, "pos_frac": 0.75, "sample": [205.39503479003906, -13.657052993774414, 143.2181396484375, 196.60179138183594, 30.142513275146484, -86.61349487304688, 335.523681640625, 473.2808532714844, 457.1466369628906, 262.96124267578125, 48.557044982910156, -122.73375701904297, 113.20930480957031, -80.5657730102539, 10.4918212890625, -111.81959533691406, 346.678466796875, -52.78562927246094, 469.47991943359375, 207.61099243164062, 183.81182861328125, 127.82720947265625, 85.7242431640625, 68.87562561035156, 316.92425537109375, 40.36407470703125, 66.61927795410156, 44.71791076660156, 549.8701171875, -241.7370147705078, 71.84353637695312, 33.84490203857422, 522.3052368164062, 205.19345092773438, 74.1817855834961, -144.0356903076172, 158.96368408203125, 316.88189697265625, -162.78201293945312, 343.15472412109375, -174.63040161132812, 20.906959533691406, 365.3197021484375, -94.92402648925781, 110.55221557617188, 42.72869873046875, 374.9032897949219, 297.912841796875, -206.56968688964844, 163.03074645996094, -43.13822937011719, 293.2327880859375, 282.38494873046875, -66.89591979980469, 388.1512451171875, -68.28736114501953, 118.03176879882812, 310.5922546386719, 151.07931518554688, -37.94261932373047, 76.7049560546875, 30.955257415771484, 174.8680419921875, 15.081672668457031], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000259.npy"}
{"epoch": 0.3803230543318649, "step": 260, "batch_size": 64, "mean": 167.13815307617188, "std": 189.3502197265625, "min": -209.04742431640625, "p10": -24.080447959899896, "median": 130.8702163696289, "p90": 405.9152984619141, "max": 1003.4569091796875, "pos_frac": 0.875, "sample": [245.688720703125, -26.614648818969727, 42.18970489501953, -110.49346923828125, 4.764089584350586, 158.430908203125, 146.9073028564453, 200.05088806152344, 349.0518798828125, 54.64994812011719, 38.886844635009766, 481.747802734375, 115.32070922851562, 115.86067199707031, 150.9715576171875, 1003.4569091796875, 186.59088134765625, 256.03558349609375, 24.1163330078125, 475.08099365234375, -209.04742431640625, 158.67251586914062, -31.155485153198242, 150.98092651367188, -88.9204330444336, -28.48218536376953, 140.7269287109375, -18.167312622070312, 244.5888671875, 128.41795349121094, 111.87983703613281, 360.51763916015625, 5.250999450683594, 229.2290802001953, 130.3538360595703, 386.1307373046875, 394.1971130371094, 168.65631103515625, 110.90748596191406, 105.6844482421875, 108.87843322753906, 60.031341552734375, 344.3634033203125, 119.98067474365234, 162.7998046875, 24.280193328857422, 37.89207458496094, 79.47869873046875, 182.41458129882812, 215.8113250732422, 209.53936767578125, 206.2054443359375, 117.14960479736328, 439.08673095703125, 16.150779724121094, -37.586036682128906, 56.5062255859375, 410.9373779296875, 24.33527374267578, 612.7806396484375, 131.3865966796875, 415.3709411621094, 58.648406982421875, 337.284912109375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000260.npy"}
{"epoch": 0.38179148311306904, "step": 261, "batch_size": 64, "mean": 155.6842041015625, "std": 205.00607299804688, "min": -319.1692810058594, "p10": -59.038606643676744, "median": 129.26173400878906, "p90": 405.9144165039063, "max": 880.7112426757812, "pos_frac": 0.78125, "sample": [416.7480773925781, 408.55999755859375, 47.27112579345703, -217.9115447998047, 114.57801818847656, 880.7112426757812, 47.03736877441406, 356.51171875, -21.424793243408203, 226.72299194335938, 48.79667663574219, 73.86261749267578, -97.42840576171875, 399.74139404296875, 133.27171325683594, 223.00167846679688, 242.96743774414062, 358.8330078125, 395.11529541015625, -15.051078796386719, 125.25175476074219, 82.16964721679688, 152.2229766845703, 181.08981323242188, 183.70957946777344, 299.36517333984375, 509.231689453125, 17.006195068359375, 320.5686340332031, 212.22462463378906, 49.048095703125, 30.60346221923828, 4.977333068847656, 90.22758483886719, 244.18075561523438, 57.27771759033203, 291.23388671875, 364.08148193359375, 258.3778076171875, 108.04389190673828, 249.23663330078125, -63.53713607788086, 481.1345520019531, 413.353271484375, -3.6540775299072266, 382.7010498046875, -199.980224609375, 236.88690185546875, -32.58684539794922, 247.01419067382812, 115.1430892944336, 497.3114013671875, -9.611457824707031, -319.1692810058594, -183.7517547607422, 117.72834014892578, -154.60128784179688, -16.790205001831055, 156.60733032226562, 212.7572784423828, 52.97418975830078, 73.12664031982422, 157.23092651367188, -48.54203796386719], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000261.npy"}
{"epoch": 0.3832599118942731, "step": 262, "batch_size": 64, "mean": 129.4208984375, "std": 167.50112915039062, "min": -249.91700744628906, "p10": -73.16872253417966, "median": 130.11138916015625, "p90": 292.5226928710938, "max": 735.4868774414062, "pos_frac": 0.796875, "sample": [37.443328857421875, 401.1820983886719, 95.03839874267578, 211.74462890625, -249.91700744628906, 323.8866882324219, 474.5330810546875, 113.20799255371094, 162.81765747070312, 244.68307495117188, -158.9423828125, -237.26077270507812, 140.83856201171875, 735.4868774414062, 129.11666870117188, -178.08560180664062, 182.55850219726562, 279.4002685546875, -25.911596298217773, 229.83114624023438, 160.35919189453125, 70.11846923828125, -96.53691101074219, 280.0035705566406, 157.94813537597656, -17.504959106445312, 52.75454330444336, 417.42041015625, 153.07901000976562, 331.1758117675781, 59.46546936035156, 86.89717864990234, 82.01071166992188, -46.49225997924805, 98.19291687011719, 72.05828094482422, 154.9703369140625, -34.44648742675781, 204.97149658203125, 160.4117889404297, 131.10610961914062, 153.83644104003906, 40.263511657714844, 264.12738037109375, 217.03465270996094, 104.07329559326172, 259.451171875, 121.76366424560547, -82.49193572998047, -51.41455841064453, 96.18754577636719, 81.58671569824219, 61.54704284667969, 297.8880310058594, -143.05133056640625, -34.441184997558594, 117.92428588867188, 155.06591796875, 260.5444641113281, 265.6076354980469, 186.26983642578125, 37.22941589355469, 264.0460205078125, 220.27444458007812], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000262.npy"}
{"epoch": 0.38472834067547723, "step": 263, "batch_size": 64, "mean": 184.63003540039062, "std": 211.26284790039062, "min": -332.68701171875, "p10": -48.45418167114257, "median": 158.38462829589844, "p90": 461.1761749267579, "max": 746.587890625, "pos_frac": 0.8125, "sample": [321.3268127441406, -332.68701171875, 155.7579803466797, 166.3199920654297, 381.14410400390625, 46.68920135498047, 306.01812744140625, 50.92877197265625, 24.089492797851562, 96.20943450927734, -248.2447509765625, 191.4101104736328, 81.58028411865234, 325.5684814453125, 299.1419982910156, 175.39027404785156, 19.423091888427734, 382.095947265625, 117.16605377197266, 150.22723388671875, -2.812479019165039, -66.31695556640625, 64.79792022705078, 366.5445556640625, -27.73886489868164, -91.68051147460938, 321.74658203125, 157.86273193359375, -4.908683776855469, 92.01419067382812, 28.973709106445312, 34.28562927246094, 356.5970153808594, 58.42626190185547, -8.506546020507812, 527.9666137695312, 271.7786865234375, 110.6591796875, 404.73443603515625, 469.0233154296875, 113.4795913696289, 538.8191528320312, 97.93598937988281, 286.8605651855469, 319.08367919921875, 171.41041564941406, 483.71240234375, -52.11341857910156, 239.8646240234375, 301.3168029785156, 396.3373718261719, 442.8661804199219, 686.8690795898438, -159.2811279296875, 0.3313751220703125, 548.10302734375, -39.91596221923828, -76.04456329345703, 258.38104248046875, 239.76715087890625, 157.24131774902344, 746.587890625, 158.90652465820312, 182.80136108398438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000263.npy"}
{"epoch": 0.38619676945668135, "step": 264, "batch_size": 64, "mean": 153.5868377685547, "std": 203.8961181640625, "min": -240.804443359375, "p10": -59.44566116333007, "median": 122.91144561767578, "p90": 451.2365905761719, "max": 666.877197265625, "pos_frac": 0.78125, "sample": [225.47607421875, 332.3329162597656, 49.4660530090332, 666.877197265625, 76.85043334960938, -218.28208923339844, -144.5726318359375, 386.6570129394531, 64.07470703125, 509.21435546875, -43.72063446044922, 42.75495147705078, 148.47674560546875, -150.40771484375, 167.95162963867188, 169.20223999023438, 271.6653747558594, 419.7105712890625, 434.13311767578125, 417.50079345703125, 77.64501953125, 228.76202392578125, 263.5162353515625, 198.64837646484375, 75.31097412109375, 22.475814819335938, -53.45513153076172, -5.2244110107421875, -14.717277526855469, 201.69056701660156, -62.013031005859375, 16.90906524658203, 73.84664154052734, 570.2484741210938, 2.999603271484375, 474.0444641113281, 116.4012222290039, 192.92391967773438, 44.85548400878906, 264.85614013671875, 51.07887268066406, 201.25697326660156, 121.32919311523438, 124.49369812011719, -22.451011657714844, 598.8013916015625, 347.97515869140625, -240.804443359375, 239.05337524414062, -136.78790283203125, 179.74891662597656, 27.630451202392578, 235.12649536132812, 458.566650390625, -41.44189453125, 153.3296661376953, 33.622657775878906, 79.34799194335938, -199.3275604248047, 482.4938049316406, 313.0662841796875, 252.0373077392578, -8.435409545898438, 64.76113891601562], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000264.npy"}
{"epoch": 0.3876651982378855, "step": 265, "batch_size": 64, "mean": 151.9520263671875, "std": 242.4590606689453, "min": -503.61468505859375, "p10": -88.56967163085936, "median": 126.22060775756836, "p90": 454.9584594726564, "max": 841.07080078125, "pos_frac": 0.75, "sample": [138.08547973632812, 409.5807800292969, 353.9091491699219, 230.72398376464844, 82.60453033447266, 170.8604736328125, 352.3214111328125, 34.938941955566406, 690.1947021484375, 92.6011962890625, 233.91268920898438, -278.8737487792969, 16.073123931884766, -18.819034576416016, -107.12301635742188, 43.07901382446289, 303.3143615722656, -91.946044921875, 111.54532623291016, -69.28256225585938, 251.76922607421875, 841.07080078125, -503.61468505859375, 539.05126953125, -80.69146728515625, 211.92959594726562, 180.99847412109375, 468.7804870605469, -71.43051147460938, 750.1122436523438, -240.02452087402344, 351.4099426269531, 70.32916259765625, -64.22898864746094, 80.97207641601562, -98.39903259277344, 8.908737182617188, -56.026397705078125, 379.98590087890625, -39.08183288574219, 472.5395812988281, 108.65069580078125, 74.15998840332031, 401.62603759765625, -35.47785186767578, 192.02064514160156, 20.203590393066406, 156.54388427734375, -22.511978149414062, 286.94219970703125, 114.3557357788086, 111.65264892578125, 153.25067138671875, 422.7070617675781, 177.4291229248047, 152.95166015625, 224.9441375732422, 171.93405151367188, 47.32357406616211, -318.2179260253906, 488.1240234375, 87.9476318359375, 338.47900390625, 217.8306121826172], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000265.npy"}
{"epoch": 0.3891336270190896, "step": 266, "batch_size": 64, "mean": 119.895263671875, "std": 195.84727478027344, "min": -265.3891906738281, "p10": -104.04676437377928, "median": 78.21189498901367, "p90": 401.7353973388672, "max": 658.9937744140625, "pos_frac": 0.71875, "sample": [-27.24102783203125, 564.3976440429688, 76.30052947998047, 116.60265350341797, 239.44761657714844, 142.90077209472656, -88.90925598144531, 287.2501220703125, 44.19818115234375, 413.8106689453125, 135.19569396972656, 337.0391845703125, -36.72117614746094, 312.1675109863281, -109.01499938964844, 62.48773193359375, 270.5372619628906, 80.12326049804688, 213.789794921875, 61.118125915527344, 173.7179412841797, -244.61607360839844, -39.553688049316406, 374.44390869140625, 182.46595764160156, 399.0066833496094, 242.0929412841797, 2.0423812866210938, 32.074684143066406, 402.90484619140625, 476.93377685546875, 296.51007080078125, 658.9937744140625, 432.46282958984375, 219.28822326660156, 55.2979736328125, 0.8638725280761719, 219.6187744140625, 29.889381408691406, 95.1332015991211, 45.48944091796875, 153.482666015625, 63.797401428222656, 39.114662170410156, 141.64505004882812, -18.101757049560547, -265.3891906738281, -223.08251953125, 260.2247314453125, 37.867393493652344, -161.4075927734375, -92.45421600341797, -14.947402954101562, -120.5518798828125, -56.289024353027344, -30.9400634765625, -19.693918228149414, -30.791099548339844, -142.59359741210938, 444.9905090332031, 225.89410400390625, 233.37368774414062, 14.283706665039062, 84.3243408203125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000266.npy"}
{"epoch": 0.39060205580029367, "step": 267, "batch_size": 64, "mean": 145.87728881835938, "std": 211.58094787597656, "min": -527.7257080078125, "p10": -65.69191436767578, "median": 116.78258514404297, "p90": 343.37127685546875, "max": 748.6859741210938, "pos_frac": 0.796875, "sample": [77.57463836669922, 187.6361083984375, 343.812744140625, 8.641490936279297, 113.42684936523438, -21.49250602722168, 262.7776794433594, 190.8550567626953, 24.21051025390625, 51.12019729614258, -63.8609619140625, -74.41647338867188, -24.549423217773438, 334.76165771484375, 163.56878662109375, 108.90180206298828, -73.95391082763672, 210.14907836914062, 492.2315673828125, 748.6859741210938, 101.069580078125, 65.55936431884766, 147.21572875976562, -3.59661865234375, -91.99018859863281, 118.35847473144531, 71.58485412597656, 87.0393295288086, 342.3411865234375, 216.58984375, 558.8985595703125, 425.1083068847656, 703.6143798828125, 47.597625732421875, -2.0154762268066406, 305.4122009277344, 37.32712936401367, -73.6844253540039, 7.549129486083984, 215.25534057617188, 257.5311279296875, 18.850738525390625, -6.160213470458984, 597.86181640625, -527.7257080078125, 8.548896789550781, -386.7144775390625, 216.93313598632812, 199.92202758789062, 115.20669555664062, 205.18397521972656, 251.28271484375, 77.54727172851562, 280.11749267578125, 106.47394561767578, 158.0121307373047, -66.47660827636719, 329.579345703125, 99.7251968383789, 135.2764892578125, 323.50579833984375, 171.48944091796875, 295.91632080078125, 134.94314575195312], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000267.npy"}
{"epoch": 0.3920704845814978, "step": 268, "batch_size": 64, "mean": 144.32022094726562, "std": 213.49301147460938, "min": -354.69049072265625, "p10": -70.02733726501464, "median": 141.0375213623047, "p90": 389.09351196289066, "max": 854.8500366210938, "pos_frac": 0.84375, "sample": [22.14691925048828, 458.0223388671875, 87.53760528564453, 15.077156066894531, 25.686599731445312, 67.15113067626953, 162.26219177246094, 78.30812072753906, 390.7999267578125, 136.01759338378906, 178.99588012695312, -339.31695556640625, 171.9798583984375, 17.085556030273438, 289.9302978515625, -165.02914428710938, 120.7517318725586, 59.51582336425781, -306.7873840332031, 142.5433807373047, 157.77090454101562, 854.8500366210938, 62.80588912963867, 167.62844848632812, 51.60236740112305, 264.45233154296875, 458.12774658203125, 556.1342163085938, -29.781532287597656, 325.74005126953125, -101.3626937866211, 70.01429748535156, -73.15603637695312, -29.666568756103516, 566.4907836914062, 151.92459106445312, 168.7229766845703, 155.8584747314453, 302.40618896484375, 223.03733825683594, 190.22174072265625, 262.13116455078125, 9.494148254394531, -92.43904113769531, 365.08795166015625, 4.63360595703125, 212.98985290527344, 385.11187744140625, 7.391204833984375, 269.5014953613281, 173.0817413330078, 110.65141296386719, 674.3155517578125, 190.95481872558594, 87.78860473632812, 103.56562805175781, 139.5316619873047, 16.568519592285156, 293.7987060546875, 0.4107379913330078, 175.14324951171875, -354.69049072265625, -62.7270393371582, 157.69898986816406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000268.npy"}
{"epoch": 0.3935389133627019, "step": 269, "batch_size": 64, "mean": 174.90939331054688, "std": 218.8176727294922, "min": -416.6075744628906, "p10": -79.87607955932617, "median": 186.6360855102539, "p90": 463.3854156494141, "max": 620.594482421875, "pos_frac": 0.765625, "sample": [137.2760772705078, 331.5475769042969, 184.6629638671875, 320.1646728515625, -75.31151580810547, 211.03814697265625, 343.3276672363281, 206.68585205078125, -49.2373046875, -32.829490661621094, 299.1661682128906, -12.64599609375, 502.4109802246094, 47.418609619140625, 375.4678649902344, 263.3570556640625, -15.569648742675781, 188.6092071533203, -96.67570495605469, -416.6075744628906, 277.2717590332031, 44.20296096801758, 219.12014770507812, -249.143310546875, -102.12693786621094, 397.504638671875, -7.381065368652344, 570.8900146484375, 485.82501220703125, 99.1890869140625, 166.92626953125, 120.559814453125, 443.0390625, 273.8496398925781, 4.3210601806640625, 372.26214599609375, 390.0579833984375, 111.70732879638672, 313.46331787109375, 44.333740234375, 8.510444641113281, 148.43141174316406, -180.49131774902344, 39.940460205078125, 146.02854919433594, 601.3309326171875, 298.3624267578125, 466.94482421875, 318.9083251953125, -21.742462158203125, 307.3155212402344, 282.41925048828125, 16.969757080078125, 620.594482421875, -81.83232116699219, 465.7071838378906, 5.006561279296875, 43.97161102294922, -58.29597091674805, 198.6259765625, 457.96795654296875, 390.67572021484375, -189.2157745361328, 219.93881225585938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000269.npy"}
{"epoch": 0.39500734214390604, "step": 270, "batch_size": 64, "mean": 127.51922607421875, "std": 201.89947509765625, "min": -430.9273986816406, "p10": -74.92207336425781, "median": 115.40621185302734, "p90": 366.439404296875, "max": 730.8496704101562, "pos_frac": 0.78125, "sample": [242.89498901367188, 315.3934020996094, 476.2918701171875, -316.541015625, -430.9273986816406, -20.110702514648438, 52.251670837402344, 259.1951599121094, 109.95384979248047, 268.1772155761719, 346.12353515625, -45.980926513671875, 107.13436126708984, 204.88455200195312, 42.59735107421875, 48.400054931640625, 120.85857391357422, 159.87518310546875, -143.62481689453125, 202.93063354492188, 281.7731628417969, 123.64654541015625, -25.750633239746094, 245.450927734375, 459.08551025390625, 267.8416748046875, -1.6187076568603516, 97.42230987548828, 213.04501342773438, 89.21704864501953, -247.9136505126953, 358.28515625, 730.8496704101562, -76.08786010742188, -148.553466796875, 226.13320922851562, 379.6217041015625, 52.594512939453125, 73.38529968261719, 18.819271087646484, 593.6932373046875, -224.48712158203125, 188.01779174804688, 9.687297821044922, 147.058837890625, 135.9005126953125, 55.15187072753906, 216.79840087890625, -72.201904296875, -34.31736373901367, 191.03231811523438, 295.2140808105469, 107.03927612304688, 7.46729850769043, 1.0416717529296875, 181.74972534179688, 17.765213012695312, 98.36599731445312, -40.27497100830078, 369.93408203125, 151.3899688720703, 150.33135986328125, 68.56144714355469, 429.286865234375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000270.npy"}
{"epoch": 0.3964757709251101, "step": 271, "batch_size": 64, "mean": 157.23764038085938, "std": 176.83050537109375, "min": -136.71234130859375, "p10": -30.790550231933594, "median": 109.97701263427734, "p90": 411.3792297363282, "max": 654.493408203125, "pos_frac": 0.765625, "sample": [654.493408203125, 376.6571350097656, 100.58717346191406, -59.454219818115234, 205.4671173095703, -10.977497100830078, 368.2872009277344, 328.6770935058594, 5.912147521972656, -31.110389709472656, 351.87054443359375, -136.71234130859375, 420.52520751953125, -94.46149444580078, 94.02574920654297, -30.04425811767578, 85.38125610351562, -7.295890808105469, 170.5121307373047, -6.1339874267578125, -67.03189849853516, 40.54670715332031, 466.48858642578125, 372.8304443359375, 478.13983154296875, 390.07635498046875, 55.365135192871094, -37.99049377441406, 83.08966827392578, 114.47991180419922, -7.209651947021484, 249.0145721435547, 146.324951171875, -10.569923400878906, -43.831268310546875, 134.71836853027344, 61.767425537109375, 10.762702941894531, 338.060302734375, 106.83045959472656, 28.370071411132812, 532.8599853515625, 59.21642303466797, -5.62596321105957, 97.15083312988281, 28.523204803466797, 113.12356567382812, 176.4046173095703, 340.5040588378906, 163.67410278320312, 459.36895751953125, 244.79046630859375, 100.41606903076172, 24.279373168945312, 420.509033203125, -0.04131317138671875, 224.812744140625, 347.1661682128906, 18.78240966796875, 226.2016143798828, 151.29940795898438, 279.948974609375, 140.81411743164062, 222.5911407470703], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000271.npy"}
{"epoch": 0.39794419970631423, "step": 272, "batch_size": 64, "mean": 198.25775146484375, "std": 187.06304931640625, "min": -172.41297912597656, "p10": -52.81837882995605, "median": 195.50526428222656, "p90": 428.6201507568359, "max": 664.3883056640625, "pos_frac": 0.796875, "sample": [333.57489013671875, 249.02874755859375, 4.260786056518555, -101.70930480957031, -1.1513519287109375, 378.719970703125, 219.36390686035156, 457.61444091796875, 475.66326904296875, 283.83984375, 308.900390625, 385.6377258300781, -53.436397552490234, 161.3138427734375, -161.6438446044922, 95.64210510253906, 76.78511810302734, 226.5758514404297, 153.10037231445312, 172.02371215820312, -55.750823974609375, 353.34234619140625, 345.821044921875, 212.69439697265625, 290.38739013671875, 106.0521011352539, -13.09952163696289, 265.0136413574219, 144.45663452148438, 633.2036743164062, 60.36613464355469, 193.29733276367188, 224.07298278808594, 171.35549926757812, 63.07411193847656, 375.4798278808594, 426.2228088378906, -172.41297912597656, 197.71319580078125, 224.977294921875, 664.3883056640625, -8.346832275390625, 489.32659912109375, 154.11392211914062, 291.23583984375, -66.9654769897461, 301.85064697265625, 93.78738403320312, 250.34178161621094, 349.32562255859375, -27.819046020507812, 174.04977416992188, -13.87396240234375, -110.81333923339844, -51.37633514404297, 309.36248779296875, 246.53857421875, 76.54254913330078, 429.6475830078125, 164.67356872558594, 183.22198486328125, 424.93804931640625, 515.5381469726562, 138.43612670898438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000272.npy"}
{"epoch": 0.39941262848751835, "step": 273, "batch_size": 64, "mean": 195.46688842773438, "std": 205.177490234375, "min": -270.54327392578125, "p10": -65.70724792480469, "median": 185.55870819091797, "p90": 445.9255004882813, "max": 697.8560791015625, "pos_frac": 0.8125, "sample": [244.66650390625, 304.44378662109375, 88.98922729492188, 150.7701416015625, 608.7577514648438, 446.256591796875, 169.12747192382812, -47.657875061035156, 633.2512817382812, -67.19367980957031, 445.1529541015625, 204.14291381835938, -11.3594970703125, 58.33507537841797, 5.909730911254883, 169.67752075195312, 1.2419853210449219, 242.94085693359375, 202.62632751464844, 375.7244873046875, 246.8641815185547, 52.97407531738281, 62.765625, 165.06703186035156, 106.17793273925781, 92.19583892822266, 337.991943359375, 200.8426971435547, 217.76803588867188, 181.94378662109375, 351.740478515625, -10.858299255371094, 340.0394592285156, 362.8276062011719, -15.219476699829102, 612.9793701171875, 560.4656372070312, 239.32408142089844, 189.1736297607422, 295.11083984375, -200.26889038085938, -104.00274658203125, 243.22642517089844, -73.41754150390625, 258.6148986816406, 324.4107666015625, 367.37567138671875, 38.504371643066406, 175.4766387939453, -270.54327392578125, 697.8560791015625, 142.748291015625, 130.53196716308594, -94.14447021484375, -77.98129272460938, 312.1150817871094, 52.86247253417969, 101.67312622070312, 309.2315979003906, -62.23890686035156, 378.61785888671875, 147.4046630859375, 417.97271728515625, 477.8779602050781], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000273.npy"}
{"epoch": 0.4008810572687225, "step": 274, "batch_size": 64, "mean": 146.9497528076172, "std": 195.35658264160156, "min": -179.39756774902344, "p10": -107.30868530273438, "median": 148.43601989746094, "p90": 447.46154785156256, "max": 646.9305419921875, "pos_frac": 0.734375, "sample": [106.39771270751953, 108.43344116210938, 37.74311065673828, -147.71482849121094, 213.37374877929688, 175.3507080078125, 226.6079559326172, -32.59741973876953, 76.0964126586914, -27.68609619140625, -101.40499877929688, 430.6470642089844, 372.34552001953125, 248.97171020507812, 264.5513610839844, 154.6108856201172, 40.23638153076172, 491.49847412109375, -179.39756774902344, 269.763427734375, -142.3204803466797, 151.4856414794922, -92.36996459960938, 79.25518798828125, -40.89691925048828, 301.098388671875, 497.3074645996094, -109.83883666992188, 106.15441131591797, 646.9305419921875, -41.6816291809082, 90.26431274414062, 308.55078125, 157.550537109375, -4.355047225952148, 272.11932373046875, 454.6677551269531, -47.931793212890625, 176.3096466064453, 79.73487854003906, 240.99960327148438, -45.49967956542969, 162.91468811035156, 284.34246826171875, 333.474609375, -62.619102478027344, 110.92631530761719, 27.713987350463867, 178.11538696289062, 531.771484375, 321.5708923339844, 145.3863983154297, 11.202362060546875, 171.519775390625, -156.6600341796875, 293.994873046875, -127.2646484375, -135.57923889160156, 532.7171630859375, 219.78515625, 127.8505859375, 463.97039794921875, 1.794464111328125, 202.4947509765625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000274.npy"}
{"epoch": 0.4023494860499266, "step": 275, "batch_size": 64, "mean": 158.92544555664062, "std": 200.8318634033203, "min": -226.32015991210938, "p10": -86.39121093749999, "median": 152.01934051513672, "p90": 434.7708038330079, "max": 592.4501953125, "pos_frac": 0.78125, "sample": [208.00172424316406, 64.18232727050781, 304.7650451660156, 11.876110076904297, 8.825490951538086, -123.54251861572266, 414.9560852050781, 129.76158142089844, 510.0124206542969, -160.39810180664062, 265.522705078125, 457.3617858886719, 551.3499755859375, 80.48771667480469, 211.40655517578125, 306.389404296875, -27.49478530883789, -6.069183349609375, 139.4237060546875, 196.6206512451172, -131.43927001953125, 225.35104370117188, 472.2493591308594, -17.45690155029297, 19.457801818847656, -195.79989624023438, 260.777587890625, 174.57559204101562, 188.18377685546875, 69.60869598388672, -76.99535369873047, 58.86474609375, 350.40814208984375, 250.16299438476562, 1.0434856414794922, 17.89947509765625, -213.163330078125, -30.3399658203125, -30.278722763061523, 27.548477172851562, 95.31912231445312, 417.80377197265625, -226.32015991210938, 164.61497497558594, 223.22889709472656, 442.0423889160156, 208.34149169921875, 592.4501953125, 212.7630615234375, 514.4752807617188, 412.11383056640625, -54.404205322265625, 274.0531311035156, 407.57672119140625, 52.339088439941406, 390.51568603515625, 132.2859344482422, 344.4322204589844, -90.41800689697266, 200.45042419433594, 54.12689971923828, 298.84674072265625, 59.86395263671875, 80.63052368164062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000275.npy"}
{"epoch": 0.40381791483113066, "step": 276, "batch_size": 64, "mean": 158.36476135253906, "std": 224.60455322265625, "min": -286.7627258300781, "p10": -50.14526634216308, "median": 147.6091079711914, "p90": 460.21464843750005, "max": 859.5872192382812, "pos_frac": 0.796875, "sample": [113.52494049072266, 859.5872192382812, 154.89076232910156, 467.1328125, 511.71099853515625, 13.704046249389648, -25.781143188476562, 671.8424072265625, 33.94413757324219, 170.10340881347656, 165.93890380859375, 16.68958282470703, -32.18170166015625, 264.4090576171875, 301.7294616699219, 271.3647766113281, 188.38262939453125, -286.7627258300781, 120.48004150390625, 9.299976348876953, 34.35737228393555, 399.43817138671875, 188.77723693847656, 356.40655517578125, -37.56553649902344, -251.6810302734375, -43.85628128051758, 68.31312561035156, 190.39093017578125, 352.3307189941406, -52.840545654296875, 183.4158935546875, 59.859458923339844, 158.5897216796875, 340.19903564453125, 269.72662353515625, -255.65106201171875, -207.41297912597656, 9.225214004516602, 13.794246673583984, -132.68609619140625, 52.679931640625, 48.95573043823242, 482.9588317871094, 97.36573028564453, 132.77056884765625, 273.15631103515625, 226.3921356201172, 318.5195007324219, 35.775794982910156, 484.934326171875, -27.757003784179688, 140.32745361328125, 251.22097778320312, -18.188323974609375, 241.00804138183594, 222.18556213378906, 159.95892333984375, 438.48284912109375, 88.6094970703125, -233.39071655273438, 604.4051513671875, 37.76029968261719, 444.072265625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000276.npy"}
{"epoch": 0.4052863436123348, "step": 277, "batch_size": 64, "mean": 164.54754638671875, "std": 215.87669372558594, "min": -274.06512451171875, "p10": -108.51144332885738, "median": 149.09385681152344, "p90": 476.7580932617188, "max": 581.092041015625, "pos_frac": 0.765625, "sample": [-17.465763092041016, 305.45001220703125, -175.80528259277344, 127.86920166015625, 28.336666107177734, 359.98748779296875, 395.1436767578125, 152.23275756835938, -124.95176696777344, 227.26416015625, 171.30093383789062, -10.602819442749023, 389.7440185546875, -37.32501983642578, 226.7090606689453, 393.81671142578125, 12.085470199584961, 128.43165588378906, 209.35638427734375, 311.49322509765625, -70.15068817138672, 410.4955749511719, 55.74287414550781, 42.150306701660156, -227.08551025390625, -57.30518341064453, 581.092041015625, 33.75634765625, -19.630584716796875, 260.6013488769531, -274.06512451171875, 513.1363525390625, 171.10537719726562, 342.349853515625, -236.30902099609375, -193.70098876953125, 546.9080200195312, 304.1463623046875, 145.9549560546875, -30.77239990234375, 276.907470703125, 50.6239013671875, 538.7166748046875, 41.02382278442383, 11.895782470703125, 116.62525177001953, 199.44406127929688, 424.7384948730469, 431.8720703125, 521.1275024414062, -158.90509033203125, 284.609619140625, 61.840187072753906, 103.4480209350586, 13.055511474609375, -7.142131805419922, 266.6680908203125, 159.23721313476562, 484.1623840332031, 480.7822570800781, 104.31590270996094, 173.11181640625, 114.02469635009766, 467.3683776855469], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000277.npy"}
{"epoch": 0.4067547723935389, "step": 278, "batch_size": 64, "mean": 191.06597900390625, "std": 212.43609619140625, "min": -158.03900146484375, "p10": -41.8957633972168, "median": 151.7436981201172, "p90": 509.4983642578125, "max": 659.5801391601562, "pos_frac": 0.78125, "sample": [-120.95610046386719, 356.288818359375, -53.18737030029297, 58.78897476196289, 359.3575134277344, 334.4716796875, 244.08706665039062, 55.68476486206055, 193.04835510253906, 117.4493179321289, 340.71142578125, 189.7234344482422, 44.11779022216797, -71.68923950195312, 103.07733154296875, 237.2952117919922, -23.1142578125, 323.6515808105469, 525.73046875, 12.841484069824219, -15.22496223449707, 653.865234375, 440.156005859375, 232.37237548828125, 340.53363037109375, 284.56060791015625, 133.92706298828125, -26.663246154785156, 171.84780883789062, 547.53369140625, 508.3251037597656, -91.71882629394531, 115.07290649414062, 47.30999755859375, 608.2230834960938, 503.54010009765625, 32.407493591308594, -37.86370849609375, 156.65444946289062, 236.44732666015625, 83.01416778564453, 444.25054931640625, 201.66192626953125, 43.64186096191406, 42.379638671875, 510.0011901855469, 137.08682250976562, 246.02699279785156, -42.49655532836914, 616.4512329101562, 208.62442016601562, 129.10972595214844, 33.11181640625, 17.52635955810547, 369.745849609375, 330.34979248046875, -30.871728897094727, -18.027334213256836, 659.5801391601562, -158.03900146484375, 356.6162109375, 146.83294677734375, -126.54531860351562, -40.49391555786133], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000278.npy"}
{"epoch": 0.40822320117474303, "step": 279, "batch_size": 64, "mean": 188.96109008789062, "std": 244.12782287597656, "min": -372.378662109375, "p10": -77.81048583984372, "median": 146.44042205810547, "p90": 500.25877685546874, "max": 911.9229736328125, "pos_frac": 0.796875, "sample": [83.78203582763672, 48.05895233154297, 66.19233703613281, 74.0850830078125, -52.2342529296875, -46.10875701904297, 410.795166015625, 514.6226806640625, 41.21788787841797, -47.50016403198242, 18.97869110107422, 305.23681640625, 148.89056396484375, 36.036712646484375, 497.2706604003906, 651.5120849609375, 58.95438766479492, 164.69204711914062, 272.7003479003906, 312.5380554199219, 711.2935791015625, 96.14058685302734, 18.753942489624023, -275.3804016113281, 31.89671516418457, 83.80760955810547, -23.63226318359375, 180.05636596679688, 89.26193237304688, 622.2075805664062, 136.0535888671875, 57.23137664794922, 248.68453979492188, -48.305259704589844, 283.4238586425781, 478.9476318359375, 285.302978515625, 911.9229736328125, 477.38836669921875, 312.61651611328125, -29.256820678710938, -88.771728515625, -159.00843811035156, -372.378662109375, 180.6861114501953, 508.4914855957031, 135.55136108398438, 262.39361572265625, 379.32623291015625, 143.9902801513672, 441.33782958984375, 353.6243896484375, 81.2706069946289, -126.70529174804688, 501.5393981933594, 193.89962768554688, 457.01763916015625, 289.35504150390625, -127.24696350097656, -134.51747131347656, 370.4696044921875, 290.1160583496094, 6.609279632568359, 298.3221130371094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000279.npy"}
{"epoch": 0.40969162995594716, "step": 280, "batch_size": 64, "mean": 194.56700134277344, "std": 196.5094757080078, "min": -238.09835815429688, "p10": -81.32214355468749, "median": 177.7982635498047, "p90": 455.73517150878905, "max": 588.8502197265625, "pos_frac": 0.796875, "sample": [140.7296905517578, 316.13897705078125, 137.42724609375, 210.6753692626953, 60.04170227050781, -63.922889709472656, 378.3176574707031, 513.5999755859375, 581.9307250976562, 26.40362548828125, 203.6557159423828, -6.969932556152344, -105.16069030761719, 416.70086669921875, 413.2937316894531, 408.5894775390625, 149.62213134765625, 287.521728515625, 184.04925537109375, 299.9064025878906, 127.74812316894531, 588.8502197265625, 452.0533142089844, 209.64356994628906, 280.0224609375, 69.7816162109375, 368.7352294921875, 314.9398193359375, -2.6273193359375, 295.4573974609375, 167.17892456054688, 171.54727172851562, 538.05126953125, 168.83309936523438, -89.6055908203125, 164.83494567871094, -146.9752197265625, 385.6891174316406, 251.27085876464844, 98.98307037353516, 209.03932189941406, 126.94293975830078, -65.32594299316406, 585.4181518554688, -50.44804763793945, -238.09835815429688, 159.14608764648438, 469.4710998535156, 184.5675048828125, 95.04781341552734, 214.50022888183594, 457.3131103515625, -88.17765808105469, 149.35206604003906, -25.001007080078125, -94.07329559326172, 212.19668579101562, 155.10987854003906, 77.002197265625, 368.33111572265625, 306.20556640625, -143.8101806640625, 87.24397277832031, 333.37213134765625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000280.npy"}
{"epoch": 0.4111600587371512, "step": 281, "batch_size": 64, "mean": 173.30999755859375, "std": 187.84364318847656, "min": -573.7628173828125, "p10": -6.742523956298828, "median": 167.90028381347656, "p90": 416.3775482177735, "max": 589.661376953125, "pos_frac": 0.859375, "sample": [89.29059600830078, 284.8836669921875, 171.98495483398438, 230.65538024902344, 71.46176147460938, 213.960205078125, 71.62356567382812, 217.8370361328125, 172.92738342285156, 60.38207244873047, 266.92352294921875, 400.37579345703125, 294.132568359375, 70.98029327392578, 251.2126007080078, -97.67791748046875, 179.13687133789062, 589.661376953125, 34.67347717285156, 154.51097106933594, -6.754638671875, 395.477294921875, 280.38299560546875, 542.8733520507812, 235.76388549804688, 247.00250244140625, 206.02499389648438, 280.04400634765625, -32.5567512512207, 580.8109130859375, 329.30517578125, 110.16513061523438, 214.13516235351562, 451.93218994140625, 243.91046142578125, -573.7628173828125, 81.44563293457031, 47.64387512207031, 121.1933822631836, 84.60955810546875, 481.84320068359375, 15.166671752929688, 1.9083099365234375, 279.554443359375, 163.81561279296875, 230.5149688720703, 100.0341796875, 118.91160583496094, 470.96844482421875, 103.25279998779297, -0.6134872436523438, 136.44326782226562, -98.76890563964844, 62.37811279296875, -40.743011474609375, -6.714256286621094, -47.444488525390625, 384.14788818359375, 267.67730712890625, 74.1683578491211, 135.08657836914062, 264.17279052734375, 4.260763168334961, 423.2354431152344], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000281.npy"}
{"epoch": 0.41262848751835535, "step": 282, "batch_size": 64, "mean": 133.78237915039062, "std": 215.61375427246094, "min": -261.46112060546875, "p10": -100.16380538940429, "median": 109.5755615234375, "p90": 372.3944396972657, "max": 1136.048583984375, "pos_frac": 0.765625, "sample": [458.91015625, 177.0931396484375, 93.68826293945312, -261.46112060546875, 188.2343292236328, 242.00320434570312, 100.122314453125, 31.785507202148438, 151.18865966796875, 417.7947998046875, 256.88555908203125, 215.65692138671875, 340.911376953125, 96.54837036132812, -41.95048522949219, 279.105712890625, 1136.048583984375, 152.01791381835938, 239.3671875, 119.02880859375, 156.45220947265625, 174.13351440429688, 30.668453216552734, -132.76123046875, 558.934326171875, 380.122314453125, 90.61253356933594, 213.49354553222656, 134.19580078125, 525.0223999023438, 300.38604736328125, 218.5548095703125, 283.0032653808594, -26.50421142578125, 12.630992889404297, 158.34280395507812, -47.02494812011719, -105.59537506103516, 70.60693359375, -102.78857421875, -34.33416748046875, 41.60723876953125, 175.38951110839844, 79.04495239257812, 238.75401306152344, 6.835289001464844, -142.67861938476562, 396.16741943359375, 171.59805297851562, 120.80340576171875, -66.8424301147461, -65.45396423339844, 28.237533569335938, 50.45351028442383, 94.51338195800781, -192.72946166992188, 291.814208984375, 4.205488204956055, -94.03934478759766, 354.36273193359375, -208.8720703125, 46.89582443237305, -52.043128967285156, 32.91734313964844], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000282.npy"}
{"epoch": 0.41409691629955947, "step": 283, "batch_size": 64, "mean": 193.72015380859375, "std": 208.96673583984375, "min": -361.97650146484375, "p10": -53.85848350524902, "median": 165.50507354736328, "p90": 485.2361541748047, "max": 623.4049072265625, "pos_frac": 0.8125, "sample": [-118.91412353515625, -90.94087982177734, 182.00120544433594, 39.521400451660156, 337.0249938964844, 332.67584228515625, -34.08209228515625, 292.09564208984375, 374.5603942871094, 56.48479461669922, 45.099693298339844, 88.74794006347656, -49.546077728271484, 96.53883361816406, 89.51176452636719, 159.725830078125, -361.97650146484375, 39.34762191772461, 413.88214111328125, -55.70665740966797, 478.8548278808594, 455.87200927734375, 165.5381622314453, 345.74102783203125, 87.72645568847656, 182.038330078125, -68.10411071777344, 505.12451171875, 411.26190185546875, 487.97100830078125, 13.40576171875, 316.1440734863281, -135.8023223876953, 211.23117065429688, 457.42462158203125, 47.28443908691406, 46.432891845703125, -8.913715362548828, 88.3866195678711, -68.17658996582031, 518.2123413085938, 164.6616973876953, 338.67449951171875, 238.24691772460938, 293.5731201171875, 149.66812133789062, 84.94695281982422, 226.35440063476562, 429.76361083984375, 298.0095520019531, 605.2508544921875, -17.585731506347656, 387.28729248046875, 165.47198486328125, 68.3458023071289, 306.2527770996094, 580.0018310546875, 179.97694396972656, 249.63275146484375, 623.4049072265625, 503.67144775390625, 95.92620086669922, 87.6124038696289, -34.76387023925781], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000283.npy"}
{"epoch": 0.4155653450807636, "step": 284, "batch_size": 64, "mean": 187.64157104492188, "std": 204.42652893066406, "min": -206.83831787109375, "p10": -43.215397262573234, "median": 156.73512268066406, "p90": 474.54623718261746, "max": 772.005126953125, "pos_frac": 0.859375, "sample": [-87.90316772460938, 148.9864044189453, 278.71990966796875, 38.85972595214844, 133.46682739257812, 58.44915008544922, 382.24822998046875, 19.742034912109375, 269.0996398925781, 410.2398681640625, 337.30767822265625, -47.09244155883789, 503.9074401855469, 10.317224502563477, 295.2286376953125, 23.080432891845703, 39.202354431152344, 223.14520263671875, 350.6855163574219, 502.1061096191406, 89.26272583007812, 705.9447631835938, 159.16290283203125, 309.0919494628906, 163.38470458984375, -34.16896057128906, 176.4344940185547, 74.828369140625, 238.6553497314453, -16.968490600585938, 54.889434814453125, 673.0858764648438, 205.7360076904297, -93.8618392944336, 142.34677124023438, 115.56085205078125, -50.785499572753906, 67.47283172607422, 216.65811157226562, 772.005126953125, 48.79112243652344, 181.45138549804688, 198.9486083984375, 37.68980407714844, -115.37570190429688, -206.83831787109375, 341.59326171875, 352.2875671386719, 112.97447204589844, 286.6568603515625, 516.9727783203125, 154.30734252929688, 322.7322692871094, 246.87757873535156, 77.17057800292969, 19.406139373779297, -59.28478240966797, 5.951927185058594, 322.71771240234375, 258.5933837890625, 29.980262756347656, 561.156982421875, 52.52318572998047, 403.244140625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000284.npy"}
{"epoch": 0.4170337738619677, "step": 285, "batch_size": 64, "mean": 164.9952850341797, "std": 193.39703369140625, "min": -550.6981811523438, "p10": -43.645685577392555, "median": 167.80145263671875, "p90": 363.083203125, "max": 591.9623413085938, "pos_frac": 0.875, "sample": [362.62664794921875, 241.38009643554688, 45.82887268066406, 363.27886962890625, 124.43708038330078, 124.32563781738281, 177.7088623046875, 52.899269104003906, 176.00927734375, 579.4395751953125, 261.17218017578125, 186.4564666748047, 356.4088439941406, 116.8433609008789, 155.3994140625, 569.0712280273438, 41.0225830078125, 84.22576904296875, 159.5936279296875, 246.546875, 434.5775451660156, -194.45889282226562, 87.2719955444336, 301.23577880859375, 420.8929748535156, 3.1932449340820312, -53.186588287353516, 66.64276123046875, 51.777923583984375, 89.71048736572266, 276.84393310546875, 156.5969696044922, -68.0873794555664, 49.49581527709961, 267.630859375, 358.8970642089844, 62.64466094970703, -550.6981811523438, 245.4224853515625, 59.008270263671875, 193.43336486816406, 17.958648681640625, 591.9623413085938, 273.4294128417969, 1.2397079467773438, 298.58038330078125, 270.28662109375, 155.74331665039062, 238.67376708984375, 339.8691711425781, 213.15328979492188, -209.14529418945312, 286.6933288574219, 217.78994750976562, -157.1343994140625, -21.38357925415039, 453.93878173828125, 72.2647705078125, 93.3958969116211, 183.7571258544922, -103.52294158935547, 322.66168212890625, 205.80642700195312, 130.16036987304688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000285.npy"}
{"epoch": 0.4185022026431718, "step": 286, "batch_size": 64, "mean": 187.3711700439453, "std": 223.2196044921875, "min": -349.8279113769531, "p10": -81.38945922851562, "median": 201.33332061767578, "p90": 495.0781066894532, "max": 694.3087158203125, "pos_frac": 0.78125, "sample": [88.7552719116211, -76.58629608154297, 508.9061584472656, 160.49951171875, 204.94309997558594, 197.72354125976562, 93.96416473388672, 251.6739044189453, 325.2784729003906, -349.8279113769531, 281.5730285644531, 218.53298950195312, 20.816024780273438, -96.98665618896484, 431.6409606933594, 220.09005737304688, 568.1705322265625, -26.68328857421875, 222.12274169921875, 176.20407104492188, 501.73968505859375, 49.35699462890625, -290.6814270019531, 454.52215576171875, 304.84368896484375, -79.3409423828125, -68.37953186035156, 376.92315673828125, 299.2326354980469, 583.7457275390625, 447.0152587890625, -164.268798828125, 248.02169799804688, 49.115570068359375, 29.956405639648438, 62.20716857910156, 414.24664306640625, 157.98509216308594, 42.90802001953125, -34.49449157714844, 236.4333038330078, -77.09334564208984, 247.79177856445312, -89.42826080322266, 91.00674438476562, 426.4721374511719, 479.534423828125, 163.215576171875, -107.24928283691406, 551.2177734375, -49.72418975830078, 221.35037231445312, 61.96617126464844, 524.6607666015625, 74.9569091796875, 333.13092041015625, 122.62411499023438, 249.0806884765625, 310.8601989746094, -82.26739501953125, 694.3087158203125, 405.2914733886719, 291.6544189453125, 106.4958267211914], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000286.npy"}
{"epoch": 0.4199706314243759, "step": 287, "batch_size": 64, "mean": 149.5043182373047, "std": 195.03195190429688, "min": -290.8348693847656, "p10": -76.22454681396482, "median": 127.07719421386719, "p90": 440.92284545898445, "max": 699.7320556640625, "pos_frac": 0.796875, "sample": [175.2323455810547, 24.759490966796875, -6.673379898071289, 47.93596649169922, 120.26009368896484, -57.897003173828125, 310.41943359375, 550.0513305664062, 152.30955505371094, 137.56915283203125, 557.7200317382812, 81.71797943115234, 209.2126922607422, 518.2883911132812, -84.07920837402344, -50.70915222167969, 83.1369857788086, 465.4923095703125, 104.40618896484375, 332.22186279296875, 202.0728759765625, 6.768100738525391, 137.9921875, -290.8348693847656, 699.7320556640625, 65.95957946777344, 77.42388916015625, -7.3116455078125, -16.741409301757812, 19.638450622558594, 142.13595581054688, -11.496040344238281, -105.65557861328125, 138.86708068847656, 80.90220642089844, 233.8911895751953, 68.16511535644531, 249.2390899658203, 145.71304321289062, 251.4219207763672, 52.3900146484375, 296.123046875, 241.60943603515625, 128.9768524169922, 10.997268676757812, 125.17753601074219, 82.85049438476562, 66.39045715332031, 380.33905029296875, 365.6265869140625, -163.2978057861328, 50.81011199951172, 237.59327697753906, 115.20345306396484, 415.0658264160156, -203.36166381835938, -175.74386596679688, -88.2458724975586, 195.05828857421875, 301.2184753417969, 467.82611083984375, 452.0044250488281, 195.64120483398438, 258.7651062011719], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000287.npy"}
{"epoch": 0.42143906020558003, "step": 288, "batch_size": 64, "mean": 188.77154541015625, "std": 196.6981964111328, "min": -291.787109375, "p10": 2.6559745788574283, "median": 174.22957611083984, "p90": 402.4691864013672, "max": 685.646728515625, "pos_frac": 0.90625, "sample": [182.16104125976562, 175.9810791015625, -16.378047943115234, 171.14230346679688, 323.3966064453125, 402.9073486328125, 32.791748046875, 685.646728515625, 34.91032409667969, 401.4468078613281, 238.052734375, 155.57943725585938, -98.1315689086914, 17.1571044921875, 81.57814025878906, 179.76515197753906, 131.74285888671875, 393.7131652832031, 220.79913330078125, 316.5777893066406, 175.5381317138672, 275.8583984375, 299.94134521484375, 14.7266845703125, 89.53031921386719, 236.09396362304688, 197.29248046875, 26.828989028930664, 172.9210205078125, 10.389999389648438, 182.46963500976562, 222.88821411132812, 632.4684448242188, 170.24627685546875, 400.4921875, 351.667236328125, 163.22174072265625, 8.823822021484375, 261.53948974609375, 97.22586822509766, -291.787109375, 586.8309326171875, 109.4134521484375, 282.139892578125, 71.60490417480469, 169.8397216796875, 78.28821563720703, 40.903656005859375, 556.8779296875, 114.12584686279297, 229.96852111816406, 555.4392700195312, 278.78985595703125, -274.40850830078125, 118.02577209472656, 503.576171875, 279.0876770019531, 91.38917541503906, 72.73304748535156, 372.77276611328125, 0.01261138916015625, 382.162109375, -142.07582092285156, -125.33570861816406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000288.npy"}
{"epoch": 0.42290748898678415, "step": 289, "batch_size": 64, "mean": 208.65264892578125, "std": 220.25396728515625, "min": -181.81216430664062, "p10": -25.89182319641113, "median": 165.96237182617188, "p90": 510.805697631836, "max": 727.1712036132812, "pos_frac": 0.828125, "sample": [167.26251220703125, -8.908172607421875, 604.0603637695312, 677.9888916015625, -176.93075561523438, 577.0504150390625, 350.17095947265625, 304.78155517578125, 36.939208984375, 533.6783447265625, 503.8321228027344, 414.70391845703125, 262.97161865234375, 501.38006591796875, 411.7052001953125, 287.55682373046875, 54.34709167480469, 727.1712036132812, -56.50438690185547, 286.27679443359375, 140.9654541015625, 36.778594970703125, 182.89764404296875, -21.239360809326172, -115.80818176269531, 80.46163940429688, 258.7418518066406, 409.4559631347656, 69.38111877441406, 148.7047576904297, 423.9242858886719, 649.6441040039062, 354.3210754394531, -26.258724212646484, 73.47418975830078, 16.950515747070312, 474.24432373046875, 76.68235778808594, -96.80740356445312, 19.11707878112793, 99.413818359375, 414.98284912109375, -25.035720825195312, 255.11526489257812, 160.027587890625, 426.176025390625, 401.4076843261719, 1.9767475128173828, 100.82682800292969, -84.18887329101562, 174.17225646972656, 164.6622314453125, 93.1602783203125, 41.93601989746094, 513.7943725585938, 197.45416259765625, 34.48296356201172, -7.651252746582031, 132.55307006835938, 228.61520385742188, 278.3479919433594, -181.81216430664062, 229.00119018554688, 89.1865463256836], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000289.npy"}
{"epoch": 0.4243759177679883, "step": 290, "batch_size": 64, "mean": 213.74664306640625, "std": 262.31243896484375, "min": -310.6365966796875, "p10": -51.740492630004866, "median": 197.78682708740234, "p90": 588.6980346679688, "max": 988.450927734375, "pos_frac": 0.8125, "sample": [548.98291015625, 201.48236083984375, 118.25902557373047, 140.40585327148438, -310.6365966796875, 380.9624938964844, 468.83203125, -31.44677734375, -19.565505981445312, 133.1722412109375, -221.45932006835938, 380.14959716796875, 244.63189697265625, 41.124298095703125, 278.5189208984375, 228.88751220703125, 513.79296875, -12.4468994140625, 234.53750610351562, 6.401828765869141, 41.82026672363281, 30.382980346679688, -170.8790283203125, 174.97158813476562, 128.22744750976562, -276.3343505859375, 122.733642578125, 271.2451171875, 988.450927734375, 751.6607666015625, 72.87677001953125, 263.79571533203125, 322.00518798828125, 687.5091552734375, 192.75421142578125, 16.57359504699707, 277.7412414550781, 84.61933898925781, 265.4648132324219, 206.8087158203125, 0.20263671875, 206.10914611816406, 194.09129333496094, 340.67840576171875, 55.06306457519531, 285.77032470703125, -35.91564178466797, -66.15797424316406, 319.5599365234375, -111.1220703125, 680.4033203125, 449.07476806640625, 255.75503540039062, 629.6700439453125, 17.068477630615234, 595.802490234375, 572.1209716796875, 329.1323547363281, -58.5225715637207, 89.70121765136719, 82.75042724609375, 828.148681640625, 268.0196838378906, -24.63321304321289], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000290.npy"}
{"epoch": 0.42584434654919234, "step": 291, "batch_size": 64, "mean": 186.96609497070312, "std": 225.4898681640625, "min": -434.4881896972656, "p10": -74.51049804687497, "median": 176.11073303222656, "p90": 476.00030822753916, "max": 622.0753173828125, "pos_frac": 0.78125, "sample": [182.71014404296875, 168.88475036621094, 0.87811279296875, 35.46906280517578, 3.847412109375, 300.559326171875, 622.0753173828125, 296.3050842285156, 495.34967041015625, 290.2300720214844, 48.95466613769531, 456.886962890625, 282.0101623535156, -84.90605926513672, 598.5157470703125, -49.575164794921875, 443.3253479003906, 234.5415802001953, -50.254188537597656, 227.7058563232422, 415.310302734375, -9.435165405273438, 331.89642333984375, 87.11404418945312, 264.37744140625, 420.07342529296875, 130.39495849609375, 146.23423767089844, 453.5294494628906, 37.91042709350586, 232.76608276367188, 76.10883331298828, 323.2490234375, -225.78981018066406, -20.7646484375, -204.85049438476562, -17.418052673339844, 38.605918884277344, -434.4881896972656, 618.207763671875, 308.59600830078125, 484.1917419433594, 416.5129089355469, 237.93060302734375, -114.30220031738281, -30.297027587890625, -85.19881439208984, 388.9744873046875, 101.86602783203125, 169.51132202148438, 124.77021789550781, 168.12014770507812, 357.1431884765625, 293.2074890136719, 8.968902587890625, 215.31033325195312, 385.8453674316406, 84.50434875488281, 524.352783203125, -34.719688415527344, -130.88009643554688, 407.0875244140625, 515.6890869140625, 2.0987281799316406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000291.npy"}
{"epoch": 0.42731277533039647, "step": 292, "batch_size": 64, "mean": 203.70681762695312, "std": 268.582275390625, "min": -744.5189208984375, "p10": -62.361013793945304, "median": 190.36669158935547, "p90": 526.829281616211, "max": 826.6534423828125, "pos_frac": 0.796875, "sample": [253.04931640625, 9.061466217041016, -7.0415496826171875, 205.11917114257812, 282.0487365722656, 180.8843994140625, 179.184326171875, 107.22750091552734, 184.2896728515625, -744.5189208984375, 723.7734375, 62.303550720214844, 638.9796142578125, 120.67062377929688, 336.8044128417969, -65.09281921386719, -285.16119384765625, 438.2001037597656, 68.90921783447266, 346.6023254394531, 5.684669494628906, 802.3848266601562, 459.4077453613281, 201.47097778320312, 451.9984436035156, 511.1235656738281, 277.12933349609375, 484.862060546875, 170.93121337890625, 153.18936157226562, -120.08584594726562, 244.52426147460938, 9.815444946289062, 533.560302734375, -6.480155944824219, 132.74957275390625, -55.98680114746094, 181.3528594970703, 110.97178649902344, 116.35374450683594, 38.865875244140625, 506.2301025390625, 257.76177978515625, 219.51560974121094, 227.97076416015625, 223.46539306640625, 488.2015686035156, 348.79541015625, 826.6534423828125, -34.843841552734375, 184.94525146484375, 313.1012878417969, 238.6007537841797, -224.0845489501953, 195.7881317138672, -172.4248046875, 731.4232177734375, -74.97720336914062, -24.405818939208984, 564.015380859375, 216.1162872314453, 225.33590698242188, 87.48103332519531, -26.54478645324707], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000292.npy"}
{"epoch": 0.4287812041116006, "step": 293, "batch_size": 64, "mean": 159.80654907226562, "std": 251.48907470703125, "min": -305.4943542480469, "p10": -120.94404220581055, "median": 135.41266250610352, "p90": 486.93957824707036, "max": 983.1134033203125, "pos_frac": 0.703125, "sample": [38.151790618896484, 65.23043823242188, 31.157400131225586, 52.292816162109375, 234.8221435546875, 219.1837921142578, 481.6713562011719, 164.3135528564453, 446.84088134765625, 51.81977081298828, 490.5764465332031, -72.56355285644531, 238.68292236328125, -14.324581146240234, 983.1134033203125, -305.4943542480469, 336.1222229003906, 176.56552124023438, 77.74752807617188, 263.7253112792969, 365.8487854003906, -170.90985107421875, 258.1194152832031, -56.822059631347656, 44.187217712402344, 54.85607147216797, -103.36080932617188, 159.9056396484375, -16.491561889648438, 530.501220703125, 504.4033203125, 320.7445983886719, 322.48675537109375, 436.75372314453125, 302.4436950683594, -161.36976623535156, 42.48988342285156, 171.0486297607422, 164.95330810546875, -29.112112045288086, -121.93383026123047, 110.91968536376953, -32.017822265625, 260.8895568847656, 313.277587890625, 397.1668395996094, 219.57174682617188, 338.8336181640625, 110.54118347167969, -234.64810180664062, -8.428497314453125, -293.098876953125, -118.63453674316406, 489.1973876953125, 843.1558837890625, 94.5312271118164, -100.42678833007812, 160.99029541015625, 72.55982971191406, -98.14009094238281, -144.2269287109375, 375.7169494628906, 533.6419067382812, -42.129966735839844], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000293.npy"}
{"epoch": 0.4302496328928047, "step": 294, "batch_size": 64, "mean": 183.0571746826172, "std": 247.13404846191406, "min": -318.3774108886719, "p10": -121.83783874511712, "median": 105.88597106933594, "p90": 524.2427124023437, "max": 913.450439453125, "pos_frac": 0.859375, "sample": [57.14350128173828, 70.5767822265625, 320.9577331542969, 432.9874267578125, 77.43988037109375, 32.27439498901367, -150.33041381835938, 141.54122924804688, 4.715240478515625, 81.9772720336914, 21.877822875976562, 52.507537841796875, -56.61446762084961, 571.2433471679688, 82.6072006225586, 283.8547668457031, 248.1181182861328, 450.4617004394531, 60.4177131652832, 186.40797424316406, 714.5092163085938, -182.52867126464844, 308.1756896972656, 89.20327758789062, 681.9991455078125, 523.0191650390625, 343.5911865234375, 171.2119140625, 192.09420776367188, -149.79071044921875, 297.4577331542969, 318.115478515625, 321.92327880859375, 913.450439453125, -251.94442749023438, -157.2821502685547, 604.364501953125, 185.5393524169922, 474.6796569824219, 15.377540588378906, -318.3774108886719, -172.09564208984375, 270.02276611328125, 305.5537414550781, 596.4386596679688, 122.56866455078125, 22.47269058227539, 155.6211395263672, 524.76708984375, 8.99681282043457, 352.8268737792969, 352.9156799316406, 31.99620819091797, 163.27601623535156, -17.060138702392578, 72.16302490234375, 60.9672966003418, 43.71057891845703, 19.964435577392578, 82.92337036132812, 521.7040405273438, 62.293060302734375, 23.790679931640625, 44.88898468017578], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000294.npy"}
{"epoch": 0.43171806167400884, "step": 295, "batch_size": 64, "mean": 205.2287139892578, "std": 273.3934326171875, "min": -337.83270263671875, "p10": -139.53794326782224, "median": 176.48563385009766, "p90": 551.5036865234375, "max": 1001.17333984375, "pos_frac": 0.78125, "sample": [-316.8210754394531, -66.4739990234375, 659.20068359375, 12.118410110473633, -337.83270263671875, 167.28704833984375, -200.86387634277344, 306.4330139160156, 373.3380432128906, -242.41357421875, 360.23358154296875, -10.849288940429688, 168.77529907226562, 404.0787353515625, 446.0076904296875, 256.24951171875, 204.9907989501953, 276.437255859375, -103.07244110107422, 143.2225341796875, 361.98175048828125, -75.70296478271484, 308.08563232421875, 443.9115905761719, 34.011131286621094, 150.83514404296875, 551.804931640625, -155.166015625, 804.1375732421875, 153.29690551757812, 506.6347351074219, 75.47515869140625, -32.79935073852539, -181.90463256835938, 465.34002685546875, 618.5626220703125, 72.13603973388672, -59.96466827392578, -221.1505889892578, 38.89164733886719, 266.4671630859375, 417.5477294921875, 368.33209228515625, 108.68678283691406, -33.25049591064453, 97.92051696777344, 241.8189239501953, 83.62586212158203, 214.07867431640625, 550.80078125, 338.6587829589844, 425.27667236328125, 35.986785888671875, 1001.17333984375, 554.889892578125, 443.0732116699219, 172.49490356445312, 180.4763641357422, 204.93446350097656, 195.63023376464844, 31.96832847595215, 123.15817260742188, 736.53759765625, 15.888221740722656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000295.npy"}
{"epoch": 0.4331864904552129, "step": 296, "batch_size": 64, "mean": 136.5897979736328, "std": 295.1600036621094, "min": -793.0938720703125, "p10": -293.4412170410156, "median": 137.51750946044922, "p90": 514.31748046875, "max": 852.36181640625, "pos_frac": 0.71875, "sample": [123.55745697021484, 49.17993927001953, 203.91888427734375, 852.36181640625, -153.95806884765625, 621.0537109375, -296.51165771484375, -73.35234069824219, -53.307273864746094, -793.0938720703125, 110.20773315429688, 236.28778076171875, -298.39813232421875, 230.49563598632812, 150.5067596435547, -286.27685546875, 158.73561096191406, 362.36590576171875, 634.8729248046875, 422.01220703125, 218.12371826171875, 275.53607177734375, 18.975242614746094, -113.26361083984375, 125.69854736328125, 380.20098876953125, 34.367523193359375, -304.0048828125, 18.251434326171875, -310.49908447265625, 679.2666625976562, -3.7195587158203125, 255.7254180908203, 100.30916595458984, 285.67071533203125, 574.555419921875, 141.14930725097656, 98.8411865234375, -359.2823791503906, 228.82867431640625, -378.6367492675781, 269.0220031738281, -103.50440216064453, -22.463411331176758, 512.4287109375, 124.31299591064453, 141.12615966796875, 231.33644104003906, 158.83059692382812, 169.676025390625, 113.8367691040039, -20.92101287841797, -29.642539978027344, 829.2742309570312, 186.5205535888672, 238.55072021484375, 90.14151763916016, 34.332603454589844, 515.126953125, -95.11880493164062, 133.9088592529297, 265.0482482910156, 465.1304931640625, 368.04107666015625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000296.npy"}
{"epoch": 0.434654919236417, "step": 297, "batch_size": 64, "mean": 162.59478759765625, "std": 211.89878845214844, "min": -282.6131896972656, "p10": -54.992729568481444, "median": 117.13555145263672, "p90": 455.50895996093755, "max": 674.8621215820312, "pos_frac": 0.765625, "sample": [156.8982696533203, 249.24134826660156, 378.4864196777344, 576.0145263671875, 27.418254852294922, 5.528741836547852, 319.94793701171875, 196.51568603515625, -175.36087036132812, -54.07988357543945, 44.93939208984375, 117.36376953125, 236.1353759765625, 439.73614501953125, 385.4659118652344, 20.152013778686523, 16.11285400390625, 30.377532958984375, -98.98788452148438, -13.69759750366211, 288.15399169921875, -136.32728576660156, 116.90733337402344, -18.452714920043945, 552.9349365234375, 94.4169921875, 288.217041015625, 69.8829345703125, 215.84121704101562, 386.4877014160156, 176.79751586914062, 96.15756225585938, 484.49822998046875, 199.0077667236328, 449.27398681640625, -0.13791275024414062, 104.80009460449219, -1.5565109252929688, 531.1182861328125, 296.3382873535156, -6.2601165771484375, 169.52313232421875, -248.08642578125, 363.5428466796875, 75.24088287353516, -55.383949279785156, 121.88496398925781, 46.32684326171875, 674.8621215820312, -61.10209655761719, 123.75677490234375, 458.18109130859375, 262.8914794921875, -17.185546875, -282.6131896972656, 344.2086181640625, 27.93042755126953, 49.93611145019531, 231.31369018554688, 603.134765625, 96.61152648925781, 399.5772705078125, -38.04155731201172, 13.24907112121582], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000297.npy"}
{"epoch": 0.43612334801762115, "step": 298, "batch_size": 64, "mean": 213.7909698486328, "std": 322.3824157714844, "min": -675.9561157226562, "p10": -116.88758392333983, "median": 154.53738403320312, "p90": 633.7767578125, "max": 1518.670166015625, "pos_frac": 0.75, "sample": [146.98825073242188, -119.69149780273438, 27.074321746826172, -675.9561157226562, 20.080657958984375, -122.19000244140625, 7.46025276184082, 650.8672485351562, 295.8125305175781, -12.05059814453125, 163.91329956054688, 472.070556640625, 0.7606658935546875, -112.35481262207031, -169.15896606445312, 73.27922821044922, 192.610107421875, 127.95770263671875, 131.47695922851562, 281.9972229003906, 815.8963623046875, -8.0528564453125, 397.09051513671875, -166.16085815429688, 646.5142822265625, 276.3673095703125, -106.85719299316406, 580.6058959960938, 615.1170654296875, 742.4189453125, -88.59111785888672, -118.8302001953125, -23.011436462402344, 1.3723983764648438, -27.328758239746094, 637.65185546875, 675.53955078125, 144.3324432373047, 372.55548095703125, 200.99301147460938, 8.311796188354492, 7.9592437744140625, 1518.670166015625, 277.96319580078125, 624.73486328125, 217.46975708007812, 458.83319091796875, 436.9645080566406, -5.676235198974609, 83.12394714355469, 216.84506225585938, 461.00115966796875, 321.121337890625, 101.03443145751953, 552.3748779296875, 143.7024383544922, -57.6324462890625, 393.7354431152344, -128.87599182128906, 277.8360290527344, 162.08651733398438, 238.9447479248047, 59.09843444824219, 364.42608642578125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000298.npy"}
{"epoch": 0.43759177679882527, "step": 299, "batch_size": 64, "mean": 183.4227294921875, "std": 263.05072021484375, "min": -354.2772216796875, "p10": -115.2706314086914, "median": 152.91412353515625, "p90": 556.2069335937501, "max": 858.6271362304688, "pos_frac": 0.75, "sample": [-107.39063262939453, -295.33837890625, -354.2772216796875, 324.49774169921875, 61.40509033203125, -35.69839859008789, 359.3606262207031, 132.540283203125, -12.985754013061523, 110.96475982666016, 582.9197998046875, 571.2388305664062, 788.7965087890625, -58.76573944091797, 176.22352600097656, 781.1231079101562, -206.1373291015625, 389.7911071777344, 174.0859832763672, 231.07872009277344, -118.64777374267578, 339.1361083984375, 126.50804901123047, 148.4608612060547, 432.8478698730469, 119.227294921875, 858.6271362304688, 14.137714385986328, 249.62826538085938, -24.028690338134766, -37.13495635986328, 147.07679748535156, -138.67630004882812, 73.71482849121094, 195.2706298828125, 521.1325073242188, 133.4379119873047, 207.9773406982422, -54.422943115234375, 159.2542724609375, 70.15877532958984, 329.3852233886719, 249.02915954589844, -18.21249771118164, 229.55130004882812, -107.35462188720703, -195.9983673095703, 483.7250671386719, 33.901123046875, 811.1285400390625, -145.13833618164062, 73.47948455810547, 256.8031921386719, 315.3132629394531, 254.6085205078125, 320.3564453125, 60.99163818359375, 673.4271240234375, 272.2191162109375, 157.3673858642578, 53.048980712890625, 363.4769592285156, 36.040992736816406, 194.7864990234375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000299.npy"}
{"epoch": 0.4390602055800294, "step": 300, "batch_size": 64, "mean": 222.55722045898438, "std": 285.59686279296875, "min": -392.91436767578125, "p10": -173.02040405273436, "median": 234.223876953125, "p90": 596.7128173828125, "max": 800.21728515625, "pos_frac": 0.796875, "sample": [313.1877746582031, 296.6683654785156, 154.9563446044922, 379.58587646484375, 601.0057983398438, -372.72711181640625, -156.0899658203125, 657.8386840820312, 417.13580322265625, -4.331943511962891, -88.29549407958984, 300.41943359375, 184.1913299560547, 496.1178283691406, 274.42486572265625, 432.3711242675781, 132.4249725341797, 524.1130981445312, 46.059814453125, 53.57939910888672, -113.8780517578125, 221.67422485351562, 256.6419372558594, 21.438255310058594, 469.4508361816406, 365.7694091796875, -143.28823852539062, 69.25489807128906, 168.73866271972656, 296.76373291015625, 94.20851135253906, 647.4871826171875, 282.7581787109375, -392.91436767578125, 481.94891357421875, -273.16339111328125, 586.6958618164062, 318.88763427734375, 456.52740478515625, 234.76309204101562, 86.32593536376953, 457.62481689453125, 639.49462890625, 371.2557678222656, 95.27157592773438, 374.3053894042969, -31.229907989501953, -180.27630615234375, 769.029541015625, 450.68408203125, 443.42633056640625, -268.58917236328125, 233.68466186523438, -267.41961669921875, 10.235233306884766, 217.52041625976562, 494.3818054199219, 76.82217407226562, 711.9332275390625, 800.21728515625, -203.3846435546875, 38.29448318481445, 198.31878662109375, 33.33549499511719], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000300.npy"}
{"epoch": 0.44052863436123346, "step": 301, "batch_size": 64, "mean": 214.12109375, "std": 256.1773376464844, "min": -179.1638946533203, "p10": -61.91326904296875, "median": 157.8402862548828, "p90": 565.391394042969, "max": 916.3235473632812, "pos_frac": 0.828125, "sample": [191.20689392089844, 306.3600158691406, -71.57947540283203, 365.2630920410156, 387.3919982910156, 262.2044677734375, 842.7662353515625, 41.69999694824219, 527.8984985351562, -39.66258239746094, 828.893798828125, 394.6725158691406, 75.4772720336914, 253.44822692871094, -73.18360900878906, 155.47357177734375, 99.69805145263672, 581.4597778320312, 49.81413269042969, 355.4619140625, 799.5911254882812, 28.551502227783203, -58.935523986816406, 58.25993347167969, 411.5314025878906, 49.502532958984375, 75.1455307006836, 24.747756958007812, 279.7001037597656, 399.213623046875, 356.2625732421875, 308.1778564453125, 40.994869232177734, 165.74203491210938, 55.27594757080078, 79.2576904296875, 195.69630432128906, -156.5039825439453, 11.208244323730469, 22.104248046875, 108.95989990234375, 47.29541778564453, 249.97689819335938, -63.18944549560547, 45.21183776855469, 916.3235473632812, 354.9178466796875, 239.70681762695312, 592.6919555664062, 13.596893310546875, 527.2742309570312, 160.20700073242188, 308.3251037597656, 515.8063354492188, 429.010986328125, 72.13880920410156, -81.4915771484375, -179.1638946533203, 607.0830078125, -172.99435424804688, 97.99076843261719, -57.75897216796875, -3.9132347106933594, 295.45574951171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000301.npy"}
{"epoch": 0.4419970631424376, "step": 302, "batch_size": 64, "mean": 178.4041748046875, "std": 234.73150634765625, "min": -254.1911163330078, "p10": -89.22050933837889, "median": 155.19847106933594, "p90": 494.03751525878914, "max": 867.176513671875, "pos_frac": 0.75, "sample": [15.545700073242188, 385.4400939941406, 710.9442138671875, -12.30868148803711, 196.55145263671875, 5.117158889770508, 243.9486846923828, -0.2501220703125, 170.23878479003906, -0.43634033203125, 312.7873229980469, 31.080245971679688, 48.842918395996094, 51.393531799316406, 164.50469970703125, 54.13429260253906, 402.513671875, -5.702732086181641, 386.8939208984375, 300.71380615234375, 228.24314880371094, 188.05087280273438, 197.717041015625, 255.46121215820312, -97.29959106445312, -124.4955062866211, 381.00201416015625, 597.0450439453125, 90.10118865966797, -254.1911163330078, 198.09860229492188, -66.91226959228516, 119.9530029296875, -4.0445098876953125, -78.63044738769531, 499.7527770996094, 867.176513671875, 509.59912109375, 168.94876098632812, -79.70382690429688, -26.682952880859375, 480.701904296875, -225.5850830078125, 220.80987548828125, 426.0982666015625, 71.81249237060547, 121.60292053222656, 651.7185668945312, 254.9427947998047, 318.44189453125, 23.07975196838379, 264.02490234375, -156.60488891601562, 292.4830322265625, 101.56517028808594, 267.98260498046875, -93.29908752441406, 454.5206604003906, 87.01683044433594, 650.703857421875, -152.76040649414062, 98.04905700683594, 145.89224243164062, 83.5284652709961], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000302.npy"}
{"epoch": 0.4434654919236417, "step": 303, "batch_size": 64, "mean": 159.99368286132812, "std": 240.21298217773438, "min": -473.1531066894531, "p10": -122.09793624877929, "median": 133.6910858154297, "p90": 503.9039031982422, "max": 648.7640991210938, "pos_frac": 0.703125, "sample": [301.46905517578125, -13.201343536376953, 85.73112487792969, 571.98046875, -118.84418487548828, 407.8822021484375, 460.89862060546875, 639.9293823242188, -16.799104690551758, -123.49240112304688, 382.25177001953125, 182.92733764648438, 297.1487731933594, -95.52337646484375, 80.5815658569336, 176.36862182617188, 99.51692199707031, 36.78010559082031, -54.49474334716797, 396.36724853515625, -473.1531066894531, -15.600139617919922, -92.92891693115234, 330.1385192871094, 431.5083312988281, 504.99871826171875, -124.6539077758789, 277.9144287109375, -225.34988403320312, 129.28135681152344, -56.2994384765625, 305.07904052734375, 648.7640991210938, 213.68283081054688, 463.9720458984375, 194.18975830078125, -161.25711059570312, 129.601318359375, 113.61243438720703, 637.8954467773438, 569.016845703125, -14.098876953125, 199.90274047851562, 501.3493347167969, 137.78085327148438, 205.45895385742188, 248.42901611328125, 294.6923522949219, 150.26547241210938, -65.92095184326172, 105.22689819335938, -39.645912170410156, 209.91380310058594, 51.5871467590332, 108.36318969726562, 543.4553833007812, -194.07388305664062, -45.04801940917969, 47.70650100708008, 39.25221633911133, 249.5624542236328, 49.450408935546875, -183.08096313476562, 141.17672729492188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000303.npy"}
{"epoch": 0.44493392070484583, "step": 304, "batch_size": 64, "mean": 176.21649169921875, "std": 159.255126953125, "min": -234.4119110107422, "p10": -5.435072326660148, "median": 168.35804748535156, "p90": 374.8068542480469, "max": 587.6840209960938, "pos_frac": 0.890625, "sample": [11.077316284179688, 221.76963806152344, 289.08270263671875, 149.03350830078125, -8.991104125976562, -21.271957397460938, 2.862335205078125, 5.1171875, 184.59884643554688, 154.90728759765625, 281.47601318359375, 376.5869140625, 180.75320434570312, 311.53021240234375, 72.85606384277344, 257.2333984375, 280.263916015625, 278.6555480957031, 15.156791687011719, 88.8643798828125, 147.14126586914062, 362.17498779296875, 39.35224914550781, 69.27386474609375, 268.1565246582031, 89.61795043945312, 270.921142578125, 59.90391540527344, 410.04669189453125, 105.20944213867188, 127.5946044921875, 105.56167602539062, 587.6840209960938, 129.63507080078125, -36.040184020996094, -234.4119110107422, 228.64764404296875, -32.0557861328125, 43.27308654785156, 256.406494140625, -66.84591674804688, 446.61431884765625, 22.079513549804688, 56.27507781982422, 425.578369140625, 114.69625854492188, 12.774024963378906, 64.08565521240234, 218.68905639648438, 346.1197814941406, 275.12371826171875, -132.64028930664062, 197.95677185058594, 454.5703430175781, 398.4083251953125, 79.09941101074219, 241.02499389648438, 155.962890625, 370.65338134765625, 364.0090026855469, 335.54833984375, 224.30203247070312, 331.5086364746094, 212.60662841796875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000304.npy"}
{"epoch": 0.44640234948604995, "step": 305, "batch_size": 64, "mean": 159.7189178466797, "std": 221.27178955078125, "min": -265.2970275878906, "p10": -70.65944519042968, "median": 129.81434631347656, "p90": 436.6955810546875, "max": 1019.568603515625, "pos_frac": 0.78125, "sample": [369.4222717285156, 240.76739501953125, 121.66706848144531, 228.8957977294922, 150.6127166748047, 598.6942138671875, 243.43212890625, 39.42584228515625, 266.35321044921875, 421.5235290527344, 115.16670227050781, 81.65000915527344, 93.22301483154297, -59.614471435546875, -99.55975341796875, 16.755516052246094, 299.62237548828125, 427.2950134277344, 73.41085052490234, 59.712425231933594, 1019.568603515625, 324.3818054199219, 43.798397064208984, 273.541259765625, 270.0330505371094, -22.47865104675293, 3.4859066009521484, 200.5444793701172, 228.0324249267578, -103.45870971679688, -42.462440490722656, 274.9560546875, -56.65751647949219, 287.18115234375, 155.7301025390625, 152.04421997070312, 551.45263671875, 266.8060302734375, 13.739259719848633, 157.34869384765625, 119.84667205810547, -13.561882019042969, 137.9616241455078, 76.8373031616211, 566.5147094726562, 510.8538818359375, 300.2677001953125, -27.3536376953125, -198.34512329101562, 138.81161499023438, -265.2970275878906, 484.73919677734375, 440.7243957519531, 60.796226501464844, -75.39300537109375, 31.2159423828125, -97.59961700439453, 41.873748779296875, 245.44418334960938, -228.54885864257812, 47.45728302001953, 268.88201904296875, -47.78230285644531, 17.62313461303711], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000305.npy"}
{"epoch": 0.447870778267254, "step": 306, "batch_size": 64, "mean": 156.1304931640625, "std": 192.00767517089844, "min": -357.5015869140625, "p10": -37.59999542236327, "median": 120.25780868530273, "p90": 411.9909606933594, "max": 872.349609375, "pos_frac": 0.8125, "sample": [4.498950958251953, -45.4621696472168, 90.72348022460938, 31.236228942871094, 6.58544921875, -65.45967102050781, 267.28167724609375, 33.08509063720703, 79.78619384765625, -26.09272003173828, 224.86790466308594, 351.3678894042969, 145.052490234375, 419.57086181640625, 46.52406311035156, 73.35065460205078, -4.376461029052734, 457.3199768066406, 51.64995574951172, 69.52577209472656, 78.24425506591797, 17.130836486816406, -9.37921142578125, 403.4070739746094, 292.3580627441406, 289.3414611816406, 7.625942230224609, -7.40155029296875, 415.6697692871094, 131.03529357910156, 124.41763305664062, 147.73919677734375, 256.42401123046875, 364.012451171875, 61.315155029296875, 171.62855529785156, 116.09798431396484, -70.04918670654297, 32.04240036010742, 263.6358642578125, -357.5015869140625, 292.00518798828125, 96.11072540283203, -76.45454406738281, 0.5265064239501953, 279.8278503417969, -42.53168487548828, 153.15652465820312, 546.1016845703125, 445.1732177734375, 206.54989624023438, 322.5223388671875, 240.2174530029297, -62.652671813964844, 425.5022277832031, -15.234333038330078, 57.43991470336914, 9.514366149902344, 252.21234130859375, 155.94842529296875, 872.349609375, 254.082763671875, 283.53521728515625, 357.6186218261719], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000306.npy"}
{"epoch": 0.44933920704845814, "step": 307, "batch_size": 64, "mean": 159.51007080078125, "std": 203.6453857421875, "min": -426.2929992675781, "p10": -71.72502975463865, "median": 102.11538696289062, "p90": 437.606332397461, "max": 748.7535400390625, "pos_frac": 0.8125, "sample": [372.61444091796875, 100.13359832763672, 133.35140991210938, 186.65045166015625, 177.12564086914062, 130.1896209716797, -124.2603759765625, 50.47589874267578, -77.26667785644531, 506.8374328613281, 90.9217529296875, 12.375862121582031, 94.9615478515625, -111.76406860351562, 383.20513916015625, -426.2929992675781, 538.9320068359375, -125.38801574707031, 234.57723999023438, -58.794517517089844, 202.50192260742188, 304.7064514160156, 77.88274383544922, 154.93785095214844, 495.14910888671875, 265.6943054199219, 91.00944519042969, 100.3011474609375, 103.92962646484375, 503.3353576660156, 91.16783142089844, 448.8238220214844, 57.85334014892578, 75.31547546386719, -29.481948852539062, 90.6744384765625, -25.496551513671875, 85.86643981933594, 403.4097900390625, 411.43218994140625, -13.916007995605469, 305.5000305175781, 13.859031677246094, 266.4849853515625, 261.51055908203125, 69.48249816894531, 242.62071228027344, 280.07147216796875, 22.9714298248291, 83.51191711425781, 242.36863708496094, 14.545711517333984, 341.41888427734375, 560.3071899414062, 264.1045837402344, 748.7535400390625, 93.14476776123047, -101.86296081542969, 239.40480041503906, 111.64801025390625, -54.35395812988281, 295.45513916015625, 45.94813537597656, -121.93285369873047], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000307.npy"}
{"epoch": 0.45080763582966227, "step": 308, "batch_size": 64, "mean": 180.3531036376953, "std": 190.30332946777344, "min": -198.6342010498047, "p10": -55.3882270812988, "median": 165.0775604248047, "p90": 473.99103088378916, "max": 567.4791870117188, "pos_frac": 0.828125, "sample": [167.72732543945312, 400.5205078125, -198.6342010498047, 204.1484375, 31.752700805664062, 567.4791870117188, -186.14144897460938, -83.65167236328125, 0.32806396484375, 121.41663360595703, 444.1683349609375, -8.63515853881836, 208.9046630859375, 110.55917358398438, 268.21246337890625, 486.7721862792969, -107.83876037597656, 144.4207305908203, 66.80648040771484, 19.702369689941406, 190.5363006591797, 502.9247741699219, 338.70916748046875, 497.3421630859375, -65.76943969726562, 295.302978515625, 134.15576171875, 139.22007751464844, 119.32379150390625, -6.109907150268555, 267.10211181640625, 139.19781494140625, 162.42779541015625, 229.24940490722656, 218.43873596191406, 439.7378845214844, 537.7973022460938, 105.33480834960938, -92.89353942871094, 67.22050476074219, 83.77421569824219, 350.205322265625, 309.82415771484375, 208.94747924804688, 149.18357849121094, 515.3609619140625, 186.35189819335938, 49.72108459472656, 29.58733367919922, 334.63409423828125, 541.096923828125, 259.10498046875, 179.30126953125, 10.316438674926758, -28.943178176879883, 73.00469970703125, 353.3930969238281, 343.64483642578125, 272.72515869140625, -31.16539764404297, 389.3741455078125, 198.93319702148438, -129.98260498046875, 16.93859100341797], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000308.npy"}
{"epoch": 0.4522760646108664, "step": 309, "batch_size": 64, "mean": 188.28118896484375, "std": 216.8057098388672, "min": -529.2852172851562, "p10": -54.60814361572265, "median": 168.54273986816406, "p90": 462.49720153808596, "max": 709.6312255859375, "pos_frac": 0.828125, "sample": [172.96798706054688, 87.87237548828125, 100.68172454833984, -56.3316650390625, 270.2435302734375, 160.7906494140625, 115.30406951904297, -17.727527618408203, 506.178466796875, 7.726837158203125, 280.06475830078125, 397.2574462890625, 116.11363983154297, 503.8432312011719, 45.78931427001953, 52.888877868652344, -66.98008728027344, 322.3071594238281, 466.0339660644531, 350.6182556152344, 251.75035095214844, -529.2852172851562, 254.30035400390625, -90.56996154785156, 56.957977294921875, 142.40481567382812, 244.6768341064453, 282.58160400390625, 75.83988952636719, 30.242982864379883, -121.21399688720703, 629.67041015625, 108.9617919921875, 43.36778259277344, 164.11749267578125, 454.2447509765625, 709.6312255859375, 427.083984375, 145.34860229492188, -21.918106079101562, -11.262386322021484, 81.26722717285156, 65.59840393066406, 418.09521484375, 333.359375, 306.5115051269531, 477.77496337890625, 398.582275390625, 333.91729736328125, 2.874889373779297, 429.5797424316406, -190.63063049316406, 245.53598022460938, 292.0232849121094, -110.08098602294922, 57.41060256958008, 480.4474792480469, 423.739013671875, -50.58659362792969, 179.63565063476562, 197.89503479003906, 21.07266616821289, 386.16168212890625, 207.23780822753906], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000309.npy"}
{"epoch": 0.45374449339207046, "step": 310, "batch_size": 64, "mean": 172.1704864501953, "std": 190.66905212402344, "min": -271.911376953125, "p10": -41.45940628051757, "median": 163.60421752929688, "p90": 428.87709045410156, "max": 576.6652221679688, "pos_frac": 0.796875, "sample": [-95.71824645996094, 171.24819946289062, 392.66497802734375, -32.331687927246094, 146.0812225341797, 345.4319152832031, 43.69972610473633, -9.982000350952148, 37.06623840332031, 193.57327270507812, -203.10430908203125, 423.11920166015625, 431.3447570800781, -25.601459503173828, 48.126319885253906, 262.68109130859375, 203.3931884765625, 247.40476989746094, 74.7623291015625, 160.83575439453125, -19.086872100830078, 105.85171508789062, 365.20025634765625, 77.14126586914062, 133.12786865234375, 562.0148315429688, 327.79443359375, -5.994289398193359, 36.61805725097656, 233.68460083007812, -271.911376953125, 576.6652221679688, -45.12663650512695, 108.30322265625, 544.4385375976562, 206.46066284179688, 295.657470703125, -44.00825500488281, -35.51209259033203, 7.393424987792969, 93.44296264648438, 166.3726806640625, 339.19512939453125, 279.63800048828125, 13.685859680175781, 307.84619140625, 445.10211181640625, -90.32782745361328, 378.3116455078125, 218.02206420898438, 287.6399230957031, -77.24231719970703, 276.456298828125, 115.62071228027344, 31.743515014648438, 61.625038146972656, 72.98779296875, 531.8958740234375, 526.3271484375, 223.43345642089844, 70.20965576171875, 335.02099609375, 269.47650146484375, 169.01971435546875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000310.npy"}
{"epoch": 0.4552129221732746, "step": 311, "batch_size": 64, "mean": 194.36155700683594, "std": 249.73983764648438, "min": -227.40623474121094, "p10": -106.82852783203123, "median": 160.74566650390625, "p90": 533.5041259765625, "max": 870.255859375, "pos_frac": 0.75, "sample": [445.27691650390625, 407.6166687011719, 408.81500244140625, -30.032188415527344, -167.683837890625, 527.6138916015625, -123.938232421875, 272.7977294921875, 189.7836456298828, 529.209716796875, 293.67974853515625, 110.91984558105469, 354.3914489746094, 747.5083618164062, 304.00946044921875, 104.01087188720703, 444.68035888671875, 165.32493591308594, -7.374092102050781, 577.18408203125, 184.17022705078125, 228.4669189453125, 531.2052001953125, 215.54774475097656, 23.17491912841797, 148.49757385253906, -63.32832336425781, -174.87191772460938, 215.07342529296875, 86.08281707763672, -84.0367431640625, 301.9765930175781, -149.1016845703125, 156.16639709472656, 256.3592529296875, -116.596435546875, -8.218887329101562, 870.255859375, 0.31854248046875, 257.5729064941406, 21.414161682128906, 107.23426055908203, 239.87387084960938, 22.937299728393555, 620.177490234375, 396.3167419433594, 33.13161087036133, -49.68316650390625, 18.980792999267578, 95.73793029785156, -22.34872817993164, 542.153564453125, 131.45062255859375, -1.6627273559570312, -227.40623474121094, 460.8946533203125, 153.07516479492188, 228.009521484375, 10.065391540527344, 238.9618377685547, 534.4893798828125, -197.56935119628906, -47.02301788330078, 697.419921875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000311.npy"}
{"epoch": 0.4566813509544787, "step": 312, "batch_size": 64, "mean": 182.33462524414062, "std": 257.72406005859375, "min": -362.12310791015625, "p10": -86.53764038085934, "median": 141.5223846435547, "p90": 509.42380981445314, "max": 1076.7845458984375, "pos_frac": 0.8125, "sample": [124.07371520996094, 1076.7845458984375, 509.96343994140625, 191.20199584960938, 141.8014678955078, 95.22486114501953, -50.353797912597656, 102.66900634765625, 100.29871368408203, 350.09002685546875, -249.57791137695312, 403.2472229003906, 929.584228515625, 342.1400146484375, 233.91726684570312, 229.36236572265625, -188.23541259765625, 256.27923583984375, 389.7598876953125, 177.49996948242188, 540.2463989257812, 84.79341125488281, 130.9010772705078, 276.06854248046875, 62.52490234375, 18.643022537231445, 39.914390563964844, 715.445068359375, 358.9877014160156, 538.0161743164062, 137.28073120117188, 424.5389709472656, 176.09698486328125, 105.63041687011719, -146.93182373046875, 141.24330139160156, 175.2729949951172, 508.1646728515625, -9.623123168945312, 37.15269088745117, 304.71405029296875, 110.2713851928711, -46.390472412109375, 263.4769287109375, 158.7555389404297, 14.07177734375, -28.249298095703125, 650.257080078125, -99.69923400878906, 37.11457061767578, 215.558837890625, 147.3823699951172, -239.9830322265625, -55.82725524902344, -123.23091888427734, 125.59434509277344, 144.62477111816406, 27.026046752929688, -362.12310791015625, 298.5184020996094, 245.67172241210938, 16.274337768554688, 20.765705108642578, 364.744384765625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000312.npy"}
{"epoch": 0.4581497797356828, "step": 313, "batch_size": 64, "mean": 188.029541015625, "std": 263.4723205566406, "min": -470.00787353515625, "p10": -100.81963043212889, "median": 170.8473358154297, "p90": 488.8036590576172, "max": 855.4716796875, "pos_frac": 0.765625, "sample": [62.71771240234375, 307.7505187988281, 105.64060974121094, 301.46600341796875, -151.22830200195312, -23.537124633789062, 406.78228759765625, 161.75650024414062, 397.2517395019531, -46.53937911987305, 13.899307250976562, -179.18275451660156, 459.6439208984375, 169.7377166748047, 22.153554916381836, 220.97019958496094, 493.5712585449219, 189.1551055908203, 69.1654052734375, 225.50704956054688, 46.41107940673828, 327.1004333496094, -470.00787353515625, -109.99595642089844, 217.136962890625, 582.3521728515625, 263.28753662109375, 77.92787170410156, 231.53884887695312, 855.4716796875, 352.255859375, 114.35302734375, -16.481809616088867, 171.9569549560547, -225.34725952148438, 117.36671447753906, -79.408203125, -428.9019775390625, 428.0720520019531, 270.76861572265625, -12.421379089355469, 255.24359130859375, 58.59660720825195, 366.6822814941406, 300.3855285644531, 106.2708511352539, -8.400108337402344, 321.1138916015625, 294.54742431640625, -151.34495544433594, 477.67926025390625, 107.813720703125, 145.73825073242188, -0.4450569152832031, 189.68792724609375, 616.2908935546875, 633.0257568359375, -11.011245727539062, 16.623580932617188, 79.19316101074219, 285.854248046875, 363.48748779296875, 852.1612548828125, 814.5799560546875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000313.npy"}
{"epoch": 0.45961820851688695, "step": 314, "batch_size": 64, "mean": 176.62509155273438, "std": 279.3013610839844, "min": -350.9794921875, "p10": -154.40374298095702, "median": 121.91443634033203, "p90": 514.8562896728517, "max": 976.40771484375, "pos_frac": 0.71875, "sample": [762.42431640625, -263.783447265625, 104.07137298583984, 27.87615203857422, 34.927825927734375, 243.63851928710938, 20.2265625, 352.8273010253906, 69.55485534667969, -131.16517639160156, 319.860595703125, 16.823257446289062, 253.71746826171875, 55.84770202636719, 174.22268676757812, -350.9794921875, -173.20664978027344, 18.064468383789062, 737.2991333007812, 771.7003784179688, 33.33213806152344, -25.177698135375977, -161.72769165039062, 109.18711853027344, 188.61669921875, -168.40782165527344, -26.296127319335938, -132.6261749267578, -57.37084197998047, 412.3706359863281, 447.2913818359375, 40.0673828125, 273.36773681640625, 531.0903930664062, 300.28790283203125, 170.12754821777344, -112.51384735107422, 134.64175415039062, 25.776123046875, 790.778564453125, 476.9767150878906, 332.69512939453125, 976.40771484375, 461.954345703125, -71.1390380859375, 377.2506103515625, -182.32870483398438, 460.73291015625, -27.073455810546875, 250.92080688476562, 308.0294189453125, 61.724395751953125, 339.02337646484375, 340.5792541503906, 402.5361022949219, 531.353515625, 76.46058654785156, 206.613525390625, 225.73179626464844, -12.776802062988281, -137.3145294189453, -33.09883499145508, 319.5840148925781, -197.5992889404297], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000314.npy"}
{"epoch": 0.461086637298091, "step": 315, "batch_size": 64, "mean": 163.78216552734375, "std": 211.92840576171875, "min": -178.5395050048828, "p10": -72.59332885742187, "median": 137.0259017944336, "p90": 498.8488677978517, "max": 690.9887084960938, "pos_frac": 0.75, "sample": [634.6129150390625, -10.021291732788086, 292.04473876953125, -4.093292236328125, -33.07645797729492, 76.70401000976562, 53.07444763183594, 148.8510284423828, 280.80072021484375, -30.82354736328125, 300.171875, 21.270540237426758, -86.69374084472656, 250.40673828125, 84.40142059326172, 519.638427734375, 375.087646484375, 169.05679321289062, -149.32992553710938, 340.9930725097656, 248.6395263671875, 512.8500366210938, 521.1489868164062, 26.947265625, -134.33868408203125, 196.06640625, 205.05419921875, 163.65933227539062, 396.4337463378906, 147.8971405029297, 466.1794738769531, 115.22846984863281, 110.50469970703125, 136.7992706298828, 209.67251586914062, 333.4289855957031, -148.7681884765625, -13.8707275390625, 218.73460388183594, 20.299331665039062, 63.76726150512695, -62.785125732421875, 690.9887084960938, 40.851806640625, 217.84976196289062, -14.756126403808594, -155.83309936523438, 7.873445510864258, -38.64674377441406, 560.693603515625, 111.98660278320312, 215.51829528808594, 611.5235595703125, 137.25253295898438, 140.8887176513672, -76.79684448242188, 293.4589538574219, 36.15311050415039, 403.8733825683594, 461.8783264160156, -178.5395050048828, -53.63252258300781, 64.57173156738281, 38.276206970214844], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000315.npy"}
{"epoch": 0.46255506607929514, "step": 316, "batch_size": 64, "mean": 229.9477081298828, "std": 262.9196472167969, "min": -251.00921630859375, "p10": -22.551457595825184, "median": 164.60208892822266, "p90": 609.6040832519532, "max": 946.818359375, "pos_frac": 0.875, "sample": [360.153076171875, 106.82963562011719, 501.38623046875, 46.07319641113281, 46.878562927246094, 946.818359375, 161.93374633789062, 682.6895141601562, 484.24078369140625, 617.218505859375, 362.4810791015625, 325.38189697265625, 45.486053466796875, 23.67807388305664, 30.375518798828125, 224.43104553222656, 86.34193420410156, 106.88212585449219, 21.156734466552734, 417.76904296875, 180.63705444335938, 22.173465728759766, 637.5213623046875, 425.148681640625, 268.3177795410156, 94.75675201416016, 31.341922760009766, 675.5030517578125, 147.30226135253906, 500.35443115234375, 33.51239776611328, 89.74480438232422, 51.529876708984375, 141.09971618652344, 207.38809204101562, 347.9508056640625, 415.816650390625, 65.59969329833984, 390.24151611328125, 26.69342803955078, 591.8370971679688, 326.35992431640625, 35.764617919921875, 394.9868469238281, 205.91635131835938, 149.56069946289062, -244.11805725097656, -27.554950714111328, 920.0856323242188, 167.2704315185547, 932.3897705078125, -251.00921630859375, 183.01742553710938, 178.41717529296875, 366.15020751953125, -10.876640319824219, -65.9431381225586, -59.10337829589844, -29.079124450683594, 80.4586410522461, 212.85012817382812, -151.45880126953125, 106.077392578125, 353.8149719238281], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000316.npy"}
{"epoch": 0.46402349486049926, "step": 317, "batch_size": 64, "mean": 254.23016357421875, "std": 245.82586669921875, "min": -157.82156372070312, "p10": -31.49203243255615, "median": 223.19707489013672, "p90": 570.1170654296875, "max": 807.2319946289062, "pos_frac": 0.8125, "sample": [331.6177978515625, 16.11725616455078, 102.32584381103516, 475.6669006347656, 452.9862060546875, 104.49642944335938, 451.82244873046875, 416.7850341796875, -157.82156372070312, 573.4114990234375, 728.9720458984375, 143.0185089111328, 252.7801513671875, 22.01055908203125, 145.96768188476562, 385.31866455078125, 193.7584686279297, -4.181938171386719, 96.1935806274414, 399.46624755859375, -32.734375, 355.3455810546875, 366.6194152832031, 556.8270263671875, 29.446514129638672, 558.9112548828125, -16.22555160522461, 102.3974380493164, 737.5796508789062, 419.605712890625, 631.0354614257812, 488.3778076171875, 214.6092071533203, 562.4300537109375, 407.28289794921875, -66.88685607910156, 249.69163513183594, 133.0117950439453, 280.3222961425781, 120.1851806640625, 211.01666259765625, 556.1282958984375, -23.802698135375977, -28.593233108520508, 807.2319946289062, 215.5048065185547, 622.9727783203125, 326.87872314453125, 230.88934326171875, -128.9656982421875, 99.68162536621094, 91.78308868408203, 251.88429260253906, -4.284111022949219, -95.96062469482422, 117.56436920166016, 89.06704711914062, -155.591552734375, 667.99609375, 419.94976806640625, 91.29938507080078, 445.488525390625, -153.80506896972656, 387.8533630371094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000317.npy"}
{"epoch": 0.4654919236417034, "step": 318, "batch_size": 64, "mean": 186.30563354492188, "std": 245.23635864257812, "min": -358.6517333984375, "p10": -87.53584747314453, "median": 158.58511352539062, "p90": 502.32821044921883, "max": 881.1334228515625, "pos_frac": 0.796875, "sample": [51.420509338378906, -120.50709533691406, 881.1334228515625, 106.54632568359375, 240.42721557617188, 109.41319274902344, -72.18103790283203, 171.69979858398438, 139.3013916015625, 662.0151977539062, 21.013612747192383, 147.41082763671875, 354.18017578125, 486.09063720703125, 622.0704956054688, 233.73423767089844, -18.043869018554688, 131.01300048828125, -124.85411834716797, -88.3517837524414, 598.6336669921875, 67.1567611694336, -134.4108123779297, 509.28717041015625, 835.9159545898438, 322.68157958984375, -38.0730094909668, -100.66990661621094, 225.27635192871094, 362.5252990722656, 68.78384399414062, 368.1368103027344, 56.18230438232422, 413.4674987792969, 224.33827209472656, 132.04681396484375, 553.4669189453125, 20.322612762451172, -64.85049438476562, 435.28155517578125, 431.7608642578125, 206.535888671875, 11.143115997314453, -85.63199615478516, 382.08404541015625, 212.9498291015625, 130.87661743164062, 192.221923828125, 153.2574005126953, -61.431915283203125, -331.13232421875, 51.922950744628906, 194.72409057617188, 194.66009521484375, 361.3500671386719, 373.8814697265625, 210.99005126953125, 155.74545288085938, 57.93861389160156, 299.1669006347656, 161.42477416992188, 182.84414672851562, -358.6517333984375, 5.899009704589844], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000318.npy"}
{"epoch": 0.4669603524229075, "step": 319, "batch_size": 64, "mean": 226.8285369873047, "std": 266.0675354003906, "min": -371.3077392578125, "p10": -119.14757080078125, "median": 214.22503662109375, "p90": 552.8462097167969, "max": 1042.9033203125, "pos_frac": 0.8125, "sample": [307.1232604980469, 125.76008605957031, 336.7287292480469, 242.770263671875, 28.978713989257812, 290.6366271972656, 354.2691345214844, 553.2940673828125, 236.16250610351562, 84.97650909423828, -83.83165740966797, -261.7403869628906, 355.1036682128906, 124.47587585449219, 328.8241271972656, 181.1525421142578, 69.6939468383789, 267.64697265625, 297.995361328125, 139.06314086914062, 189.6302947998047, 512.1032104492188, 206.91293334960938, 269.69140625, 297.1602783203125, 3.83624267578125, 341.1520080566406, 834.4122314453125, 407.8407287597656, 373.03948974609375, -120.26998901367188, 182.29237365722656, 180.5768280029297, 551.8012084960938, -314.65191650390625, 425.4326171875, -371.3077392578125, -129.63772583007812, -116.52859497070312, 402.182373046875, 616.427978515625, 450.702880859375, 632.1848754882812, -19.299619674682617, 186.78184509277344, 56.920806884765625, 579.775390625, 581.5020751953125, 281.7735290527344, 545.3626708984375, 1042.9033203125, -153.1315460205078, 149.96177673339844, 445.750244140625, -12.905963897705078, 37.37358856201172, -243.90655517578125, -49.88452911376953, 431.0745849609375, 150.0979461669922, 146.44708251953125, 180.31597900390625, 221.53713989257812, 154.50958251953125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000319.npy"}
{"epoch": 0.4684287812041116, "step": 320, "batch_size": 64, "mean": 250.36404418945312, "std": 274.046630859375, "min": -318.25823974609375, "p10": -73.07018966674805, "median": 228.75359344482422, "p90": 562.7494201660156, "max": 1095.3519287109375, "pos_frac": 0.78125, "sample": [186.10919189453125, 184.2982177734375, 923.0845947265625, 128.32444763183594, 304.6636657714844, -134.16195678710938, 194.702392578125, -318.25823974609375, 848.9923706054688, 222.3607940673828, 458.5753173828125, 386.1118469238281, -39.64647674560547, 385.4319152832031, 187.80349731445312, 566.2217407226562, 631.7671508789062, 423.8840026855469, 389.04132080078125, -153.98110961914062, 368.1433410644531, 368.449951171875, -47.82457733154297, 212.62493896484375, 79.42176055908203, -37.89173126220703, 323.4783935546875, 324.6957092285156, -11.540115356445312, 554.6473388671875, -69.48497009277344, 255.65164184570312, -50.635841369628906, 117.40001678466797, 496.0357666015625, 151.77207946777344, 200.89346313476562, 235.14639282226562, 1095.3519287109375, 395.7718811035156, 413.8883056640625, 144.20626831054688, 91.76213073730469, -88.25822448730469, 660.7025756835938, 59.693260192871094, 420.794189453125, 219.9279327392578, -10.98505973815918, 22.473129272460938, 366.96966552734375, 437.33026123046875, 301.55657958984375, 432.7042236328125, 526.9501953125, 624.3232421875, 508.85626220703125, 28.055845260620117, -150.4193572998047, -74.6067123413086, 47.641204833984375, 242.76548767089844, 244.8724365234375, -185.33700561523438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000320.npy"}
{"epoch": 0.4698972099853157, "step": 321, "batch_size": 64, "mean": 210.19265747070312, "std": 287.0802917480469, "min": -382.10064697265625, "p10": -124.89369964599608, "median": 196.7539520263672, "p90": 537.6637756347657, "max": 1097.8570556640625, "pos_frac": 0.78125, "sample": [727.4830932617188, 124.09465789794922, 337.32269287109375, 267.9765319824219, -139.81454467773438, 285.4106140136719, 342.3778076171875, 326.4598693847656, 196.40625, 501.58062744140625, 396.0833435058594, 396.65838623046875, 89.93478393554688, 7.485523223876953, -231.47398376464844, 274.96728515625, 446.51910400390625, 251.92807006835938, 210.21408081054688, 521.2789306640625, 872.7155151367188, 161.42440795898438, -382.10064697265625, 313.3880310058594, 226.0664520263672, 320.89324951171875, 757.1285400390625, -284.40948486328125, -109.33238220214844, 141.36988830566406, 53.64833068847656, 544.6858520507812, -92.14376831054688, 359.56597900390625, 180.14584350585938, 442.8738098144531, 261.92962646484375, -253.91470336914062, 1097.8570556640625, 106.75675201416016, -73.17355346679688, 67.01712799072266, 6.472101211547852, 106.98725891113281, -14.023262023925781, 108.33047485351562, -108.0372314453125, -30.043140411376953, 146.06936645507812, 242.8848114013672, 745.2046508789062, 659.5806884765625, -15.311626434326172, -320.0997009277344, 211.08966064453125, 197.10165405273438, 154.40597534179688, 174.03480529785156, 8.255403518676758, 518.4984130859375, 172.59893798828125, 309.7396545410156, 264.86981201171875, -131.56283569335938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000321.npy"}
{"epoch": 0.4713656387665198, "step": 322, "batch_size": 64, "mean": 231.65281677246094, "std": 243.9152374267578, "min": -315.3796691894531, "p10": -54.15164337158202, "median": 260.5352783203125, "p90": 545.3587585449219, "max": 808.6083984375, "pos_frac": 0.78125, "sample": [-186.15478515625, 808.6083984375, 306.628173828125, 775.672119140625, 92.49938201904297, -28.029300689697266, 539.6051635742188, 393.2556457519531, 311.846923828125, -84.80735778808594, -28.353775024414062, -40.746604919433594, 318.82879638671875, 281.0895080566406, 364.599609375, 382.5447082519531, -59.89665985107422, 243.45260620117188, -22.443580627441406, -125.98782348632812, 659.626708984375, 579.0009765625, 245.98220825195312, 578.058837890625, -1.935302734375, 226.14700317382812, 397.32269287109375, 524.4927368164062, 36.85095977783203, 547.8245849609375, -87.49873352050781, 223.61532592773438, -20.319011688232422, 369.1622314453125, -28.363983154296875, 8.486919403076172, 277.60113525390625, 180.97744750976562, 0.3559761047363281, 305.06024169921875, 480.75799560546875, 466.0938720703125, 377.36602783203125, -118.99605560302734, 451.30450439453125, 117.84193420410156, 322.9715576171875, 3.107624053955078, 259.7619323730469, 80.01356506347656, 118.13825225830078, 671.9845581054688, 262.1266174316406, 327.928466796875, 379.7699890136719, 385.11212158203125, 88.20394897460938, 93.77710723876953, 76.87679290771484, -315.3796691894531, 22.419708251953125, 342.5965576171875, 406.0345458984375, 261.3086242675781], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000322.npy"}
{"epoch": 0.47283406754772395, "step": 323, "batch_size": 64, "mean": 194.42291259765625, "std": 257.8011474609375, "min": -402.9685363769531, "p10": -99.35482711791992, "median": 200.08599853515625, "p90": 553.6526245117187, "max": 856.0299682617188, "pos_frac": 0.796875, "sample": [371.20562744140625, 244.14447021484375, 597.7575073242188, -402.9685363769531, 629.7362670898438, -321.1958923339844, 613.9656372070312, 856.0299682617188, 64.42973327636719, 77.29705047607422, 235.552001953125, 47.83538818359375, 188.91830444335938, 123.4024887084961, 14.698661804199219, 348.74273681640625, 97.7619857788086, 260.5168151855469, -141.76629638671875, 647.8539428710938, -17.090858459472656, 72.58473205566406, 519.7896118164062, -123.12874603271484, -99.15339660644531, 320.8451232910156, 79.69889068603516, 145.634033203125, 440.9847412109375, -5.02618408203125, 555.357666015625, 272.786865234375, 504.44378662109375, -60.216094970703125, 340.91180419921875, 22.697769165039062, 30.489425659179688, 218.55264282226562, 329.7027587890625, 238.80111694335938, 344.4729919433594, 71.56033325195312, 282.22943115234375, 520.1857299804688, -344.46466064453125, 148.9796905517578, 331.82806396484375, 617.1612548828125, 362.6293640136719, -63.603294372558594, 211.25369262695312, 549.6741943359375, -99.44115447998047, 305.4865417480469, -253.6865997314453, 220.77981567382812, 112.65960693359375, 223.86663818359375, 47.393333435058594, 310.84881591796875, 47.85857391357422, -8.653003692626953, 147.36782836914062, 14.0965576171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000323.npy"}
{"epoch": 0.47430249632892807, "step": 324, "batch_size": 64, "mean": 190.73269653320312, "std": 285.35772705078125, "min": -410.3531494140625, "p10": -139.81797027587888, "median": 158.45526123046875, "p90": 546.6569000244143, "max": 967.022705078125, "pos_frac": 0.765625, "sample": [-199.36239624023438, 611.3944702148438, 11.59197998046875, 192.55450439453125, 57.692405700683594, 165.66326904296875, 242.61578369140625, 282.897216796875, 88.29069519042969, 202.79132080078125, 842.290283203125, -46.579803466796875, 100.08572387695312, 108.67025756835938, -40.38325881958008, 695.8131103515625, -195.5752410888672, -410.3531494140625, 315.098388671875, 270.1871032714844, 25.5032958984375, 151.24725341796875, -186.27728271484375, 231.52029418945312, 319.8809814453125, 894.2649536132812, 427.08905029296875, -119.87043762207031, 104.70745849609375, 496.3399353027344, 483.6145935058594, 140.0627899169922, 112.19316864013672, 130.98321533203125, 358.4126281738281, 301.9688720703125, 176.8694610595703, 334.5966491699219, 224.20285034179688, 18.41840171813965, 435.26483154296875, -401.81427001953125, -90.96228790283203, 407.2102966308594, 568.2213134765625, 410.802734375, 143.16668701171875, -363.2565612792969, -148.36691284179688, 95.71759796142578, 281.5112609863281, 75.34581756591797, 261.93243408203125, -50.82485580444336, 127.03221130371094, 40.974945068359375, 355.9732666015625, -33.581939697265625, -62.712894439697266, 967.022705078125, 685.6358642578125, 240.5209503173828, -29.929458618164062, 370.8975830078125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000324.npy"}
{"epoch": 0.47577092511013214, "step": 325, "batch_size": 64, "mean": 160.074951171875, "std": 226.30267333984375, "min": -378.3366394042969, "p10": -81.95445251464844, "median": 130.7279815673828, "p90": 434.3146026611328, "max": 864.7681884765625, "pos_frac": 0.828125, "sample": [-364.8695983886719, -378.3366394042969, 245.76446533203125, -36.76227569580078, 319.0888977050781, 280.9888916015625, 109.04926300048828, -76.70381164550781, 167.92611694335938, 326.42205810546875, 308.16680908203125, 497.64794921875, 303.84259033203125, 137.3098907470703, 110.97126770019531, 105.1451187133789, 8.992561340332031, 99.18751525878906, 50.48371887207031, 247.75408935546875, -149.16531372070312, 284.22698974609375, 283.6888122558594, 567.2703247070312, 23.74453353881836, 31.435752868652344, 147.13851928710938, 95.67430877685547, 124.14607238769531, 30.711814880371094, 431.77874755859375, -84.20472717285156, 864.7681884765625, 11.714115142822266, 523.0983276367188, 83.44414520263672, 527.8182373046875, -84.88572692871094, 226.94686889648438, 207.28346252441406, -70.3302993774414, 166.1392822265625, 363.2330017089844, -30.152559280395508, 271.7447509765625, 329.5450744628906, 268.8428039550781, 265.46746826171875, 105.05923461914062, 139.28823852539062, 358.2969665527344, -344.681640625, 23.362159729003906, 663.18994140625, 29.9151611328125, 62.229835510253906, 435.4013977050781, 55.111183166503906, -113.11532592773438, 249.42095947265625, 108.64399719238281, 36.5416145324707, 60.09877014160156, 202.84226989746094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000325.npy"}
{"epoch": 0.47723935389133626, "step": 326, "batch_size": 64, "mean": 156.58657836914062, "std": 195.29005432128906, "min": -410.6053161621094, "p10": -36.35793266296386, "median": 137.02716827392578, "p90": 440.8150817871094, "max": 678.8482666015625, "pos_frac": 0.796875, "sample": [56.081146240234375, 43.797550201416016, 91.65684509277344, 444.406494140625, 222.14195251464844, -9.27703857421875, 182.60525512695312, 127.06491088867188, 315.7584533691406, -228.88027954101562, -110.01007080078125, -410.6053161621094, 499.42010498046875, 97.02318572998047, -150.8325958251953, 157.04701232910156, 210.53724670410156, 55.55099105834961, 301.85400390625, 252.70391845703125, -3.814666748046875, 32.299041748046875, 208.24998474121094, 252.1015625, -27.797958374023438, 137.9586639404297, 678.8482666015625, 204.0128631591797, 94.35357666015625, 122.05899047851562, 365.170166015625, 201.93603515625, 249.32994079589844, 136.09567260742188, 173.35125732421875, -7.401611328125, 58.90185546875, 464.0670166015625, -40.026493072509766, 48.999114990234375, 112.59999084472656, -43.25408935546875, 67.74333190917969, 168.82260131835938, 40.321197509765625, 89.23838806152344, 309.5694580078125, 210.36422729492188, 222.40428161621094, 252.35096740722656, -27.702590942382812, 452.9344177246094, 580.8555908203125, 85.04357147216797, 348.51068115234375, 186.17300415039062, 432.43511962890625, 331.27142333984375, -8.032342910766602, -226.10289001464844, 468.41546630859375, 313.46783447265625, 48.09614944458008, 109.27733612060547], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000326.npy"}
{"epoch": 0.4787077826725404, "step": 327, "batch_size": 64, "mean": 226.25048828125, "std": 228.50222778320312, "min": -301.54718017578125, "p10": -96.98240356445312, "median": 243.63195037841797, "p90": 508.8961181640625, "max": 745.5933227539062, "pos_frac": 0.8125, "sample": [200.97763061523438, 209.89903259277344, 502.8052978515625, 215.778076171875, 227.57176208496094, 340.32122802734375, 71.36136627197266, 276.1135559082031, 329.5722961425781, 196.70164489746094, 351.4232177734375, 323.48162841796875, 121.72026062011719, 386.086669921875, -98.82099151611328, 366.1884765625, -34.24758529663086, 407.8216857910156, 357.8125305175781, 301.51849365234375, -224.61541748046875, 618.3914184570312, 552.4227294921875, 390.0469055175781, 23.93244743347168, 511.5064697265625, 304.176025390625, -103.35749816894531, 208.89813232421875, 296.7445068359375, 391.65936279296875, -143.33837890625, -20.855743408203125, 163.08445739746094, 282.12042236328125, 219.05995178222656, 267.86785888671875, 259.692138671875, 267.947265625, 165.91336059570312, -37.51469421386719, 96.40109252929688, -96.71661376953125, 495.2233581542969, 216.9012451171875, -97.0963134765625, 48.18812942504883, 345.99566650390625, 406.8780517578125, 71.43255615234375, 672.0851440429688, 112.98155212402344, 745.5933227539062, -301.54718017578125, 474.5899353027344, -18.449220657348633, 71.39823913574219, 572.4534301757812, 11.703536987304688, 429.79736328125, 389.0541687011719, 527.0435180664062, 91.66864776611328, -233.41729736328125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000327.npy"}
{"epoch": 0.4801762114537445, "step": 328, "batch_size": 64, "mean": 202.90028381347656, "std": 251.0277862548828, "min": -401.61334228515625, "p10": -85.47906036376952, "median": 178.10714721679688, "p90": 502.5141387939454, "max": 940.4327392578125, "pos_frac": 0.765625, "sample": [187.6488494873047, -135.11233520507812, 190.55410766601562, 145.11907958984375, 23.46490478515625, 374.90185546875, 346.777587890625, 135.4268035888672, 38.9117431640625, -401.61334228515625, 64.31543731689453, 448.05975341796875, 508.01177978515625, 352.3700256347656, 287.17523193359375, 424.48150634765625, -4.404563903808594, -144.02207946777344, -6.825593948364258, 210.93304443359375, -75.74093627929688, -33.832000732421875, 352.678955078125, 940.4327392578125, 248.23672485351562, 312.422607421875, 367.5096130371094, 181.59402465820312, 686.3491821289062, 392.706298828125, -114.0177001953125, 180.75735473632812, 305.4767150878906, -89.65254211425781, 49.16868591308594, 120.03872680664062, 321.9198913574219, 455.9929504394531, 166.18426513671875, 141.56362915039062, -9.387689590454102, 98.32697296142578, -94.2376708984375, 175.45693969726562, -18.412269592285156, 478.7099609375, 96.18740844726562, 262.5836486816406, 294.9044189453125, 192.67718505859375, 590.21728515625, 489.6863098144531, 74.18055725097656, 14.229888916015625, -12.48354721069336, 751.5648803710938, 799.4921264648438, 65.47549438476562, 555.58642578125, -23.50417709350586, 14.641975402832031, 294.8894958496094, 83.82777404785156, -144.95770263671875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000328.npy"}
{"epoch": 0.48164464023494863, "step": 329, "batch_size": 64, "mean": 226.84713745117188, "std": 272.8593444824219, "min": -647.8804931640625, "p10": -52.25366325378418, "median": 211.3514633178711, "p90": 563.1756347656251, "max": 971.4133911132812, "pos_frac": 0.8125, "sample": [259.6306457519531, 393.10797119140625, 536.3704833984375, 321.07220458984375, 265.3070373535156, -141.77955627441406, 499.8395080566406, 45.136085510253906, 569.084228515625, 249.043701171875, 37.701904296875, 431.6497497558594, 119.7710952758789, -39.51106262207031, 71.54325866699219, 114.03543853759766, 236.1054229736328, 205.69302368164062, 73.89796447753906, 786.3359375, 147.33737182617188, 125.17881774902344, 206.09178161621094, 163.91885375976562, -647.8804931640625, 216.61114501953125, 50.07970428466797, 971.4133911132812, 620.2750244140625, -40.57075500488281, 267.3485412597656, -13.367916107177734, 414.3999938964844, 148.48822021484375, 266.05072021484375, 409.6174621582031, 604.5834350585938, 348.57598876953125, 750.1676635742188, 234.77244567871094, 31.07858657836914, -51.8598518371582, -107.37980651855469, 549.388916015625, 482.9302062988281, -145.28097534179688, 261.3399658203125, -105.54560852050781, 374.5826416015625, 165.85694885253906, 250.4852294921875, -52.42243957519531, 205.4080047607422, 237.3489532470703, 84.19416046142578, 51.981895446777344, 314.035400390625, 479.1700439453125, 749.2861938476562, -26.85089111328125, -252.60757446289062, 185.68521118164062, 526.0296020507812, 34.235633850097656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000329.npy"}
{"epoch": 0.4831130690161527, "step": 330, "batch_size": 64, "mean": 221.03128051757812, "std": 257.9089050292969, "min": -262.09600830078125, "p10": -117.28226089477535, "median": 222.68964385986328, "p90": 492.839569091797, "max": 882.01513671875, "pos_frac": 0.8125, "sample": [466.68621826171875, -35.9927978515625, -149.41375732421875, -132.7428741455078, 319.7552490234375, 469.15924072265625, 168.78941345214844, 224.28375244140625, -81.2074966430664, 123.12300872802734, -142.9993133544922, 108.98471069335938, 159.19461059570312, 460.90985107421875, 502.98828125, 18.94367218017578, 268.15185546875, 287.48602294921875, 255.69876098632812, 427.7328796386719, -25.522417068481445, 84.13035583496094, 323.5679626464844, 43.33158493041992, 145.54180908203125, 334.28363037109375, -262.09600830078125, 345.1932678222656, 233.5458221435547, 198.29373168945312, 420.0757751464844, 13.823795318603516, -255.34170532226562, 221.0955352783203, 529.4736328125, 184.24874877929688, 67.90901184082031, 786.899169921875, 43.724281311035156, 131.16891479492188, -36.025856018066406, 422.27606201171875, 118.48458862304688, 882.01513671875, -212.9388885498047, 815.033935546875, 765.3505859375, 326.3349914550781, 377.72900390625, 78.39613342285156, 303.18896484375, 406.80828857421875, 130.41836547851562, 706.4451904296875, 18.34398651123047, 325.04388427734375, 235.40823364257812, 356.4942626953125, -184.61782836914062, 427.5370788574219, 59.5477294921875, 322.91351318359375, 290.63037109375, -71.69448852539062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000330.npy"}
{"epoch": 0.4845814977973568, "step": 331, "batch_size": 64, "mean": 226.0296630859375, "std": 191.14158630371094, "min": -150.73233032226562, "p10": 23.341040802001967, "median": 198.9734115600586, "p90": 416.4165802001955, "max": 831.281982421875, "pos_frac": 0.90625, "sample": [123.63333129882812, 710.2371826171875, 291.3955383300781, -44.408599853515625, 323.22625732421875, 323.7286376953125, 98.41610717773438, 35.49113464355469, 208.2162628173828, 164.68408203125, 709.8173828125, 249.99453735351562, 831.281982421875, 18.13385772705078, 167.19349670410156, 294.2622375488281, 630.9786987304688, 140.7036590576172, 174.83157348632812, 202.9608154296875, 436.4402160644531, 316.87042236328125, 475.86444091796875, -40.99357223510742, 117.81118774414062, 185.4468536376953, 216.6990203857422, 277.4507141113281, 300.22930908203125, -32.66925048828125, 281.6929626464844, -81.08476257324219, 236.5565185546875, 175.80963134765625, 46.134422302246094, 168.41793823242188, -150.73233032226562, 163.6038818359375, 304.88507080078125, 636.2036743164062, 243.94674682617188, 194.9860076904297, 131.67813110351562, 352.7862548828125, 78.33551788330078, 111.12908935546875, 295.3158264160156, 175.07708740234375, 315.83489990234375, 122.55308532714844, -148.51513671875, 264.2745056152344, 124.38496398925781, 111.72489929199219, 321.7616882324219, 310.83349609375, 334.85626220703125, 97.6600112915039, 157.01780700683594, 369.69476318359375, 144.0555877685547, 265.8581237792969, 359.452880859375, 41.781585693359375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000331.npy"}
{"epoch": 0.48604992657856094, "step": 332, "batch_size": 64, "mean": 166.06466674804688, "std": 224.4556427001953, "min": -158.99774169921875, "p10": -90.11821899414062, "median": 116.0476188659668, "p90": 457.91500549316413, "max": 928.6058349609375, "pos_frac": 0.765625, "sample": [739.7239990234375, 189.063720703125, 133.66685485839844, 338.1615905761719, 39.35557556152344, 7.1893157958984375, 79.03257751464844, 412.76409912109375, 392.5804443359375, 194.29551696777344, -64.50828552246094, 67.86451721191406, -69.862060546875, 928.6058349609375, 186.5674285888672, 116.32976531982422, 115.76547241210938, 152.3639373779297, 126.17081451416016, 2.9489879608154297, 130.81591796875, 83.93052673339844, 468.0796203613281, -67.66755676269531, 387.0120849609375, 31.541156768798828, -158.99774169921875, 95.96675872802734, 609.514892578125, 206.07376098632812, 160.436279296875, 115.10563659667969, 258.9569091796875, -34.936378479003906, -53.79657745361328, 81.52605438232422, 510.24554443359375, 505.5926818847656, -90.70657348632812, 273.11456298828125, 174.70767211914062, -110.97049713134766, 322.740234375, 90.04582214355469, 321.4397277832031, 660.9623413085938, 69.79890441894531, 434.19757080078125, -88.74539184570312, 96.52365112304688, -139.60653686523438, 199.38912963867188, 349.4960021972656, -134.11813354492188, 331.81982421875, 45.259586334228516, -36.56129455566406, -2.453460693359375, -114.29736328125, 187.06161499023438, 253.1402130126953, 110.6928939819336, 107.17753601074219, -99.4493408203125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000332.npy"}
{"epoch": 0.48751835535976507, "step": 333, "batch_size": 64, "mean": 214.71702575683594, "std": 294.2414855957031, "min": -352.0252990722656, "p10": -179.96820983886718, "median": 179.05921173095703, "p90": 579.6980529785156, "max": 901.5574951171875, "pos_frac": 0.75, "sample": [-76.83805847167969, 285.2176208496094, 901.5574951171875, -19.716629028320312, 36.69940948486328, 9.035995483398438, 287.76312255859375, 230.3395538330078, 53.6748046875, 20.145931243896484, 785.2357788085938, -182.33758544921875, 178.8304443359375, -174.43966674804688, 536.2374267578125, -191.31240844726562, 638.3575439453125, -107.41259765625, 620.037109375, 463.2967834472656, 14.34627914428711, 300.4979553222656, -18.627120971679688, 352.13861083984375, 444.6146240234375, 547.731689453125, -219.6807861328125, 130.9189453125, 40.23095703125, 434.2929382324219, 131.84503173828125, -5.032915115356445, -81.16592407226562, 522.661865234375, 541.511962890625, 402.0760192871094, -192.9412841796875, 9.341720581054688, 564.9761352539062, 469.6390075683594, 446.80938720703125, -248.3960723876953, -16.18437957763672, 477.7181701660156, 504.92535400390625, 0.8506088256835938, 175.06285095214844, 179.28797912597656, 733.5903930664062, 277.2212829589844, 495.916259765625, 290.1775207519531, 586.0074462890625, 147.66519165039062, 261.8669128417969, -150.4835662841797, 392.8433837890625, 4.672672271728516, 329.0646667480469, -288.3282470703125, 635.798583984375, -352.0252990722656, 70.76704406738281, 103.31416320800781], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000333.npy"}
{"epoch": 0.4889867841409692, "step": 334, "batch_size": 64, "mean": 197.10379028320312, "std": 301.8833923339844, "min": -515.4450073242188, "p10": -156.36649780273433, "median": 140.7756576538086, "p90": 657.6854919433595, "max": 853.7113037109375, "pos_frac": 0.78125, "sample": [643.7648315429688, 70.9466552734375, -190.857421875, 508.81463623046875, 249.64529418945312, 41.907928466796875, -53.04314422607422, 66.06038665771484, 235.63546752929688, 30.80828857421875, 257.57623291015625, -16.176063537597656, 357.4796142578125, 210.50967407226562, -180.37242126464844, 203.4947967529297, 65.97929382324219, 439.3310546875, -515.4450073242188, 695.1153564453125, 46.99052810668945, 227.43870544433594, -33.617279052734375, -285.53265380859375, 392.43194580078125, 251.58364868164062, 458.1334533691406, -404.61578369140625, -232.1195526123047, -34.54020690917969, 33.947418212890625, -82.39910888671875, 129.9238739013672, 55.771270751953125, 780.3351440429688, 808.5736083984375, 309.8455810546875, 26.84403419494629, 106.71583557128906, 536.6599731445312, 83.6963119506836, 263.5869140625, 518.7882080078125, 400.41522216796875, -100.35267639160156, 85.0216064453125, 853.7113037109375, 118.60073852539062, 759.296630859375, 473.33795166015625, -71.29016876220703, -336.64208984375, 20.1619873046875, 274.8777770996094, 201.85874938964844, 663.6514892578125, 459.4214782714844, 115.99201965332031, 49.17155456542969, 375.77099609375, 115.84210205078125, 757.212158203125, 167.33897399902344, 151.62744140625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000334.npy"}
{"epoch": 0.49045521292217326, "step": 335, "batch_size": 64, "mean": 154.20657348632812, "std": 273.9264831542969, "min": -316.40594482421875, "p10": -183.8399169921875, "median": 131.03733825683594, "p90": 463.6544860839844, "max": 1035.01318359375, "pos_frac": 0.71875, "sample": [166.64508056640625, 443.018310546875, 409.785888671875, 4.777984619140625, 179.6516571044922, -82.1991958618164, 17.078989028930664, 467.61328125, 73.769775390625, 37.27342987060547, -237.26763916015625, 71.54960632324219, 285.3466491699219, 244.24879455566406, -63.68291473388672, 670.6517944335938, 0.5076828002929688, 131.92230224609375, -251.15855407714844, 226.69732666015625, -132.79833984375, 93.62892150878906, 210.51165771484375, 196.52182006835938, 664.5596923828125, 574.0382080078125, 398.9681396484375, 237.9769744873047, 11.735506057739258, 659.5730590820312, 1035.01318359375, -298.801513671875, 114.84725952148438, -316.40594482421875, 424.2974548339844, -47.873023986816406, 442.11407470703125, 53.50645446777344, 72.66314697265625, 547.0603637695312, -141.7384033203125, 329.5400695800781, -143.43325805664062, 73.14501953125, 318.0257568359375, -55.93414306640625, 275.7419128417969, -26.198287963867188, -273.5370788574219, -108.41595458984375, 131.66009521484375, -51.48066711425781, 425.4185791015625, -181.1116943359375, -185.0091552734375, 393.3418273925781, 251.82777404785156, 454.41729736328125, 18.838088989257812, -229.55776977539062, 210.1315460205078, 373.29595947265625, 130.41458129882812, 142.47140502929688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000335.npy"}
{"epoch": 0.4919236417033774, "step": 336, "batch_size": 64, "mean": 247.0634307861328, "std": 352.38372802734375, "min": -703.8919067382812, "p10": -100.76492996215819, "median": 238.02218627929688, "p90": 757.9753967285156, "max": 1181.754150390625, "pos_frac": 0.8125, "sample": [399.91741943359375, 57.51740264892578, 944.85693359375, -5.600744247436523, 464.8152770996094, 240.2386474609375, 262.6301574707031, -105.2038803100586, -90.40737915039062, 894.1959228515625, 285.97137451171875, 332.0711669921875, 176.24588012695312, 502.9799499511719, 1181.754150390625, -77.0758056640625, 764.49609375, 348.4311828613281, 240.36537170410156, 384.9984130859375, -109.36219787597656, 53.648284912109375, 131.4295196533203, 295.14862060546875, 125.4732666015625, 380.95037841796875, 121.52703857421875, 58.27334976196289, 798.5178833007812, 235.80572509765625, 182.3282928466797, 64.89482879638672, 76.88282775878906, 243.02047729492188, 126.48753356933594, -6.22625732421875, -405.482666015625, 742.7604370117188, -79.20089721679688, -259.2465515136719, 725.8353881835938, 19.28687286376953, 880.63916015625, 67.03512573242188, 447.5693359375, 76.96624755859375, -703.8919067382812, 92.3490219116211, 316.64569091796875, 322.0108337402344, 110.06175231933594, 138.8302001953125, 147.6852264404297, 70.2117691040039, 241.35552978515625, 460.81976318359375, 504.75, 1078.369140625, -364.2702941894531, 360.66986083984375, -409.02825927734375, 362.33050537109375, 329.8776550292969, 555.1235961914062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000336.npy"}
{"epoch": 0.4933920704845815, "step": 337, "batch_size": 64, "mean": 238.76651000976562, "std": 277.9673156738281, "min": -270.1160888671875, "p10": -113.70706253051755, "median": 206.38333892822266, "p90": 605.9479003906253, "max": 994.797607421875, "pos_frac": 0.828125, "sample": [940.017333984375, 252.55157470703125, -218.5514678955078, 131.51385498046875, 554.7047729492188, 531.4376220703125, -44.411964416503906, 285.8489990234375, 33.7849235534668, 445.6240234375, 10.936408996582031, 240.94760131835938, 298.0799865722656, 434.48291015625, 406.4166564941406, -126.11750030517578, 214.86338806152344, 195.6852569580078, -84.74937438964844, 449.42669677734375, 201.1707000732422, 164.623046875, 44.33191680908203, 627.9092407226562, 176.3284149169922, -159.86651611328125, 994.797607421875, 79.11126708984375, -231.40896606445312, -190.27676391601562, 211.59597778320312, 173.38230895996094, 284.824951171875, -214.1021270751953, 198.5861053466797, 115.93913269042969, 481.6736755371094, -7.4416656494140625, 286.3733215332031, 386.82171630859375, 18.409374237060547, 753.2157592773438, 94.70831298828125, 172.62857055664062, -270.1160888671875, 131.73895263671875, 872.3502807617188, 508.3065185546875, 354.3779602050781, 55.64173126220703, 191.38406372070312, 81.48958587646484, 659.163330078125, 353.85369873046875, 118.66386413574219, 690.8157958984375, 386.2047119140625, 224.88409423828125, 71.0779037475586, 362.4735107421875, 280.8114318847656, 240.403076171875, -46.4898567199707, 398.19580078125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000337.npy"}
{"epoch": 0.4948604992657856, "step": 338, "batch_size": 64, "mean": 231.73355102539062, "std": 237.22447204589844, "min": -236.91192626953125, "p10": -14.811825561523417, "median": 174.88855743408203, "p90": 546.0404907226563, "max": 832.4035034179688, "pos_frac": 0.890625, "sample": [54.75175476074219, 582.8677978515625, 114.05928039550781, -204.04742431640625, 188.34774780273438, 61.818267822265625, -112.34490966796875, 248.82765197753906, 656.1655883789062, 251.1046142578125, 119.22792053222656, 72.76927185058594, -110.69415283203125, 515.6917724609375, 118.3973388671875, 312.57275390625, 38.221954345703125, 523.5723876953125, 431.77117919921875, 530.6242065429688, 726.1328735351562, 5.812614440917969, 245.64111328125, 165.7372589111328, 259.44769287109375, 31.830429077148438, 440.58349609375, -23.65087127685547, -28.19550323486328, 19.74535369873047, 217.277587890625, 546.174560546875, 523.2275390625, 108.10394287109375, 231.33590698242188, 341.7970275878906, 184.03985595703125, 53.46787643432617, 355.4624328613281, 75.31731414794922, 409.6281433105469, 59.97135925292969, 77.15240478515625, 89.31471252441406, 539.4500732421875, 38.47076416015625, 195.67147827148438, 136.64288330078125, 345.6754150390625, 147.47068786621094, 18.044921875, 34.2025032043457, 545.7276611328125, 832.4035034179688, -57.29192352294922, 275.2034912109375, 532.5355834960938, 160.08547973632812, 364.62799072265625, 156.5281982421875, 647.4613647460938, -236.91192626953125, 64.63470458984375, 581.2545166015625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000338.npy"}
{"epoch": 0.49632892804698975, "step": 339, "batch_size": 64, "mean": 229.24349975585938, "std": 251.36050415039062, "min": -280.97625732421875, "p10": -52.812923812866195, "median": 202.66622161865234, "p90": 586.4722290039062, "max": 863.9401245117188, "pos_frac": 0.828125, "sample": [59.181396484375, -77.58948516845703, -59.121402740478516, -38.0931396484375, 638.1228637695312, 28.827102661132812, 48.17500305175781, 164.513427734375, 393.8428649902344, 27.88269805908203, 863.9401245117188, 60.62745666503906, 57.19549560546875, 166.665771484375, 351.86187744140625, 343.81658935546875, 21.98759651184082, 24.833641052246094, 77.22207641601562, 588.1595458984375, 292.11065673828125, -6.882591247558594, 269.1332092285156, -14.685600280761719, 186.50119018554688, 138.5845184326172, 701.816162109375, 26.098342895507812, 324.1785888671875, 649.8777465820312, -147.69058227539062, 291.7151794433594, -269.5963134765625, 185.11190795898438, 503.6397399902344, 327.22283935546875, 178.08831787109375, -36.928611755371094, 270.3497314453125, 472.4936218261719, 162.2347869873047, 270.42669677734375, 480.86492919921875, 530.1712646484375, 124.80713653564453, 700.2907104492188, -164.50347900390625, 684.3944091796875, 326.8331604003906, -280.97625732421875, 236.56365966796875, 196.96502685546875, 0.2090911865234375, 241.49505615234375, 39.56577682495117, 208.36741638183594, 259.77880859375, -62.652305603027344, 519.00830078125, 340.04046630859375, 506.748046875, 582.53515625, 374.93701171875, 310.28900146484375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000339.npy"}
{"epoch": 0.4977973568281938, "step": 340, "batch_size": 64, "mean": 200.8597412109375, "std": 272.3008117675781, "min": -425.04620361328125, "p10": -119.0114624023437, "median": 179.3026351928711, "p90": 551.9887573242189, "max": 966.2357177734375, "pos_frac": 0.78125, "sample": [171.41131591796875, 252.812255859375, 115.14237213134766, 210.38275146484375, 244.70042419433594, 521.7597045898438, 675.358642578125, -425.04620361328125, 272.9902648925781, -57.009727478027344, 395.9629821777344, 966.2357177734375, 79.43240356445312, -242.82357788085938, 562.954833984375, -44.60826873779297, 299.5364074707031, 130.2777099609375, 2.5601730346679688, 52.765254974365234, 197.90016174316406, -179.04876708984375, -341.2622985839844, 520.6550903320312, -143.10011291503906, 134.8906707763672, 8.196493148803711, 432.6534423828125, -33.955230712890625, 326.1715087890625, 526.4012451171875, 23.770889282226562, 862.1790161132812, 675.1783447265625, 18.644285202026367, 145.2620849609375, 180.136962890625, 258.59869384765625, 144.15260314941406, 178.4683074951172, 214.59613037109375, 140.63775634765625, 197.97576904296875, -8.062198638916016, -152.29957580566406, 272.9883728027344, 227.96044921875, -16.768211364746094, 516.1361083984375, -25.987197875976562, 99.19579315185547, 437.8248291015625, 581.60986328125, 622.409423828125, 350.4693603515625, 237.24295043945312, 171.01150512695312, 43.85470962524414, 374.3669738769531, 181.62289428710938, -242.2669219970703, 434.85504150390625, -62.80461120605469, 137.7662353515625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000340.npy"}
{"epoch": 0.49926578560939794, "step": 341, "batch_size": 64, "mean": 225.90994262695312, "std": 266.6433410644531, "min": -288.3016357421875, "p10": -86.69448776245116, "median": 214.07107543945312, "p90": 534.2040710449219, "max": 1109.6292724609375, "pos_frac": 0.78125, "sample": [1109.6292724609375, -89.5545654296875, -111.25697326660156, -201.93026733398438, 486.31494140625, 515.8440551757812, 704.116455078125, 494.1365051269531, 241.4485626220703, 169.01089477539062, -3.895265579223633, 199.29873657226562, 134.65011596679688, 255.17709350585938, 13.42603874206543, -46.11671447753906, -44.88026428222656, 367.67620849609375, 103.2612533569336, 294.5606689453125, -170.452880859375, 404.1650390625, 450.7294616699219, 244.23587036132812, 68.01378631591797, 575.2529296875, 109.82129669189453, 226.75888061523438, -72.20408630371094, -36.51560974121094, 145.83888244628906, 404.21380615234375, 642.4808349609375, 292.10150146484375, 357.22808837890625, 405.6560974121094, 94.04136657714844, -3.8428478240966797, -96.17435455322266, 502.88232421875, 538.1268310546875, 302.99774169921875, 18.589765548706055, 76.3836441040039, -288.3016357421875, 525.0509643554688, 36.56074523925781, 214.21810913085938, -253.90835571289062, 357.8442687988281, 213.92404174804688, 310.6967468261719, 430.7052001953125, -80.0209732055664, 749.2286376953125, 16.778182983398438, 318.31793212890625, 430.08154296875, 444.585205078125, 10.72506332397461, 167.49563598632812, 139.40113830566406, 553.6123657226562, 89.99681854248047], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000341.npy"}
{"epoch": 0.5007342143906021, "step": 342, "batch_size": 64, "mean": 206.4623260498047, "std": 276.11126708984375, "min": -587.0816650390625, "p10": -76.33133850097656, "median": 163.18711853027344, "p90": 486.30195922851567, "max": 1050.01513671875, "pos_frac": 0.78125, "sample": [723.1648559570312, -587.0816650390625, -215.831787109375, 491.6152038574219, 107.92803955078125, 44.276817321777344, 156.97808837890625, 377.69171142578125, 304.4686279296875, 351.35107421875, 113.33094787597656, 100.57017517089844, 197.8212432861328, 234.9113311767578, 551.87939453125, 276.959716796875, 61.02980041503906, 452.0282897949219, -197.95654296875, -34.879913330078125, 452.31365966796875, -62.160057067871094, 145.07101440429688, -58.11461639404297, -101.8084487915039, 291.175048828125, 984.2220458984375, 363.3288269042969, 123.83151245117188, -161.5513153076172, -49.82892990112305, 1050.01513671875, 225.41165161132812, 167.61737060546875, 1.4595489501953125, -68.1520004272461, 561.300537109375, 453.0819091796875, 646.6638793945312, 142.0097198486328, 25.658830642700195, 181.27340698242188, 361.6299133300781, -28.259368896484375, 473.9043884277344, 391.0219421386719, 164.03216552734375, 369.09521484375, 118.6373291015625, 432.70819091796875, 457.8194580078125, 87.45301055908203, -169.44586181640625, 124.40882873535156, 404.1905212402344, -67.36077117919922, 162.34207153320312, -79.8367691040039, 207.68695068359375, 123.68191528320312, 237.2186279296875, 430.6846923828125, 131.88694763183594, 57.015586853027344], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000342.npy"}
{"epoch": 0.5022026431718062, "step": 343, "batch_size": 64, "mean": 244.79345703125, "std": 214.41436767578125, "min": -336.98626708984375, "p10": 26.349974060058635, "median": 204.7386474609375, "p90": 480.05567321777346, "max": 831.4320678710938, "pos_frac": 0.9375, "sample": [374.2253112792969, 171.94699096679688, 185.93955993652344, 170.50909423828125, 250.98159790039062, 9.199134826660156, 280.2930908203125, 346.84521484375, 200.03741455078125, 709.3081665039062, 67.44215393066406, -141.52313232421875, 206.84994506835938, 69.14434814453125, 197.32501220703125, 71.42706298828125, 90.52144622802734, 238.23928833007812, 245.23587036132812, 667.2198486328125, 161.8247528076172, 252.55792236328125, 99.21119689941406, 308.2877502441406, 414.9371337890625, 354.28375244140625, 88.657470703125, 106.18991088867188, 831.4320678710938, 197.52549743652344, 66.36859893798828, 93.94479370117188, 339.2935485839844, 152.77883911132812, 297.2522277832031, 326.8030090332031, 225.04324340820312, 110.95234680175781, 6.482128143310547, 406.4971618652344, -336.98626708984375, 336.0151672363281, 186.03155517578125, 431.5195007324219, 271.2158508300781, 794.66357421875, 389.92327880859375, 343.5048522949219, 234.17672729492188, 344.5082092285156, 181.0601806640625, 374.6824645996094, 79.77587890625, -34.87435531616211, 479.02923583984375, 480.4955749511719, 186.7626495361328, 173.47857666015625, 179.09786987304688, 202.62734985351562, 3.2889938354492188, 675.806884765625, 611.6469116210938, -172.16033935546875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000343.npy"}
{"epoch": 0.5036710719530103, "step": 344, "batch_size": 64, "mean": 215.15504455566406, "std": 295.97857666015625, "min": -235.83932495117188, "p10": -58.18546524047851, "median": 179.8611297607422, "p90": 547.9165771484376, "max": 1608.604736328125, "pos_frac": 0.8125, "sample": [39.894248962402344, -52.81098175048828, -60.48881530761719, -11.59320068359375, 62.87367630004883, 63.90399169921875, 174.87989807128906, 119.95507049560547, 79.17143249511719, 250.10289001464844, -49.604034423828125, 897.189453125, 95.65581512451172, 212.38475036621094, 60.75382995605469, 184.8423614501953, 206.2708282470703, -190.69956970214844, -215.78878784179688, 282.688232421875, 253.23094177246094, 50.15568542480469, 165.7208709716797, 23.227222442626953, 223.9068603515625, 778.9205322265625, 589.1458740234375, 447.1406555175781, 54.653289794921875, -32.508331298828125, 416.1504211425781, 250.62872314453125, 277.1564025878906, 65.68780517578125, 231.31857299804688, -10.460723876953125, 359.501953125, -76.26150512695312, 273.4490966796875, 59.83540344238281, 552.415771484375, 311.2972106933594, 198.63233947753906, 62.74590301513672, 67.28897857666016, 1608.604736328125, 440.712158203125, 284.5402526855469, 53.36326599121094, 663.0332641601562, 11.959617614746094, 449.8670349121094, 185.5035400390625, 594.7327880859375, 39.81635665893555, -151.5823516845703, -168.49160766601562, 446.8369140625, 534.1309204101562, -235.83932495117188, 264.0611267089844, 375.4710998535156, 93.22315979003906, 537.41845703125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000344.npy"}
{"epoch": 0.5051395007342144, "step": 345, "batch_size": 64, "mean": 239.43661499023438, "std": 268.2687683105469, "min": -430.0550231933594, "p10": -52.09018363952634, "median": 199.9832534790039, "p90": 592.3355712890626, "max": 914.293701171875, "pos_frac": 0.8125, "sample": [53.27086639404297, 653.8857421875, 160.29110717773438, 378.70623779296875, 149.34324645996094, -209.03155517578125, 762.4423217773438, -154.08750915527344, 164.69937133789062, 441.9814453125, -18.353811264038086, 386.20269775390625, 147.69207763671875, 408.5862121582031, 499.3397521972656, 126.39498901367188, -271.6340637207031, 352.43597412109375, 38.877254486083984, 88.69117736816406, 453.8330993652344, 371.2276611328125, 63.88618469238281, 415.3348083496094, 254.42144775390625, 575.0521850585938, 130.39907836914062, 304.742919921875, 94.53146362304688, 198.9691925048828, 49.1107177734375, 816.3157348632812, 360.8802795410156, 114.66078186035156, 102.06088256835938, 454.5403137207031, 389.35113525390625, 508.0108642578125, 351.8531799316406, 284.67095947265625, 382.57891845703125, 914.293701171875, 84.44857788085938, 122.98550415039062, 439.45263671875, 279.1324157714844, -22.574413299560547, 739.6273193359375, 140.69729614257812, 464.5933837890625, -20.100730895996094, -68.56439971923828, 206.58444213867188, 599.7427368164062, -11.157842636108398, -178.30044555664062, -64.73979949951172, 113.00662231445312, 676.6461181640625, 200.997314453125, 212.71142578125, 97.7225341796875, -430.0550231933594, -9.370986938476562], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000345.npy"}
{"epoch": 0.5066079295154186, "step": 346, "batch_size": 64, "mean": 219.51083374023438, "std": 260.60430908203125, "min": -572.800048828125, "p10": -95.05551528930661, "median": 211.13385009765625, "p90": 582.5964172363282, "max": 798.45703125, "pos_frac": 0.796875, "sample": [292.00201416015625, 185.41444396972656, 33.39318084716797, 366.33636474609375, 710.40380859375, 478.0400085449219, 723.09716796875, 206.23313903808594, 201.77479553222656, 743.748779296875, -572.800048828125, 113.18296813964844, 390.4559631347656, -18.74935531616211, 319.9786682128906, 619.230712890625, -7.960418701171875, 372.5126953125, 798.45703125, 15.347505569458008, 333.07037353515625, 176.32400512695312, -147.34457397460938, 405.5225830078125, 303.97955322265625, 329.7936706542969, 304.01165771484375, 252.59231567382812, 337.8055419921875, 56.767059326171875, 17.572751998901367, -108.31600189208984, 201.6741485595703, 84.38169860839844, 590.9157104492188, 197.73020935058594, 216.03456115722656, 121.27825927734375, 301.3389892578125, -163.9977264404297, -23.60053253173828, 588.7632446289062, 332.0575256347656, 526.9238891601562, 170.2825469970703, 362.4136962890625, 393.5809020996094, 330.7243957519531, -26.30899429321289, -47.604976654052734, 230.01724243164062, 21.950572967529297, 80.90359497070312, -182.74057006835938, 69.86949157714844, 568.2071533203125, 124.541748046875, 124.73745727539062, 457.7279357910156, -214.52249145507812, -64.1143798828125, -174.9661865234375, 298.085205078125, 320.530517578125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000346.npy"}
{"epoch": 0.5080763582966226, "step": 347, "batch_size": 64, "mean": 225.85629272460938, "std": 283.8348693847656, "min": -290.29461669921875, "p10": -103.73359909057618, "median": 194.11590576171875, "p90": 612.1122314453128, "max": 925.5811767578125, "pos_frac": 0.765625, "sample": [-246.47354125976562, 265.238525390625, 496.70526123046875, 359.3840637207031, 91.97899627685547, -27.28961944580078, 204.26068115234375, 279.88311767578125, 637.0281982421875, 20.47002410888672, 438.1903076171875, 34.590721130371094, 128.23768615722656, -99.3714599609375, 380.1181640625, -238.53794860839844, -290.29461669921875, -53.436614990234375, 778.0352783203125, 115.3341293334961, 492.1473693847656, 28.88803482055664, 262.924560546875, 317.0179443359375, -240.86688232421875, 528.0133666992188, 553.9749755859375, 386.308837890625, -1.7221908569335938, 780.5263671875, 343.09173583984375, 675.9625244140625, 360.80523681640625, -75.28812408447266, 196.63790893554688, -25.792381286621094, 500.671875, 191.59390258789062, 137.3240966796875, -102.67312622070312, 206.73556518554688, 286.92364501953125, 86.56261444091797, 344.8924560546875, 156.25442504882812, 101.56692504882812, 121.4539794921875, 12.239311218261719, 925.5811767578125, 83.3489761352539, 329.22772216796875, 651.888916015625, 183.36570739746094, 849.29931640625, -104.1880874633789, 79.40247344970703, 334.4599914550781, -191.9830780029297, 442.5662841796875, -18.548843383789062, 464.2011413574219, -159.43785095214844, 143.73814392089844, 541.6544799804688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000347.npy"}
{"epoch": 0.5095447870778267, "step": 348, "batch_size": 64, "mean": 187.0452880859375, "std": 266.9032897949219, "min": -314.039794921875, "p10": -147.3863967895508, "median": 193.83759307861328, "p90": 435.50292968750006, "max": 1168.91552734375, "pos_frac": 0.765625, "sample": [302.8358459472656, 76.9616928100586, 174.10714721679688, 418.83612060546875, 338.3648681640625, -73.48169708251953, 99.45233917236328, 275.2994384765625, 381.8634033203125, -314.039794921875, 143.44569396972656, 97.37738037109375, 629.541259765625, 262.5682067871094, 113.32066345214844, 301.0659484863281, 8.296136856079102, 303.2356872558594, 106.15521240234375, 125.09423828125, 337.6847839355469, 331.89129638671875, -145.12619018554688, 1168.91552734375, 440.841552734375, 236.84896850585938, -65.18151092529297, 256.4974670410156, -35.18658447265625, 76.13195037841797, -83.84840393066406, -195.66734313964844, -197.30020141601562, 943.0906372070312, 591.909423828125, 369.9364013671875, 301.8397521972656, 307.15081787109375, 240.90171813964844, 39.52427673339844, -221.38758850097656, 404.268310546875, -31.41059112548828, 305.957763671875, 268.0170593261719, 230.90158081054688, -179.86083984375, 43.760765075683594, 533.7528076171875, 267.2265625, 96.76245880126953, -44.13243103027344, 423.046142578125, -250.73846435546875, 363.6437683105469, 213.5680389404297, 50.09845733642578, 2.4590187072753906, 33.42999267578125, 533.4693603515625, 109.5426025390625, 322.429443359375, -148.3550567626953, -46.704898834228516], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000348.npy"}
{"epoch": 0.5110132158590308, "step": 349, "batch_size": 64, "mean": 219.6474151611328, "std": 316.382568359375, "min": -456.4084167480469, "p10": -89.89638290405271, "median": 151.66849517822266, "p90": 583.6179260253906, "max": 1390.4326171875, "pos_frac": 0.796875, "sample": [147.28610229492188, -361.4437255859375, 446.8137512207031, 111.96671295166016, 82.00607299804688, 143.31240844726562, 475.537109375, -75.47853088378906, 229.96890258789062, 464.6586608886719, 176.69671630859375, 431.6900634765625, 150.10968017578125, -17.819725036621094, 42.57305145263672, 9.704626083374023, 30.6834659576416, 12.109580993652344, 248.63925170898438, -209.31951904296875, 502.523681640625, 369.21649169921875, 331.76654052734375, 26.514780044555664, 350.283447265625, -21.260730743408203, 477.49481201171875, -96.0754623413086, 125.2773666381836, 153.22731018066406, 779.3720092773438, 39.32917022705078, -55.50858688354492, 490.93994140625, 171.81698608398438, -456.4084167480469, -130.4693603515625, 108.53870391845703, 575.673583984375, 51.61612319946289, 805.5587158203125, -44.346832275390625, -309.6571044921875, -45.50392150878906, 587.0226440429688, 587.192626953125, 968.5301513671875, 158.38148498535156, 588.4764404296875, 173.50509643554688, 438.9669494628906, 443.39166259765625, 9.990531921386719, 289.9852600097656, 268.16021728515625, 363.7049865722656, -135.27743530273438, 54.89521789550781, 3.257183074951172, 512.94873046875, 463.6708068847656, 77.59628295898438, 72.98907470703125, 1390.4326171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000349.npy"}
{"epoch": 0.5124816446402349, "step": 350, "batch_size": 64, "mean": 204.90655517578125, "std": 296.6767883300781, "min": -586.86279296875, "p10": -188.07991790771482, "median": 203.43915557861328, "p90": 517.0533508300782, "max": 971.0626831054688, "pos_frac": 0.796875, "sample": [818.97705078125, -402.65533447265625, 127.32537078857422, 369.16497802734375, 279.19580078125, 189.79135131835938, 445.2022705078125, 92.9676513671875, -148.48023986816406, 98.21000671386719, 97.57131958007812, 417.8749694824219, -451.93756103515625, 154.0370635986328, 193.06112670898438, 26.766510009765625, 744.7002563476562, 487.0086669921875, 971.0626831054688, 260.53326416015625, 194.315185546875, 217.07249450683594, -32.258018493652344, 449.8550720214844, 261.3019104003906, 744.7075805664062, 146.74737548828125, -205.05120849609375, 123.66876983642578, -0.7318649291992188, 529.013671875, 138.76718139648438, 180.3707275390625, 161.43478393554688, 358.69195556640625, 284.5469970703125, 133.44247436523438, 289.7496337890625, -282.30633544921875, 677.7430419921875, 202.35594177246094, 483.00390625, 339.6324157714844, -137.28170776367188, 161.24652099609375, 238.5272216796875, 616.5740356445312, 442.9518127441406, 213.14306640625, 240.56097412109375, -586.86279296875, 412.52606201171875, 74.01707458496094, 261.6512145996094, 97.40465545654297, 204.52236938476562, -333.605224609375, -105.48602294921875, -36.599647521972656, 207.34786987304688, 489.14593505859375, 220.82424926757812, -205.83216857910156, 472.79266357421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000350.npy"}
{"epoch": 0.5139500734214391, "step": 351, "batch_size": 64, "mean": 258.5798645019531, "std": 322.65350341796875, "min": -350.12030029296875, "p10": -92.64245147705074, "median": 229.93690490722656, "p90": 702.3677185058594, "max": 1127.3704833984375, "pos_frac": 0.78125, "sample": [105.98828125, 327.827880859375, 605.8496704101562, 1126.6064453125, 1127.3704833984375, 158.3922576904297, 460.09429931640625, 112.77349853515625, 271.094482421875, 310.1195068359375, 44.42852783203125, 265.5003356933594, 792.5858154296875, 854.79833984375, 329.4505615234375, 346.8835754394531, 553.7274169921875, 576.6289672851562, 699.40087890625, -20.398962020874023, -201.37973022460938, 431.11395263671875, 102.2474365234375, 224.45370483398438, 189.33670043945312, 305.6770935058594, 460.7485046386719, 241.46018981933594, -350.12030029296875, -52.64643859863281, -335.2136535644531, 143.44586181640625, -166.4615020751953, 129.09646606445312, 7.405303955078125, 904.207275390625, 28.81066131591797, 119.07095336914062, 235.42010498046875, 39.051475524902344, -109.78359985351562, 397.75079345703125, 34.28926467895508, 676.6013793945312, 61.495147705078125, -3.832162857055664, 207.8757781982422, -23.02655029296875, -41.542236328125, -179.3690643310547, -39.479652404785156, 703.6392211914062, 182.3892059326172, 341.7012939453125, 758.0753173828125, -174.8203582763672, 314.89892578125, 616.7933349609375, 295.0466003417969, 329.2832336425781, 171.79527282714844, 316.03271484375, 245.2847442626953, -36.83477020263672], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000351.npy"}
{"epoch": 0.5154185022026432, "step": 352, "batch_size": 64, "mean": 159.57928466796875, "std": 289.8663635253906, "min": -648.714111328125, "p10": -128.3468894958496, "median": 138.53746032714844, "p90": 531.0373657226564, "max": 1033.32373046875, "pos_frac": 0.6875, "sample": [402.0065612792969, 375.8427734375, 147.00726318359375, 547.6063232421875, -360.7088317871094, 138.63275146484375, 74.23096466064453, 412.95928955078125, 48.23332214355469, 189.95729064941406, 690.18798828125, 482.149658203125, -64.14729309082031, 65.41356658935547, 216.6216583251953, -55.80278015136719, -109.99908447265625, 1033.32373046875, 283.647705078125, -120.67581939697266, 208.69671630859375, 706.339111328125, 179.1582489013672, 455.027099609375, 79.09022521972656, -91.70331573486328, -84.67990112304688, -234.8499755859375, 2.7387161254882812, 103.25704956054688, -34.391143798828125, 492.37646484375, 17.057035446166992, 46.81519317626953, 551.9456787109375, 138.44216918945312, -648.714111328125, 94.00606536865234, 234.31524658203125, 395.3569030761719, 209.31373596191406, 200.41275024414062, 170.25067138671875, -240.30685424804688, -84.54798889160156, -44.394744873046875, -223.24606323242188, 624.8201293945312, 490.83197021484375, -172.5757293701172, 286.7818908691406, -131.63449096679688, -70.12374114990234, 455.04534912109375, 231.33541870117188, 217.03436279296875, 418.7001647949219, -50.40331268310547, 649.1676635742188, 120.4852523803711, -76.95578002929688, -98.35319519042969, 239.37252807617188, 85.29306030273438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000352.npy"}
{"epoch": 0.5168869309838473, "step": 353, "batch_size": 64, "mean": 212.2716064453125, "std": 290.1940612792969, "min": -419.0412292480469, "p10": -105.98109207153318, "median": 190.8486557006836, "p90": 638.3189880371094, "max": 906.0914306640625, "pos_frac": 0.703125, "sample": [379.74615478515625, 392.3807678222656, -2.5659866333007812, 456.01947021484375, 180.1851806640625, 449.27777099609375, 82.52017211914062, -90.36711120605469, 612.3287353515625, 178.4488525390625, 335.8356018066406, -112.67279815673828, 252.93836975097656, 626.810302734375, 63.73681640625, 375.59552001953125, -164.0653839111328, -52.112548828125, 613.7584228515625, 643.2512817382812, 207.48757934570312, 285.17449951171875, 10.108211517333984, 92.37100219726562, -419.0412292480469, 411.43927001953125, -240.2755126953125, -115.08612823486328, -81.49341583251953, 265.498291015625, -30.83271598815918, 201.5121307373047, 332.0501403808594, 368.753662109375, 22.537811279296875, -15.14834213256836, 6.68115234375, -160.29306030273438, 857.56298828125, 1.688323974609375, -66.74862670898438, 688.6430053710938, -50.567970275878906, 50.04517364501953, 117.4138412475586, 302.2809143066406, 313.3597717285156, 214.03282165527344, -9.429473876953125, 650.501708984375, 328.106201171875, -63.81730651855469, 661.5802001953125, 384.1786193847656, 906.0914306640625, 792.5299682617188, -129.2644500732422, 343.40093994140625, 617.610595703125, -52.4505615234375, 88.21379852294922, 74.68971252441406, -26.45526123046875, 229.69400024414062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000353.npy"}
{"epoch": 0.5183553597650514, "step": 354, "batch_size": 64, "mean": 233.4363250732422, "std": 234.2032470703125, "min": -130.51133728027344, "p10": -53.978421020507795, "median": 201.7919464111328, "p90": 516.4206085205079, "max": 985.6363525390625, "pos_frac": 0.8125, "sample": [380.8291015625, 454.2052001953125, 274.146728515625, 55.892913818359375, 457.94036865234375, 209.96170043945312, 642.1920166015625, -130.51133728027344, 262.3077392578125, -97.63774108886719, -15.556373596191406, -126.54413604736328, 215.2145538330078, 985.6363525390625, 452.7732849121094, 169.76165771484375, -37.647911071777344, 366.0311584472656, 526.9730224609375, 306.9275207519531, 26.03607940673828, 367.37701416015625, 179.82354736328125, 99.48663330078125, -71.05685424804688, 118.33407592773438, 378.97015380859375, 230.73617553710938, 113.86052703857422, 390.9059753417969, 602.8656616210938, -101.81590270996094, 296.5497131347656, 272.5928649902344, 53.96535873413086, 122.05360412597656, -60.977210998535156, 419.8996887207031, 193.6221923828125, 125.75666809082031, 491.7983093261719, -13.784446716308594, 365.2447509765625, 192.74685668945312, 74.95160675048828, 231.5225067138672, 64.50411224365234, 661.5343627929688, 408.4254150390625, -17.354766845703125, 17.960195541381836, 418.6019592285156, 405.09771728515625, 99.21234130859375, -86.14234161376953, 784.4459838867188, 71.05062103271484, 363.277099609375, 371.4862976074219, 152.74476623535156, 548.0526733398438, 144.4746856689453, 106.96951293945312, -28.777347564697266], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000354.npy"}
{"epoch": 0.5198237885462555, "step": 355, "batch_size": 64, "mean": 220.28045654296875, "std": 237.13021850585938, "min": -342.4050598144531, "p10": -56.67172126770019, "median": 209.4861297607422, "p90": 519.4760803222657, "max": 979.783203125, "pos_frac": 0.8125, "sample": [126.12919616699219, 523.1891479492188, 306.6122741699219, 456.5762939453125, 32.856163024902344, 265.27978515625, 234.8619384765625, 979.783203125, -295.35284423828125, 265.9072265625, 197.3646240234375, 590.5716552734375, -39.68402862548828, -17.708833694458008, 447.04656982421875, 497.56439208984375, -98.5260009765625, 574.7866821289062, 163.36253356933594, 139.00250244140625, 389.3720703125, 247.92208862304688, -92.67103576660156, 9.359600067138672, 365.3229675292969, 574.7696533203125, 499.18365478515625, 266.314697265625, 181.09556579589844, -3.8361053466796875, -50.41897201538086, 219.22866821289062, 61.11170959472656, -342.4050598144531, 294.43902587890625, -154.12818908691406, 358.48309326171875, 537.6757202148438, 315.43646240234375, 662.442138671875, 96.16541290283203, 289.43865966796875, 125.25651550292969, 148.45111083984375, 146.0704345703125, 307.471923828125, 91.7009506225586, 427.9359436035156, 91.62506103515625, 190.5401611328125, 510.812255859375, 338.02862548828125, 309.59136962890625, 247.398193359375, 104.56465148925781, -59.351470947265625, 74.05180358886719, -135.05758666992188, 199.74359130859375, -47.78000259399414, 97.85162353515625, 330.83245849609375, 351.8984375, 172.38909912109375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000355.npy"}
{"epoch": 0.5212922173274597, "step": 356, "batch_size": 64, "mean": 132.46340942382812, "std": 313.333984375, "min": -675.87451171875, "p10": -246.83377838134766, "median": 126.52412414550781, "p90": 530.1524932861331, "max": 1055.760498046875, "pos_frac": 0.65625, "sample": [191.09359741210938, -17.901954650878906, 359.6845397949219, 133.27450561523438, -46.44926834106445, -110.48245239257812, -137.43690490722656, 201.21615600585938, -94.58065795898438, 259.44427490234375, 319.2700500488281, 0.4806175231933594, 721.0120239257812, 455.1110534667969, -113.93525695800781, 562.7840576171875, -175.37635803222656, -210.58543395996094, -675.87451171875, -251.95095825195312, 25.14977264404297, 771.1026611328125, -83.77702331542969, 120.6591796875, 410.00946044921875, 562.3131103515625, -61.587059020996094, 437.2216491699219, -302.2804260253906, 73.55394744873047, -285.94964599609375, 707.1704711914062, 187.33273315429688, -235.8790283203125, 393.91259765625, 106.00761413574219, 263.70672607421875, 383.49102783203125, -251.52867126464844, 89.21769714355469, -85.82548522949219, -105.46491241455078, 165.52163696289062, 241.02252197265625, 8.403850555419922, 323.34619140625, 358.92010498046875, -176.50721740722656, 666.366943359375, 279.9750671386719, 210.39523315429688, 45.167144775390625, -321.272705078125, -11.134773254394531, 249.60311889648438, 132.38906860351562, 167.59950256347656, 345.990234375, -431.07708740234375, 1055.760498046875, 372.3076171875, 302.2244873046875, 0.23632049560546875, 5.067481994628906], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000356.npy"}
{"epoch": 0.5227606461086637, "step": 357, "batch_size": 64, "mean": 205.66246032714844, "std": 243.4106903076172, "min": -498.96575927734375, "p10": -53.201892852783196, "median": 194.98416137695312, "p90": 567.8913635253907, "max": 798.9446411132812, "pos_frac": 0.796875, "sample": [384.4268493652344, 573.0447387695312, 13.208145141601562, 666.4180908203125, 505.0446472167969, -25.934799194335938, 94.13020324707031, 393.86016845703125, 122.9986572265625, 588.218017578125, 277.3636474609375, 86.4371566772461, -101.0552749633789, 584.4120483398438, -55.521759033203125, 149.01614379882812, 227.21554565429688, 158.12826538085938, 304.4444580078125, -47.60340881347656, 691.1827392578125, -41.911231994628906, 334.5464782714844, 26.508224487304688, -498.96575927734375, 52.47132110595703, 296.624267578125, 187.9154052734375, -47.78887176513672, 555.8668212890625, 353.23687744140625, -120.12740325927734, 427.0242004394531, -170.02627563476562, 367.31182861328125, -10.27621841430664, 358.9872131347656, 798.9446411132812, 285.6439208984375, 284.67926025390625, 5.602020263671875, -6.818811416625977, 301.6830749511719, 478.56103515625, 0.9130935668945312, 102.54469299316406, 54.90116882324219, 269.66180419921875, 4.218441009521484, 202.05291748046875, -85.333984375, 299.27996826171875, 226.74484252929688, 14.936508178710938, 297.411865234375, 315.2730712890625, 156.58541870117188, 361.09375, -64.64529418945312, 247.22315979003906, 85.87013244628906, 657.6759033203125, 49.42999267578125, 157.43377685546875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000357.npy"}
{"epoch": 0.5242290748898678, "step": 358, "batch_size": 64, "mean": 189.6993865966797, "std": 246.72645568847656, "min": -410.98828125, "p10": -49.414075088500965, "median": 122.18896484375, "p90": 475.67034301757815, "max": 870.60498046875, "pos_frac": 0.78125, "sample": [97.61016845703125, 607.5636596679688, 35.43214416503906, 253.67649841308594, 248.8257598876953, 356.4632263183594, 71.71055603027344, 302.1544189453125, 118.85873413085938, 85.01206970214844, -20.733360290527344, -21.11199951171875, -129.49288940429688, -26.307418823242188, 16.582212448120117, 694.0673828125, -11.088523864746094, 459.94171142578125, 82.42733764648438, 80.00704956054688, 100.0040512084961, -28.800491333007812, 17.71373748779297, 64.59111785888672, 119.070068359375, -410.98828125, 331.145263671875, -158.004150390625, 870.60498046875, 358.030517578125, 479.28057861328125, 186.21084594726562, 294.3094482421875, -8.944068908691406, 572.1796264648438, 204.06463623046875, 7.581550598144531, 95.72970581054688, -79.05187225341797, 377.6909484863281, 318.29302978515625, 390.84356689453125, -160.1563720703125, 277.00799560546875, 401.5120544433594, -55.779266357421875, 336.1859436035156, 802.6123046875, -34.56196212768555, 96.52507019042969, 336.1446533203125, 154.25772094726562, 94.265625, 65.28672790527344, 467.2464599609375, 862.416015625, 95.48680877685547, 145.65505981445312, 250.9796600341797, -86.57876586914062, 186.76136779785156, 153.97637939453125, 223.05625915527344, 125.307861328125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000358.npy"}
{"epoch": 0.5256975036710719, "step": 359, "batch_size": 64, "mean": 218.56814575195312, "std": 288.32183837890625, "min": -378.82879638671875, "p10": -98.4683967590332, "median": 185.53653717041016, "p90": 604.7622070312501, "max": 1080.3106689453125, "pos_frac": 0.75, "sample": [-378.82879638671875, 206.91192626953125, 228.04669189453125, 444.8939514160156, -234.08261108398438, 680.8341674804688, -164.01876831054688, 920.6212158203125, 156.61209106445312, -73.00933837890625, -111.37580871582031, -94.08634185791016, 367.00067138671875, 337.9489440917969, -231.7060546875, -27.876815795898438, 318.8248596191406, -28.11932373046875, 458.82843017578125, -44.557289123535156, 287.6908874511719, 238.0518798828125, 485.4053955078125, 54.806480407714844, 15.925495147705078, 187.20240783691406, -18.671621322631836, 474.97308349609375, 60.55999755859375, 52.00597381591797, 466.428466796875, 275.28985595703125, 1080.3106689453125, 429.66888427734375, -84.59568786621094, 52.85297393798828, 207.47824096679688, 102.9275131225586, -176.47216796875, 673.68359375, 157.0753173828125, 487.0788879394531, 225.72813415527344, 399.95697021484375, -65.52249145507812, 160.3914031982422, 449.50286865234375, 150.880859375, 130.94070434570312, 632.092529296875, 51.74663543701172, -10.663558959960938, -100.34642028808594, 296.0195617675781, 237.39776611328125, 620.0050048828125, 23.244598388671875, 207.9515838623047, 798.432861328125, 527.3922729492188, 84.40294647216797, 569.1956787109375, 173.2029571533203, 183.87066650390625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000359.npy"}
{"epoch": 0.527165932452276, "step": 360, "batch_size": 64, "mean": 173.5511474609375, "std": 198.30625915527344, "min": -174.05084228515625, "p10": -60.44139404296875, "median": 134.84943389892578, "p90": 426.9960876464844, "max": 640.155517578125, "pos_frac": 0.796875, "sample": [65.21188354492188, 457.60748291015625, 32.610984802246094, 133.78558349609375, -60.624046325683594, 344.56451416015625, 26.273902893066406, 214.550048828125, 315.0499572753906, 170.37303161621094, 128.08804321289062, 453.9599609375, -110.9642333984375, 310.33673095703125, 613.5388793945312, 135.9132843017578, 34.513431549072266, 24.874404907226562, 335.4647216796875, 100.06381225585938, 398.0960693359375, 194.2159423828125, -44.42681121826172, 251.20448303222656, 34.07088851928711, 338.0438232421875, 413.02801513671875, 57.35313415527344, 5.559955596923828, 429.82135009765625, 511.9960632324219, 334.3114318847656, 543.33740234375, -89.36874389648438, 38.640464782714844, 163.48585510253906, 197.13021850585938, 120.01077270507812, 420.40380859375, -31.541839599609375, 90.25743103027344, 640.155517578125, -114.30154418945312, -65.38128662109375, -33.493133544921875, 144.347900390625, -118.5100326538086, 185.16419982910156, -60.01520538330078, -174.05084228515625, -0.2399749755859375, 327.3752136230469, 14.955602645874023, 323.0356140136719, 373.2322998046875, 243.79103088378906, 293.67388916015625, 9.272605895996094, 110.38919067382812, 407.8109130859375, 109.97132873535156, -53.802406311035156, 41.968711853027344, 401.1019287109375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000360.npy"}
{"epoch": 0.5286343612334802, "step": 361, "batch_size": 64, "mean": 172.2711181640625, "std": 266.155517578125, "min": -430.6862487792969, "p10": -124.47952804565429, "median": 167.54246520996094, "p90": 492.9741516113281, "max": 798.848388671875, "pos_frac": 0.71875, "sample": [5.310842514038086, 184.65774536132812, 19.631084442138672, 390.51495361328125, -58.12274932861328, -269.4139404296875, -21.987709045410156, 249.4644317626953, 480.0346984863281, 144.22962951660156, 142.30075073242188, 225.12393188476562, -430.6862487792969, 422.00726318359375, -99.49617004394531, -111.89920806884766, 210.0227508544922, 167.7344512939453, -284.6890869140625, -4.406312942504883, 194.87869262695312, -181.10137939453125, 18.217226028442383, 6.179843902587891, 210.6019744873047, 791.815185546875, 293.1064453125, 489.59869384765625, 113.61273193359375, 467.71527099609375, -105.61613464355469, -45.02126693725586, 167.35047912597656, 115.51376342773438, 475.8873291015625, 494.4207763671875, -137.74085998535156, 798.848388671875, 742.4996948242188, 551.0907592773438, 234.58082580566406, 234.7967987060547, 591.5928955078125, 428.0666809082031, 175.96168518066406, 214.14993286132812, 19.95757293701172, 206.2830047607422, -15.869659423828125, -99.8856430053711, 50.522987365722656, -52.93771743774414, 465.1502380371094, 19.072120666503906, 320.597900390625, 348.16015625, 145.33822631835938, -129.87109375, 279.1779479980469, -52.85212707519531, 112.64227294921875, -265.608642578125, 566.4993286132812, 407.63800048828125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000361.npy"}
{"epoch": 0.5301027900146843, "step": 362, "batch_size": 64, "mean": 273.7876281738281, "std": 263.2894592285156, "min": -143.87652587890625, "p10": -39.32108917236326, "median": 231.9535140991211, "p90": 600.2761108398438, "max": 1315.4599609375, "pos_frac": 0.828125, "sample": [119.81396484375, 150.08840942382812, 526.190673828125, 547.4124755859375, 283.0472717285156, -19.41844940185547, 1315.4599609375, 259.3072204589844, 117.63577270507812, 663.83740234375, 523.2999267578125, 117.0915756225586, 50.263771057128906, 164.60073852539062, 270.22650146484375, 496.6723937988281, 202.5531463623047, -79.44178009033203, 91.49139404296875, 371.07220458984375, 473.91357421875, 133.75564575195312, 50.59117889404297, -59.889251708984375, -15.915512084960938, -14.117891311645508, 160.35202026367188, 202.3966827392578, -96.53205108642578, 657.6236572265625, 609.35986328125, 105.40202331542969, -87.958984375, 320.2881164550781, 270.6651611328125, 184.67918395996094, 549.29833984375, -47.850791931152344, 192.5234832763672, 579.0806884765625, 264.7569274902344, 66.272705078125, 758.0255737304688, 450.7612609863281, 137.40350341796875, 195.3927001953125, 423.58148193359375, 173.461669921875, 62.37898254394531, 738.0488891601562, 235.91607666015625, 644.6928100585938, 516.385498046875, 361.57305908203125, 272.1929016113281, 442.1460266113281, 488.6895751953125, -143.87652587890625, 227.99095153808594, 360.9897155761719, 273.09393310546875, 309.8404235839844, -2.659820556640625, -73.51992797851562], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000362.npy"}
{"epoch": 0.5315712187958884, "step": 363, "batch_size": 64, "mean": 272.94000244140625, "std": 266.7232666015625, "min": -265.314453125, "p10": -29.48500690460205, "median": 268.24908447265625, "p90": 595.8030944824219, "max": 1170.0435791015625, "pos_frac": 0.8125, "sample": [-265.314453125, 353.25836181640625, 558.4707641601562, 265.2861328125, 129.94602966308594, 224.5576171875, 269.13763427734375, 273.38787841796875, 285.7939453125, 42.93340301513672, 415.63568115234375, 241.5859832763672, 354.074462890625, 59.76032257080078, 365.8228759765625, 887.8887939453125, -49.54057312011719, 641.1654052734375, 81.46046447753906, 717.2921752929688, 433.69482421875, 78.53142547607422, 273.3651123046875, 391.19451904296875, 274.41314697265625, 399.05963134765625, 1170.0435791015625, 562.6232299804688, -30.2337646484375, 601.620849609375, 191.10906982421875, 234.06211853027344, -12.106903076171875, 262.46771240234375, 445.7172546386719, 66.69501495361328, 266.761474609375, -51.128387451171875, 582.2283325195312, 452.7301025390625, 370.8494873046875, -101.31575775146484, 41.90056610107422, 458.47222900390625, -27.737905502319336, 360.8515625, 578.1438598632812, 4.38165283203125, 267.36053466796875, 499.0993347167969, -24.36054229736328, 143.9561004638672, -148.86166381835938, 431.4923095703125, 271.0277099609375, -21.230796813964844, -27.61164093017578, 226.363525390625, 41.30910873413086, 41.45209503173828, 651.292724609375, 726.5012817382812, -31.64514923095703, 291.01861572265625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000363.npy"}
{"epoch": 0.5330396475770925, "step": 364, "batch_size": 64, "mean": 207.45681762695312, "std": 265.1451110839844, "min": -446.8075866699219, "p10": -55.50465011596678, "median": 190.5390167236328, "p90": 534.6271911621094, "max": 925.24658203125, "pos_frac": 0.828125, "sample": [387.3394775390625, 264.47821044921875, -236.0587615966797, 59.640419006347656, -446.8075866699219, 571.588134765625, -239.83761596679688, 333.14959716796875, 260.2305908203125, 838.9223022460938, 188.26023864746094, 592.9306030273438, 400.4261779785156, 357.6866455078125, -44.892845153808594, 71.93559265136719, 55.67425537109375, -202.1634521484375, 322.345703125, -17.933685302734375, 10.937442779541016, 443.2533264160156, 222.96829223632812, 442.2961120605469, 336.61328125, 41.81813049316406, -33.29373550415039, 493.49853515625, 192.8177947998047, 280.8620300292969, -60.05256652832031, 658.2734375, 129.8587646484375, 1.211090087890625, 62.81879425048828, 115.27191162109375, 423.784423828125, 323.5353088378906, 319.6612243652344, -129.2615966796875, 121.71672821044922, 537.9688110351562, 601.9808959960938, 81.10154724121094, 305.1021728515625, 22.34557342529297, 19.984304428100586, 402.27978515625, 526.830078125, 347.1340637207031, 45.3650016784668, 182.74044799804688, 16.752042770385742, 179.603759765625, 925.24658203125, 71.81655883789062, 319.5946044921875, 309.7647399902344, -20.181894302368164, 245.7933349609375, 51.91497802734375, 164.34121704101562, 424.6685791015625, -400.4139709472656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000364.npy"}
{"epoch": 0.5345080763582967, "step": 365, "batch_size": 64, "mean": 209.97547912597656, "std": 256.68359375, "min": -292.2744445800781, "p10": -72.70946807861326, "median": 174.33267974853516, "p90": 516.6345214843751, "max": 1220.0916748046875, "pos_frac": 0.84375, "sample": [186.78762817382812, 212.28518676757812, 82.82623291015625, 136.8175811767578, 93.80367279052734, 127.02398681640625, 320.2464904785156, 29.50678253173828, 220.35220336914062, 137.83474731445312, 184.20445251464844, 446.5513610839844, -83.72718811035156, 219.36862182617188, 51.81446838378906, -48.6383056640625, 140.3579559326172, 688.3930053710938, 576.7714233398438, 243.62814331054688, 112.58609771728516, 119.63574981689453, 52.584739685058594, 19.2098388671875, 236.06321716308594, -118.63916015625, 117.01200866699219, 34.68350601196289, 164.46090698242188, 188.9478759765625, -83.02568054199219, -17.73482894897461, 111.12610626220703, 39.261329650878906, 415.99713134765625, 622.755126953125, 374.2132873535156, 77.46173095703125, -170.30899047851562, 493.60382080078125, 275.55169677734375, -182.82904052734375, 206.24514770507812, 14.488138198852539, -29.189157485961914, 335.8122253417969, 255.13970947265625, 321.73046875, 317.3538513183594, -107.64727783203125, 46.532020568847656, 381.88763427734375, 12.284616470336914, 280.3659362792969, 223.1704559326172, 358.09124755859375, 1220.0916748046875, 526.5048217773438, 467.96636962890625, -292.2744445800781, 815.6358642578125, 720.147216796875, 113.81452941894531, 401.4552001953125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000365.npy"}
{"epoch": 0.5359765051395007, "step": 366, "batch_size": 64, "mean": 217.92330932617188, "std": 279.7445983886719, "min": -309.865966796875, "p10": -133.11940612792966, "median": 193.73069763183594, "p90": 564.9975952148437, "max": 965.1703491210938, "pos_frac": 0.796875, "sample": [-93.27169036865234, 196.85812377929688, 319.6236877441406, 104.11892700195312, -285.82098388671875, 35.43814468383789, -86.29132843017578, 57.340145111083984, -138.80886840820312, 387.518798828125, 85.85218811035156, -243.30886840820312, 512.02001953125, 681.0014038085938, 190.603271484375, 9.078170776367188, 332.1481018066406, 32.296836853027344, 143.89576721191406, -81.13563537597656, 318.12603759765625, 518.9420166015625, 329.89703369140625, 21.976669311523438, 533.3319702148438, 965.1703491210938, 336.1176452636719, 97.33211517333984, 563.8048095703125, 230.69329833984375, 119.6906509399414, -57.27960205078125, -198.27487182617188, 591.7064819335938, 332.26983642578125, 441.48956298828125, 866.6749877929688, 328.24053955078125, 244.06314086914062, 134.04188537597656, 232.03541564941406, 135.93304443359375, 54.538604736328125, -153.04115295410156, 535.2195434570312, -148.71946716308594, 342.22576904296875, 60.41704177856445, 527.0872802734375, 565.5087890625, 390.295166015625, 281.5426025390625, 725.495361328125, 340.9295349121094, 715.1693115234375, 232.9556884765625, -59.58439636230469, 36.52613830566406, -119.843994140625, 138.19235229492188, 109.56340789794922, 144.49208068847656, 292.849365234375, -309.865966796875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000366.npy"}
{"epoch": 0.5374449339207048, "step": 367, "batch_size": 64, "mean": 181.51675415039062, "std": 229.78285217285156, "min": -499.8764953613281, "p10": -59.03878479003906, "median": 128.0403823852539, "p90": 490.3047546386719, "max": 727.9361572265625, "pos_frac": 0.796875, "sample": [88.92694091796875, 74.17948913574219, 488.5242919921875, 88.91940307617188, 628.6923828125, -21.495820999145508, 356.74749755859375, 125.417724609375, 491.06781005859375, 64.9208755493164, 27.16583251953125, 1.1665973663330078, -78.06231689453125, 522.7311401367188, 467.79461669921875, 248.87078857421875, 452.76531982421875, 129.75662231445312, 188.13790893554688, -54.20311737060547, 273.3526611328125, 414.5721740722656, -61.11121368408203, -16.38166046142578, 354.5912170410156, 277.4690856933594, 727.9361572265625, 37.89414978027344, 523.2499389648438, 83.9678955078125, 435.191650390625, 195.75669860839844, 2.7897682189941406, 308.337158203125, 79.51045227050781, 210.05404663085938, 322.4151611328125, 306.4434814453125, 343.2374572753906, 45.37468338012695, 155.38949584960938, -262.92987060546875, -153.77430725097656, -27.538022994995117, -115.51376342773438, 76.07159423828125, 98.8399658203125, 409.9634704589844, 126.32414245605469, 87.662109375, -171.65988159179688, -499.8764953613281, 131.4012451171875, 119.91529846191406, 265.16650390625, 403.216552734375, 456.958740234375, 235.81700134277344, 512.5986938476562, 14.744674682617188, 123.58476257324219, -0.7716903686523438, 491.18621826171875, -16.378631591796875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000367.npy"}
{"epoch": 0.5389133627019089, "step": 368, "batch_size": 64, "mean": 282.15570068359375, "std": 220.42160034179688, "min": -205.69326782226562, "p10": 11.403990554809575, "median": 295.0249481201172, "p90": 535.2732543945312, "max": 1009.1822509765625, "pos_frac": 0.921875, "sample": [278.9085998535156, 320.2829895019531, 190.687744140625, 488.6887512207031, 383.429931640625, 570.9327392578125, 158.98431396484375, 39.86582946777344, 536.3652954101562, 204.27474975585938, -19.457332611083984, 351.04071044921875, 503.7900390625, 29.139217376708984, 279.8693542480469, 382.62506103515625, 327.6560974121094, 186.21092224121094, -46.179603576660156, 372.7358703613281, 201.68978881835938, 244.0081024169922, 591.886474609375, 349.6703186035156, 333.4191589355469, 83.86929321289062, 389.14459228515625, 9.832225799560547, 69.55186462402344, 15.071441650390625, 175.1211700439453, 429.46051025390625, 435.5520324707031, 389.895263671875, 373.4715576171875, 532.7251586914062, 406.35272216796875, 60.3232421875, 185.1466827392578, 0.00363922119140625, 299.7850036621094, 496.82208251953125, 429.4114685058594, 614.5433349609375, 155.48121643066406, 411.7857971191406, 670.4776611328125, 22.394851684570312, -205.69326782226562, 112.69882202148438, 105.97315216064453, 487.5788879394531, 431.69659423828125, 473.2118225097656, 352.7003479003906, 43.84405517578125, 201.9215087890625, 1009.1822509765625, 138.26934814453125, 219.88983154296875, 290.264892578125, -37.15617370605469, -111.69422912597656, 628.528076171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000368.npy"}
{"epoch": 0.540381791483113, "step": 369, "batch_size": 64, "mean": 150.80130004882812, "std": 285.45965576171875, "min": -685.8173828125, "p10": -167.995703125, "median": 129.34215545654297, "p90": 495.38854675292976, "max": 837.8475952148438, "pos_frac": 0.75, "sample": [-209.48114013671875, -109.9861831665039, 837.8475952148438, -65.58145904541016, -277.12799072265625, 4.26399040222168, -72.5750961303711, 305.1461181640625, 0.10802841186523438, 619.762939453125, 214.7232208251953, -161.90304565429688, 121.16803741455078, 335.29339599609375, 317.2607421875, 830.1614990234375, 522.461669921875, 150.92953491210938, 42.118534088134766, 416.8371887207031, 275.7812805175781, 120.03059387207031, 119.56008911132812, 248.05947875976562, 242.205810546875, 147.75738525390625, -455.611083984375, 307.8202209472656, 145.18539428710938, 124.24468231201172, 235.9312286376953, 378.970947265625, 126.79796600341797, 130.2274932861328, 177.85324096679688, 19.71123504638672, 348.9731140136719, 88.2591781616211, -97.25578308105469, 305.98004150390625, 503.9091491699219, -211.79527282714844, 748.880126953125, -6.2623443603515625, 453.2295227050781, 241.26312255859375, 401.55352783203125, -467.5022277832031, 6.264026641845703, 632.925048828125, 160.60794067382812, 119.92448425292969, -685.8173828125, 83.3450698852539, 190.80059814453125, -170.60684204101562, -41.51954650878906, 71.86844635009766, 128.45681762695312, 475.50714111328125, 241.48660278320312, -59.214515686035156, 72.8328857421875, -50.763328552246094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000369.npy"}
{"epoch": 0.5418502202643172, "step": 370, "batch_size": 64, "mean": 140.53900146484375, "std": 276.5737609863281, "min": -384.4647216796875, "p10": -229.5109909057617, "median": 148.41221618652344, "p90": 522.6601379394533, "max": 813.989990234375, "pos_frac": 0.703125, "sample": [78.85956573486328, 292.57171630859375, 210.17893981933594, -361.33251953125, 669.0042114257812, 345.27349853515625, 153.48135375976562, 355.62066650390625, -19.032777786254883, -136.01473999023438, 428.07208251953125, 462.6634216308594, 42.938385009765625, 109.98992919921875, -143.15585327148438, 549.746826171875, 227.85531616210938, 197.12103271484375, 233.6434326171875, 813.989990234375, 705.59326171875, 178.8814697265625, 291.1640930175781, -172.57281494140625, 247.38449096679688, -115.9023208618164, 18.25604248046875, 293.6912841796875, -198.93841552734375, 184.83929443359375, 545.6752319335938, 122.8543701171875, -3.1009368896484375, 278.76446533203125, -23.26544952392578, 349.7577209472656, 224.92160034179688, -201.53289794921875, -373.62689208984375, 18.492238998413086, 143.34307861328125, 228.77276611328125, -241.50160217285156, 76.5704345703125, 468.958251953125, 112.8990478515625, 164.76470947265625, 310.530517578125, 242.0876922607422, -384.4647216796875, -105.46369934082031, -51.25274658203125, 131.02587890625, 123.54827880859375, -314.8112487792969, -292.1332702636719, 9.27764892578125, -300.3785095214844, -108.85467529296875, 464.429443359375, 595.88232421875, 243.91845703125, 21.580078125, 572.957275390625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000370.npy"}
{"epoch": 0.5433186490455213, "step": 371, "batch_size": 64, "mean": 214.00048828125, "std": 292.9512939453125, "min": -530.5565795898438, "p10": -118.68723220825196, "median": 183.6681900024414, "p90": 573.1404174804687, "max": 933.6305541992188, "pos_frac": 0.765625, "sample": [411.4880065917969, 266.2041015625, 505.8893127441406, 86.0675277709961, -530.5565795898438, 199.40365600585938, 574.5094604492188, 215.27203369140625, 826.7787475585938, 311.50689697265625, 482.8017578125, 18.949941635131836, 353.15948486328125, 219.04238891601562, -250.85238647460938, 118.85749816894531, -7.605316162109375, 182.37237548828125, 516.692138671875, 30.74627685546875, -78.67314147949219, -230.81300354003906, 595.947509765625, -120.15055847167969, -207.34820556640625, 184.96400451660156, 900.3432006835938, 569.9459838867188, -154.7460479736328, 27.419919967651367, 704.4161376953125, -13.750350952148438, 342.5592956542969, 256.13824462890625, -152.91680908203125, -61.81965637207031, 530.432373046875, -115.2728042602539, 251.98236083984375, 434.50274658203125, -7.420198440551758, 97.50261688232422, 735.40673828125, 104.40668487548828, 21.696773529052734, 250.89096069335938, 270.44354248046875, 430.4262390136719, 530.6326904296875, -33.73167419433594, 448.9818115234375, 111.02727508544922, 76.13960266113281, 315.3084716796875, 176.5984649658203, 180.55596923828125, 104.66288757324219, 71.70790100097656, 9.085128784179688, 478.89312744140625, 256.99737548828125, 933.6305541992188, -72.57890319824219, 10.878639221191406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000371.npy"}
{"epoch": 0.5447870778267254, "step": 372, "batch_size": 64, "mean": 256.0472717285156, "std": 346.38238525390625, "min": -502.070556640625, "p10": -127.10836563110351, "median": 193.4283218383789, "p90": 756.551727294922, "max": 1366.47412109375, "pos_frac": 0.765625, "sample": [349.88818359375, 364.7823486328125, 730.140869140625, -302.94085693359375, 58.06230163574219, 767.8706665039062, 575.6336669921875, 227.60226440429688, 346.2366943359375, -18.177780151367188, 1366.47412109375, 325.468017578125, -27.734130859375, 112.07211303710938, -70.36099243164062, 111.67253875732422, 70.73309326171875, 107.2520980834961, 816.67431640625, -192.54244995117188, 111.446044921875, 217.50057983398438, 716.4259033203125, -171.09286499023438, 543.4573974609375, 146.36614990234375, 114.6469955444336, 198.7215576171875, 89.13327026367188, 152.97390747070312, 156.41836547851562, 382.3291931152344, 406.32684326171875, -110.83873748779297, 150.45260620117188, -127.51250457763672, 368.3132019042969, 780.0182495117188, 63.077919006347656, 1048.05517578125, 568.1680297851562, 404.5057373046875, 188.1350860595703, -126.16537475585938, 481.57574462890625, 501.2971496582031, 682.3765258789062, 224.2302703857422, -9.586746215820312, 341.9364013671875, 117.54678344726562, 258.4775085449219, 102.43458557128906, -95.9188461303711, 804.96728515625, 150.0887908935547, -81.5731201171875, 219.35205078125, -146.66566467285156, 909.3098754882812, -267.1404113769531, -502.070556640625, 487.52001953125, 219.19931030273438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000372.npy"}
{"epoch": 0.5462555066079295, "step": 373, "batch_size": 64, "mean": 241.53067016601562, "std": 283.54425048828125, "min": -543.20654296875, "p10": -75.89708099365232, "median": 224.31005096435547, "p90": 597.91904296875, "max": 896.496337890625, "pos_frac": 0.828125, "sample": [147.67144775390625, -203.0902557373047, -151.77532958984375, 324.15765380859375, 30.15233039855957, 896.496337890625, 449.8149108886719, 64.08872985839844, 271.7726135253906, 145.04238891601562, 600.2376708984375, 198.76316833496094, 12.411224365234375, 252.84429931640625, 383.6007385253906, 42.13483428955078, 21.441593170166016, 759.4613647460938, 368.5512390136719, 23.51592254638672, -368.02935791015625, -46.501251220703125, 233.73861694335938, -543.20654296875, 135.180908203125, 56.752769470214844, 180.5799560546875, 118.65850830078125, 549.9842529296875, 389.4815673828125, 214.88148498535156, 487.63775634765625, 344.35791015625, 495.8443603515625, 466.4612731933594, -84.97024536132812, -174.65264892578125, 592.5089111328125, 280.93359375, -92.43199157714844, 828.1957397460938, 113.34465789794922, 444.89141845703125, 256.892578125, 412.5209045410156, 479.63427734375, 317.7593688964844, 252.6713104248047, 528.4994506835938, 767.8616943359375, 154.76177978515625, -16.70001220703125, 89.98504638671875, 155.0169677734375, -4.4308319091796875, 263.89080810546875, 112.47163391113281, 117.15048217773438, 734.4027709960938, 441.61578369140625, 419.6666564941406, 654.2608642578125, -54.72636413574219, 113.82299041748047], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000373.npy"}
{"epoch": 0.5477239353891337, "step": 374, "batch_size": 64, "mean": 216.1202850341797, "std": 341.48675537109375, "min": -693.7659301757812, "p10": -104.503409576416, "median": 148.923583984375, "p90": 567.1541748046875, "max": 1251.73828125, "pos_frac": 0.765625, "sample": [-196.343505859375, 156.3621826171875, -41.26679992675781, 1251.73828125, 322.5484619140625, 1183.914794921875, 264.7918395996094, 567.5390625, 365.0738525390625, 31.14147186279297, 143.64767456054688, 330.2396545410156, 35.065185546875, 202.6678466796875, 65.55532836914062, -48.982749938964844, 525.3297119140625, 154.19949340820312, 532.700927734375, 367.1612243652344, -53.566650390625, 19.12868881225586, 393.7528991699219, 49.05946350097656, -44.89604187011719, 835.5758056640625, -9.269002914428711, 90.57176208496094, -222.25877380371094, 337.1082763671875, 260.34259033203125, 268.8755798339844, -92.93386840820312, 909.6837768554688, -386.6910400390625, 99.06193542480469, 25.149917602539062, 566.256103515625, 120.37488555908203, -37.783935546875, 176.5549774169922, 99.97376251220703, 64.1170425415039, 81.62017822265625, 223.296142578125, 136.84817504882812, -154.18809509277344, 508.3849182128906, 280.92095947265625, 433.805419921875, 106.04426574707031, 594.2300415039062, -119.70216369628906, 424.273193359375, 309.1532897949219, -5.613822937011719, -109.46178436279297, 219.6387939453125, 20.189590454101562, 109.81410217285156, 1100.447021484375, -693.7659301757812, 164.15780639648438, 520.3336791992188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000374.npy"}
{"epoch": 0.5491923641703378, "step": 375, "batch_size": 64, "mean": 208.14962768554688, "std": 302.76678466796875, "min": -567.1668701171875, "p10": -164.50341339111324, "median": 179.05352783203125, "p90": 540.4511657714844, "max": 944.263671875, "pos_frac": 0.75, "sample": [644.0487670898438, 381.3743591308594, -16.1121826171875, 327.2745666503906, -132.9258575439453, -178.03665161132812, 476.148193359375, 218.62579345703125, 236.7146759033203, -29.462421417236328, -2.740335464477539, -37.64485168457031, 61.60832977294922, -567.1668701171875, 105.4888687133789, 917.955078125, 363.390380859375, 47.63433074951172, 327.19891357421875, 139.35125732421875, 25.517730712890625, 264.0180358886719, -491.73138427734375, 76.72696685791016, 485.82843017578125, -180.53334045410156, 524.3226928710938, 96.47563934326172, 49.3447265625, 87.82059478759766, -303.9084777832031, 35.03034973144531, 544.978759765625, 610.723876953125, 60.11317825317383, 253.80026245117188, 304.0465087890625, 436.2725830078125, 520.1442260742188, 695.5223999023438, 422.3588562011719, 182.07273864746094, 62.637351989746094, -319.4461669921875, 451.73187255859375, 511.0550842285156, 290.6258544921875, -221.33547973632812, 944.263671875, 527.2219848632812, 385.80889892578125, -0.7045936584472656, -1.91204833984375, 338.61737060546875, 127.65945434570312, 557.7946166992188, 176.03431701660156, 529.8867797851562, -87.72786712646484, 156.40829467773438, 399.727294921875, 453.0649108886719, 108.64358520507812, -50.14970016479492], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000375.npy"}
{"epoch": 0.5506607929515418, "step": 376, "batch_size": 64, "mean": 262.6305236816406, "std": 304.34686279296875, "min": -442.88134765625, "p10": -102.46461639404296, "median": 220.64064025878906, "p90": 678.3485839843751, "max": 862.1532592773438, "pos_frac": 0.828125, "sample": [55.347496032714844, 741.9503784179688, 239.11166381835938, 65.70382690429688, 813.910400390625, 116.92579650878906, 435.6846923828125, -252.18019104003906, 53.65031433105469, -54.63270568847656, 395.26654052734375, 0.29924964904785156, 664.2333984375, 101.57180786132812, -262.1632385253906, 443.60992431640625, 541.1419067382812, -148.63211059570312, -92.31289672851562, 228.32908630371094, 88.68351745605469, 21.382917404174805, -185.3031005859375, 862.1532592773438, 765.1176147460938, 105.54363250732422, 426.9541015625, 285.79864501953125, 546.8511352539062, -89.69075012207031, 684.762451171875, 509.35955810546875, 141.5828399658203, 500.02764892578125, 175.29620361328125, 580.8851928710938, 515.3453369140625, 194.40628051757812, 415.19000244140625, 222.27175903320312, 217.0303192138672, 495.1922302246094, -106.81535339355469, 288.86016845703125, 620.6712646484375, 193.4075469970703, 800.6268310546875, 14.719764709472656, -157.2805938720703, 70.15460968017578, 3.2418060302734375, 61.512962341308594, 656.5413818359375, 684.39794921875, 287.08453369140625, 590.0523071289062, 200.08889770507812, 344.93072509765625, 228.62159729003906, 219.009521484375, -57.03392791748047, 98.55311584472656, 644.2329711914062, -442.88134765625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000376.npy"}
{"epoch": 0.5521292217327459, "step": 377, "batch_size": 64, "mean": 243.29318237304688, "std": 295.7868347167969, "min": -373.2279968261719, "p10": -104.32897109985346, "median": 239.34966278076172, "p90": 669.5890747070313, "max": 904.8308715820312, "pos_frac": 0.734375, "sample": [300.1886901855469, 904.8308715820312, 178.10214233398438, -52.177459716796875, 314.6452941894531, -36.700164794921875, 244.60693359375, -180.80642700195312, 289.72119140625, 58.40241241455078, 785.2191772460938, 825.1600341796875, -32.011558532714844, 869.42333984375, -33.064659118652344, 193.35543823242188, 239.6172637939453, 458.9101867675781, 244.33187866210938, -22.479734420776367, 239.08206176757812, 601.7334594726562, 422.3186340332031, -23.44971466064453, 612.182373046875, 829.0609741210938, 339.9761047363281, -233.16453552246094, 677.3546142578125, 106.77167510986328, 296.38897705078125, 279.71728515625, 27.548545837402344, 112.91000366210938, 38.327030181884766, 513.3360595703125, 81.93305969238281, 305.5189208984375, 351.51031494140625, 719.6256103515625, 273.76397705078125, -27.81085205078125, -373.2279968261719, 176.4093017578125, -136.99429321289062, 116.95964813232422, 522.0493774414062, 402.613525390625, -218.8028106689453, 384.3990173339844, 49.63566589355469, 622.7557983398438, -25.357505798339844, -129.1932373046875, -41.36322784423828, 429.4280090332031, 189.6961669921875, 651.469482421875, 372.41839599609375, -17.245437622070312, 278.186279296875, 174.09246826171875, 175.60675048828125, -126.67961883544922], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000377.npy"}
{"epoch": 0.55359765051395, "step": 378, "batch_size": 64, "mean": 270.62432861328125, "std": 273.9478454589844, "min": -499.63140869140625, "p10": -43.699035644531236, "median": 251.02676391601562, "p90": 583.5501831054688, "max": 1051.1160888671875, "pos_frac": 0.859375, "sample": [705.4793090820312, 371.6235656738281, 650.2418212890625, 547.77099609375, 480.61480712890625, 117.81407165527344, 221.04702758789062, 437.05615234375, 196.874755859375, 213.27462768554688, 341.5042724609375, 309.6123046875, 568.0589599609375, 189.57142639160156, 94.9749755859375, -31.488021850585938, 262.0760803222656, 706.5841674804688, 214.84527587890625, 314.9088134765625, 249.06344604492188, -81.11150360107422, 8.70395278930664, 68.48395538330078, 253.0916748046875, 186.98233032226562, -57.72099304199219, 533.9795532226562, 82.43254852294922, 524.0477905273438, 586.167236328125, 455.3123474121094, 245.91458129882812, 90.18927001953125, -11.870140075683594, 459.0813903808594, 252.99008178710938, 1051.1160888671875, 30.47694206237793, -248.400390625, 295.7278747558594, 4.960060119628906, 243.15525817871094, 577.4437255859375, 133.56253051757812, -48.93232727050781, 399.7422790527344, 448.9547119140625, 528.4260864257812, 399.330322265625, 111.40913391113281, 328.25482177734375, 493.68792724609375, 231.1956787109375, -499.63140869140625, 203.373779296875, 263.83233642578125, 479.7978210449219, -106.98226165771484, 762.8388061523438, 95.88841247558594, -376.149169921875, 130.6870880126953, 628.0064697265625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000378.npy"}
{"epoch": 0.5550660792951542, "step": 379, "batch_size": 64, "mean": 260.3370056152344, "std": 313.7947082519531, "min": -510.9752197265625, "p10": -85.38948440551756, "median": 205.51063537597656, "p90": 681.903192138672, "max": 1310.780029296875, "pos_frac": 0.84375, "sample": [253.40371704101562, 422.06243896484375, 16.909027099609375, 260.0822448730469, -74.23384857177734, 79.04169464111328, 178.86175537109375, 698.0263061523438, 178.28976440429688, 156.8228759765625, 395.1231994628906, 808.6064453125, 7.8735809326171875, 517.677001953125, 355.08404541015625, 218.79208374023438, -510.9752197265625, 905.6633911132812, 21.585617065429688, 280.75042724609375, 469.73486328125, 439.70989990234375, 243.83030700683594, 39.12858963012695, 140.71844482421875, -53.73577880859375, -354.051025390625, 179.77752685546875, 60.08666229248047, 383.1064453125, 755.1287841796875, 375.69873046875, -252.890625, 143.97885131835938, 448.6568298339844, 738.164794921875, 336.65167236328125, 392.9170227050781, 644.2825927734375, 122.98422241210938, -61.90288543701172, 588.4974975585938, 191.65414428710938, 9.389850616455078, 376.88800048828125, -113.37493896484375, 171.24148559570312, 498.1365051269531, 299.011474609375, 183.9828643798828, -120.7382583618164, 383.1053466796875, 780.1144409179688, 189.48805236816406, 1310.780029296875, 325.1293029785156, 192.22918701171875, 484.45941162109375, 139.24984741210938, -90.17047119140625, 14.847457885742188, 609.3045654296875, -189.76138305664062, 66.68183898925781], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000379.npy"}
{"epoch": 0.5565345080763583, "step": 380, "batch_size": 64, "mean": 288.10162353515625, "std": 315.0541687011719, "min": -746.564697265625, "p10": -69.1470260620117, "median": 292.2878875732422, "p90": 685.6867614746094, "max": 1124.33154296875, "pos_frac": 0.859375, "sample": [144.1287841796875, 110.71974182128906, 97.38591003417969, 1124.33154296875, 162.72360229492188, 111.9523696899414, 136.24237060546875, 366.05224609375, 375.12542724609375, 264.0399169921875, 123.56896209716797, 382.1796875, -54.04259490966797, 593.3289794921875, 889.7568359375, 498.50555419921875, 455.5713195800781, 563.6775512695312, 421.5025939941406, 177.8851318359375, 378.9224853515625, -274.1957092285156, 283.177001953125, 157.770751953125, 68.857421875, -75.62035369873047, 473.8133544921875, 301.3987731933594, 225.59999084472656, 688.6596069335938, 383.7186279296875, 613.6699829101562, 594.47998046875, 615.4190063476562, 791.7054443359375, 415.0211181640625, 108.01118469238281, -32.57470703125, 229.95626831054688, 72.94995880126953, 803.5506591796875, 113.60206604003906, -222.80682373046875, 211.74490356445312, -192.927001953125, 206.6847686767578, 333.608154296875, 678.7501220703125, 65.45613098144531, -312.10394287109375, 411.8118896484375, 73.28787231445312, 774.5328979492188, 345.9366455078125, 310.44317626953125, -104.82245635986328, 461.56689453125, 390.00897216796875, 175.32061767578125, 155.4927978515625, -746.564697265625, 391.0625, 754.5924072265625, 394.89947509765625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000380.npy"}
{"epoch": 0.5580029368575624, "step": 381, "batch_size": 64, "mean": 198.2617645263672, "std": 347.00701904296875, "min": -717.6475830078125, "p10": -173.42606658935546, "median": 198.1874008178711, "p90": 620.1234802246095, "max": 1075.3875732421875, "pos_frac": 0.71875, "sample": [247.97372436523438, 21.765411376953125, 8.033143997192383, 633.77685546875, 568.364013671875, 686.5250244140625, 586.9136962890625, -35.26078796386719, 120.31420135498047, 506.3462829589844, 433.28887939453125, 76.78813171386719, 385.1598205566406, 298.00714111328125, 1010.190673828125, -174.37680053710938, 390.83331298828125, 438.8740234375, 688.393798828125, 306.4026184082031, 942.7628784179688, 598.5982666015625, -203.06561279296875, -44.15180969238281, -171.2076873779297, 132.07327270507812, 629.3485717773438, 73.09870910644531, -372.7536926269531, -53.334388732910156, 129.2605438232422, 286.7397155761719, -40.30293273925781, -28.927303314208984, 262.039306640625, -98.7113037109375, 244.39373779296875, 160.61618041992188, 194.43368530273438, -73.42160034179688, 59.22228240966797, 98.82901000976562, 1075.3875732421875, 226.47772216796875, 389.48785400390625, 366.974609375, -0.23199462890625, -254.60150146484375, 529.392578125, -168.33238220214844, 3.131643295288086, 419.0746765136719, -115.86288452148438, 230.6385955810547, 86.34627532958984, -317.0675354003906, 101.586181640625, 316.3414001464844, 401.6429138183594, 210.98524475097656, 201.9411163330078, -717.6475830078125, -674.5211181640625, 453.75677490234375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000381.npy"}
{"epoch": 0.5594713656387665, "step": 382, "batch_size": 64, "mean": 244.72340393066406, "std": 263.5086975097656, "min": -315.31488037109375, "p10": -99.44692840576171, "median": 195.8606719970703, "p90": 604.7596374511719, "max": 962.5030517578125, "pos_frac": 0.8125, "sample": [962.5030517578125, 607.8582763671875, 162.427978515625, 93.41780853271484, 135.29588317871094, 730.4862060546875, -118.50591278076172, 113.40506744384766, 359.577392578125, -194.69161987304688, 162.29747009277344, 243.55398559570312, 397.99322509765625, 388.28125, 477.18267822265625, 592.017822265625, 150.5903778076172, 21.013870239257812, 444.9014587402344, 143.14120483398438, 434.47698974609375, -105.25126647949219, 28.831253051757812, 340.8892517089844, 682.9603271484375, -178.49642944335938, 328.5481872558594, -85.90347290039062, 579.1900024414062, 337.5325012207031, 702.6112060546875, -149.70167541503906, -23.037044525146484, 315.36480712890625, 457.2615661621094, -315.31488037109375, -179.85720825195312, 410.22540283203125, -7.2486724853515625, 306.8990783691406, 151.6160888671875, 253.51040649414062, 322.73089599609375, 74.3922348022461, 208.32972717285156, 123.10972595214844, 158.29957580566406, 301.9258728027344, 231.84495544433594, 183.39161682128906, 597.5294799804688, 509.0022888183594, 341.0113525390625, -2.3640193939208984, 158.75079345703125, 677.280029296875, 73.38124084472656, -0.4793434143066406, 493.40203857421875, 638.8939208984375, 126.0697250366211, 151.2803955078125, 16.383407592773438, 120.278076171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000382.npy"}
{"epoch": 0.5609397944199707, "step": 383, "batch_size": 64, "mean": 252.49868774414062, "std": 366.9219970703125, "min": -949.7139892578125, "p10": -129.02317962646484, "median": 241.0999755859375, "p90": 772.7438598632813, "max": 1017.2378540039062, "pos_frac": 0.75, "sample": [-38.214622497558594, -101.3370361328125, 135.79940795898438, -316.58807373046875, 613.9470825195312, -244.9578094482422, -92.58153533935547, 640.96826171875, 368.3359069824219, -38.05512619018555, 247.47247314453125, 198.27088928222656, 68.3843994140625, -44.59016418457031, -409.578857421875, 264.58355712890625, 937.9716796875, 48.514442443847656, 279.949951171875, 298.2425537109375, 863.081787109375, -186.21682739257812, 16.25009536743164, 782.6864013671875, 202.66946411132812, 5.823753356933594, 897.1309814453125, 69.28755187988281, -18.673583984375, 321.57293701171875, 1017.2378540039062, 234.72747802734375, 773.3712158203125, 305.94598388671875, 408.18975830078125, 372.44195556640625, 449.1025695800781, 58.52996826171875, 218.42465209960938, 291.0763244628906, -131.715576171875, -34.964874267578125, 249.32901000976562, 43.19191360473633, 418.1598815917969, 721.0645141601562, -140.19656372070312, 632.7406616210938, -6.2410736083984375, 289.1710510253906, 896.3612670898438, 109.7150650024414, -122.74092102050781, 316.8370361328125, 455.14190673828125, 71.11576080322266, 771.280029296875, 601.0843505859375, 231.8621826171875, 740.8695678710938, -949.7139892578125, 83.81655883789062, 493.7493896484375, 520.8006591796875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000383.npy"}
{"epoch": 0.5624082232011748, "step": 384, "batch_size": 64, "mean": 194.50318908691406, "std": 231.10328674316406, "min": -345.15997314453125, "p10": -104.71719894409179, "median": 207.32962799072266, "p90": 447.4849304199219, "max": 871.7481079101562, "pos_frac": 0.8125, "sample": [210.13226318359375, 11.185943603515625, 871.7481079101562, -135.98960876464844, 4.41937255859375, 118.51609802246094, -189.9742889404297, 3.2910919189453125, 707.6190185546875, 226.61688232421875, 452.0632019042969, 437.13397216796875, 204.52699279785156, -198.10525512695312, 342.2962951660156, 88.01023864746094, -97.18281555175781, 66.32588195800781, 361.37054443359375, 368.7003173828125, 335.8065490722656, 287.8122253417969, 89.42024230957031, 285.37469482421875, 366.6001892089844, 235.1385955810547, 154.59747314453125, -213.24996948242188, -345.15997314453125, 78.65816497802734, 324.4061584472656, 122.46589660644531, 51.94172668457031, 248.23721313476562, 191.01498413085938, -77.02848052978516, 400.23907470703125, 324.4093933105469, 291.29315185546875, 67.96932983398438, -4.023426055908203, 218.78871154785156, 381.235107421875, 458.7611083984375, 519.1898193359375, 291.14935302734375, 338.1134033203125, 176.32672119140625, 304.9543762207031, -120.93518829345703, 112.88463592529297, -59.945213317871094, 99.26277923583984, 443.569091796875, 94.69732666015625, -107.94622039794922, -50.577754974365234, 449.16314697265625, 434.243408203125, 56.08320236206055, 619.5115966796875, 408.0853576660156, 31.24420166015625, 281.7178649902344], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000384.npy"}
{"epoch": 0.5638766519823789, "step": 385, "batch_size": 64, "mean": 238.16543579101562, "std": 252.18600463867188, "min": -417.9302673339844, "p10": -62.42099227905268, "median": 225.0621795654297, "p90": 597.3522155761721, "max": 797.0364990234375, "pos_frac": 0.859375, "sample": [-85.9561767578125, 69.46784210205078, -7.505561828613281, 228.80770874023438, 471.1694641113281, 297.34637451171875, 435.04278564453125, 164.23529052734375, -152.92202758789062, 476.02178955078125, 119.93775939941406, 6.24311637878418, 161.20455932617188, 179.20632934570312, 469.0360412597656, 255.06158447265625, 613.8515014648438, 379.740478515625, 66.32649993896484, 31.150001525878906, 344.4508361816406, 182.48558044433594, -162.12144470214844, 342.35040283203125, -237.84628295898438, 366.0682373046875, 137.06846618652344, 667.74072265625, 148.39173889160156, 29.590789794921875, 385.53411865234375, 271.95391845703125, 142.07861328125, 797.0364990234375, 770.4051513671875, 129.8920440673828, 192.381103515625, -0.8443450927734375, 558.8538818359375, 276.7967529296875, 129.4618377685547, 118.82252502441406, 550.374755859375, 155.9783172607422, 221.316650390625, 628.52197265625, 410.0500793457031, -196.79684448242188, 679.36083984375, 344.24578857421875, -417.9302673339844, 107.08882141113281, 119.24219512939453, -276.6896667480469, 271.4373779296875, 259.6527099609375, 417.1246643066406, 146.28439331054688, 311.2627258300781, 673.7997436523438, 99.62385559082031, 375.4890441894531, 242.85916137695312, 352.2748107910156], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000385.npy"}
{"epoch": 0.5653450807635829, "step": 386, "batch_size": 64, "mean": 265.89410400390625, "std": 332.1671142578125, "min": -449.0478515625, "p10": -112.92532424926756, "median": 198.95632934570312, "p90": 767.6743103027344, "max": 1057.6285400390625, "pos_frac": 0.828125, "sample": [-35.0822639465332, 389.5083312988281, -118.96190643310547, 437.95880126953125, 793.651611328125, 210.38902282714844, 108.15396881103516, 481.4001770019531, 137.50570678710938, 567.63427734375, 420.08099365234375, 1057.6285400390625, -335.49114990234375, -7.059297561645508, 774.3953247070312, 475.66058349609375, 357.5511779785156, 459.9429626464844, 64.27464294433594, 51.25672149658203, 110.50713348388672, 292.3092041015625, 260.002197265625, 142.37884521484375, 274.2529602050781, 264.0289611816406, 141.88485717773438, -98.8399658203125, 1004.5130004882812, 602.35400390625, -66.14840698242188, 121.82540893554688, 111.35775756835938, -230.46664428710938, 90.62773132324219, 751.991943359375, 988.4064331054688, 453.228271484375, 813.8709716796875, 174.46913146972656, 249.53634643554688, 620.5247802734375, -282.64404296875, 131.6065216064453, 49.646942138671875, 186.53692626953125, 154.13636779785156, 187.5236358642578, -222.96998596191406, 455.9621276855469, -449.0478515625, 802.622802734375, 45.434059143066406, 41.41881561279297, 67.90359497070312, 154.01304626464844, 496.6685791015625, 352.6524353027344, -372.89447021484375, 594.0142822265625, 529.0821533203125, 227.43043518066406, 363.34783935546875, 141.76312255859375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000386.npy"}
{"epoch": 0.566813509544787, "step": 387, "batch_size": 64, "mean": 211.16876220703125, "std": 274.27960205078125, "min": -486.3799133300781, "p10": -82.7726379394531, "median": 170.76555633544922, "p90": 533.1275024414062, "max": 895.823486328125, "pos_frac": 0.796875, "sample": [283.6276550292969, -64.83040618896484, 101.24436950683594, 126.33457946777344, 246.5767822265625, 895.823486328125, 79.45346069335938, 138.8375701904297, 278.47430419921875, 495.9198303222656, -245.42620849609375, 287.7730712890625, 90.22317504882812, 70.10005187988281, -367.531494140625, 209.11233520507812, 408.2904052734375, 528.8464965820312, -90.46216583251953, 322.8303527832031, 801.0906372070312, 51.41727828979492, 227.35977172851562, 872.0325927734375, 444.6299743652344, 117.19082641601562, 156.33486938476562, 177.76052856445312, 118.71044158935547, 143.42274475097656, 70.8227310180664, 163.7705841064453, 557.3297119140625, 25.056468963623047, 660.5234375, 147.99884033203125, 361.27978515625, 64.96940612792969, 348.7712707519531, 335.44146728515625, 24.640464782714844, -14.53034496307373, 614.7210693359375, 534.9622192382812, 390.51983642578125, 29.67083740234375, 455.2011413574219, -51.672847747802734, 521.5979614257812, -50.39796447753906, 387.23431396484375, -486.3799133300781, 59.019187927246094, 354.9144287109375, -179.82650756835938, 233.13973999023438, -9.7022705078125, 466.89166259765625, -97.37423706054688, -28.627700805664062, 282.86370849609375, -252.34622192382812, 281.8390808105469, 407.3125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000387.npy"}
{"epoch": 0.5682819383259912, "step": 388, "batch_size": 64, "mean": 296.793212890625, "std": 349.7825927734375, "min": -534.7120361328125, "p10": -57.8051773071289, "median": 222.25848388671875, "p90": 715.891735839844, "max": 1455.0760498046875, "pos_frac": 0.84375, "sample": [1455.0760498046875, 619.080078125, -84.33009338378906, 549.7941284179688, 556.0718383789062, 885.1317749023438, 139.46080017089844, -293.7112121582031, 110.69066619873047, 328.565185546875, 199.97152709960938, 995.9169921875, 591.5842895507812, 242.9118194580078, -534.7120361328125, 55.092811584472656, 320.718017578125, 293.60772705078125, 486.94580078125, 13.681007385253906, 153.60787963867188, 819.3424682617188, 186.47244262695312, -38.08722686767578, 221.9171905517578, 209.17355346679688, 108.6179428100586, -53.93525695800781, 173.96810913085938, 388.4346618652344, 102.48522186279297, 133.67845153808594, 276.2156982421875, 130.79525756835938, 477.724365234375, 172.85174560546875, 100.31370544433594, 123.51123809814453, 527.2078247070312, 129.84378051757812, 482.1388244628906, 72.67988586425781, 420.1336975097656, 148.63113403320312, 162.08514404296875, -28.577781677246094, 453.6298828125, 276.9106140136719, -238.01394653320312, 399.2622375488281, 339.3797607421875, 1353.365966796875, -198.7517852783203, 407.081298828125, 262.3382568359375, 36.484619140625, -59.463714599609375, 579.2451782226562, 744.2200927734375, 820.484375, 649.792236328125, 485.7273254394531, 222.5997772216797, -72.30393981933594], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000388.npy"}
{"epoch": 0.5697503671071953, "step": 389, "batch_size": 64, "mean": 241.64466857910156, "std": 313.03753662109375, "min": -375.909912109375, "p10": -113.8736503601074, "median": 246.52861785888672, "p90": 610.9105834960938, "max": 966.6968994140625, "pos_frac": 0.734375, "sample": [72.64041137695312, 492.3434143066406, 404.0166320800781, -356.9451599121094, 115.80177307128906, 337.11614990234375, 158.8634796142578, 612.659423828125, 430.80059814453125, 235.93255615234375, 237.95681762695312, -123.6459732055664, 307.6363220214844, -259.5730895996094, 195.77687072753906, 533.1378173828125, 144.61712646484375, 484.00531005859375, 419.1387023925781, -91.07156372070312, -293.9978942871094, -18.71523666381836, 606.8299560546875, 38.820125579833984, 296.5010070800781, 650.6942749023438, -375.909912109375, -1.4325714111328125, 104.61165618896484, 121.19734191894531, 334.8481140136719, -30.811859130859375, 140.66481018066406, 294.46734619140625, -324.8922424316406, 374.2740478515625, 281.470458984375, 495.9308166503906, 405.04473876953125, 966.6968994140625, 474.17999267578125, -68.12959289550781, 471.7321472167969, 434.73321533203125, 279.42816162109375, 383.92926025390625, -27.715333938598633, 99.22215270996094, 472.3491516113281, 181.29269409179688, 99.20368194580078, 959.0283203125, -82.59565734863281, -63.18395233154297, 240.78460693359375, 868.423828125, -50.036407470703125, 912.9965209960938, 325.34393310546875, -52.82664108276367, 364.8674011230469, -209.28839111328125, 252.2726287841797, 781.7464599609375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000389.npy"}
{"epoch": 0.5712187958883994, "step": 390, "batch_size": 64, "mean": 241.12167358398438, "std": 375.294189453125, "min": -534.5858154296875, "p10": -119.9414596557617, "median": 241.9431915283203, "p90": 574.2462036132813, "max": 2182.66845703125, "pos_frac": 0.765625, "sample": [195.76026916503906, 294.029541015625, 614.3118896484375, 510.232177734375, -534.5858154296875, -102.49764251708984, -244.17681884765625, 126.11540222167969, 411.3277282714844, 232.20343017578125, 505.5683288574219, 57.16731262207031, 340.0992736816406, 408.70196533203125, 251.68295288085938, 54.663055419921875, -44.63768768310547, 2182.66845703125, -127.4173812866211, 402.46307373046875, 113.13334655761719, 534.9384765625, 574.2940673828125, 119.32957458496094, 204.1717987060547, -71.74639892578125, 138.98826599121094, -74.93646240234375, 181.95555114746094, 296.93853759765625, 143.9623260498047, -40.877532958984375, -247.61370849609375, -64.9449234008789, 163.8505096435547, 574.134521484375, 830.6521606445312, 582.775146484375, -388.39813232421875, 314.1444396972656, 560.7806396484375, 578.8533325195312, 406.8809509277344, 117.72710418701172, 597.9130859375, 170.1555938720703, 130.71688842773438, 77.74067687988281, 554.9644165039062, 377.226318359375, 478.9527893066406, 307.24609375, 389.8835144042969, 294.95458984375, 331.5581359863281, -35.85796356201172, -42.504180908203125, -256.31103515625, -530.077880859375, 259.577392578125, 480.3297119140625, 66.16348266601562, 279.18743896484375, 417.2948913574219], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000390.npy"}
{"epoch": 0.5726872246696035, "step": 391, "batch_size": 64, "mean": 140.00616455078125, "std": 290.0155029296875, "min": -684.3495483398438, "p10": -144.41297607421873, "median": 93.2385139465332, "p90": 474.9092224121094, "max": 1119.4993896484375, "pos_frac": 0.71875, "sample": [-45.58720779418945, 10.624862670898438, 63.331932067871094, -49.91337585449219, 203.19247436523438, 325.6122131347656, 22.97977638244629, 145.063232421875, 5.984180450439453, 362.9678649902344, -135.54971313476562, 164.55055236816406, -242.0703125, -51.39183807373047, 59.84324645996094, 36.49378967285156, 275.96148681640625, 122.1368408203125, 53.12897491455078, -175.70672607421875, 164.21153259277344, 471.0799865722656, 645.5748291015625, 476.5503234863281, -109.82010650634766, 54.87016296386719, 1119.4993896484375, -189.7627410888672, -111.74661254882812, 312.03021240234375, 304.81005859375, -117.88396453857422, 83.09359741210938, 34.13058853149414, -87.96440887451172, 242.31800842285156, 103.38343048095703, -31.191429138183594, 552.1765747070312, 154.973388671875, 7.103294372558594, -133.42044067382812, 183.87374877929688, 203.65963745117188, 665.5994873046875, -184.96511840820312, 596.50341796875, -45.85204315185547, -532.226806640625, 51.482933044433594, 56.20903015136719, 300.7891540527344, -148.21151733398438, 426.5638427734375, 376.71063232421875, 63.36183166503906, 778.7164916992188, 384.8768310546875, 313.7706298828125, -684.3495483398438, 255.94793701171875, 251.8050537109375, 316.8680725097656, 233.59266662597656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000391.npy"}
{"epoch": 0.5741556534508077, "step": 392, "batch_size": 64, "mean": 280.42364501953125, "std": 281.1254577636719, "min": -242.83287048339844, "p10": -36.953772735595685, "median": 233.82681274414062, "p90": 598.7764648437502, "max": 1198.196044921875, "pos_frac": 0.859375, "sample": [38.859535217285156, 160.23025512695312, -148.65769958496094, 60.88430404663086, 219.6388702392578, -112.47232818603516, 617.5731201171875, 348.5162353515625, 294.2513122558594, 359.46368408203125, 539.3659057617188, -242.83287048339844, 279.7313232421875, 189.4637451171875, 182.6197052001953, 29.312131881713867, 624.7767333984375, 782.249267578125, 533.3016967773438, 215.9744873046875, -125.64275360107422, 423.16033935546875, 425.1860046386719, 57.70201110839844, 353.6842041015625, 342.436767578125, 196.73292541503906, 1198.196044921875, -20.33984375, 241.66452026367188, 229.0396270751953, 327.878662109375, 482.00750732421875, 190.36370849609375, 554.9176025390625, 238.61399841308594, 297.8095703125, 177.07655334472656, 200.3361053466797, 148.06002807617188, 348.537109375, 436.7503662109375, 67.19821166992188, -44.07402801513672, 541.5036010742188, 925.417724609375, 413.31719970703125, 445.5497131347656, 396.1774597167969, -146.810302734375, 130.905517578125, 126.95567321777344, 102.70733642578125, 407.79833984375, 207.57855224609375, 156.83071899414062, 305.67095947265625, 41.59959030151367, 1031.996337890625, -2.6451873779296875, -145.35671997070312, 83.06591796875, 392.7148742675781, 812.591552734375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000392.npy"}
{"epoch": 0.5756240822320118, "step": 393, "batch_size": 64, "mean": 216.95748901367188, "std": 347.7161560058594, "min": -296.16937255859375, "p10": -204.67606506347656, "median": 239.49549865722656, "p90": 602.6011962890628, "max": 1579.2958984375, "pos_frac": 0.6875, "sample": [271.545166015625, 728.98193359375, -35.02402114868164, -123.8451919555664, 421.34246826171875, 485.701904296875, -194.30233764648438, 411.54852294921875, 264.41839599609375, 243.78939819335938, 21.35601806640625, 268.9619140625, 180.63259887695312, 367.59027099609375, 270.9674072265625, 42.89636993408203, 1092.078125, 964.58984375, 658.84033203125, -7.381439208984375, 206.44773864746094, 498.76806640625, -236.8955078125, -296.16937255859375, -104.7503662109375, -11.774120330810547, -218.42221069335938, 269.6156311035156, 297.6968994140625, 235.20159912109375, 246.23220825195312, 447.783935546875, 460.71014404296875, 234.650390625, 527.1209106445312, 293.0995178222656, 252.27938842773438, 145.01828002929688, 413.8177185058594, -44.53987121582031, -15.082077026367188, 156.39788818359375, 634.9498901367188, -190.83377075195312, -234.8966064453125, 508.5921630859375, 320.1074523925781, -209.1219482421875, 16.30156707763672, 165.08474731445312, 146.94403076171875, 255.61248779296875, -190.0905303955078, -143.9410400390625, -210.11322021484375, 402.7334289550781, 432.68212890625, -277.2902526855469, 312.2091064453125, 17.995567321777344, -119.47126007080078, 1579.2958984375, 679.8123779296875, -103.17816162109375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000393.npy"}
{"epoch": 0.5770925110132159, "step": 394, "batch_size": 64, "mean": 234.6966094970703, "std": 302.21087646484375, "min": -296.38214111328125, "p10": -114.81839218139648, "median": 172.23265075683594, "p90": 658.4907348632814, "max": 1087.570068359375, "pos_frac": 0.75, "sample": [106.99783325195312, 345.5948486328125, -189.77699279785156, 392.24713134765625, 189.33416748046875, -116.58984375, 88.99705505371094, 120.94256591796875, 535.609375, 108.0821533203125, 578.715576171875, -19.53907012939453, -296.38214111328125, 67.48863220214844, -40.066497802734375, -148.04959106445312, -110.68500518798828, 826.5327758789062, 672.427001953125, -162.37290954589844, 500.53912353515625, -238.41796875, 290.5189208984375, 275.7641906738281, 197.99815368652344, 362.4661865234375, 209.8520050048828, 58.04664611816406, -61.16857147216797, 625.9727783203125, 560.6568603515625, 749.5048217773438, -89.60130310058594, 392.53143310546875, 152.39947509765625, 49.15472412109375, 151.50387573242188, -1.1192092895507812, 138.5819091796875, 675.7734375, 20.80005645751953, 522.4014892578125, -4.376232147216797, 222.8905487060547, 97.04820251464844, -24.22064208984375, 503.8562927246094, 916.193603515625, 248.46897888183594, 145.87100219726562, 526.5799560546875, 318.14654541015625, 1087.570068359375, 906.7985229492188, 339.39227294921875, 50.46019744873047, 75.22019958496094, -116.60530090332031, 237.79598999023438, 393.0739440917969, 200.80801391601562, -69.12446594238281, 155.13113403320312, 315.93768310546875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000394.npy"}
{"epoch": 0.57856093979442, "step": 395, "batch_size": 64, "mean": 253.37454223632812, "std": 337.18341064453125, "min": -533.7626953125, "p10": -104.96495361328124, "median": 224.07679748535156, "p90": 767.2992919921876, "max": 1005.1070556640625, "pos_frac": 0.765625, "sample": [492.0422668457031, -268.7701416015625, 242.25262451171875, 69.21537017822266, 218.11805725097656, 324.69976806640625, 289.9062194824219, 248.99734497070312, -302.41363525390625, 254.42762756347656, -37.10847473144531, 564.3197631835938, 490.35943603515625, -8.841255187988281, 118.89290618896484, -51.93022537231445, 732.361328125, 333.2549133300781, 935.5853271484375, -60.67681884765625, 623.4901733398438, 13.948341369628906, 516.0458374023438, -111.82603454589844, 298.4046325683594, 4.594696044921875, 209.41818237304688, 427.2306823730469, -70.60408782958984, 27.434165954589844, 856.0963134765625, 990.6569213867188, 502.46795654296875, 131.05923461914062, 350.1060791015625, 32.450531005859375, 166.22808837890625, 291.754638671875, 940.0421142578125, 186.098388671875, 488.9798583984375, 519.109619140625, -533.7626953125, -122.78059387207031, -245.61373901367188, 315.48883056640625, -67.79978942871094, 861.542724609375, 17.07743263244629, -88.95576477050781, 130.87625122070312, -191.00997924804688, 72.14311981201172, 1005.1070556640625, 782.272705078125, 230.03553771972656, 305.9107666015625, 541.2243041992188, 149.5753936767578, 490.8727111816406, 437.9352111816406, 67.32567596435547, 123.98511505126953, -43.35857009887695], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000395.npy"}
{"epoch": 0.580029368575624, "step": 396, "batch_size": 64, "mean": 224.01565551757812, "std": 255.9213104248047, "min": -430.2630615234375, "p10": -35.674264907836914, "median": 175.3723373413086, "p90": 526.6750274658204, "max": 1008.4512939453125, "pos_frac": 0.875, "sample": [196.9877166748047, 542.962158203125, 36.55211639404297, 191.83792114257812, 128.63427734375, 556.6514282226562, 143.20037841796875, 117.93246459960938, 146.50689697265625, 192.93661499023438, 368.3088684082031, 319.0166931152344, 224.41531372070312, 409.5169982910156, -130.70599365234375, 188.68922424316406, 184.30874633789062, 58.86247253417969, 71.58943176269531, 152.7279052734375, 144.967041015625, 412.22869873046875, 973.8207397460938, 335.8314208984375, 121.20413208007812, 461.8811340332031, 335.3759765625, 182.2490234375, 155.24737548828125, 684.4251098632812, -35.2735481262207, 390.8229675292969, 69.66878509521484, 267.9355773925781, 62.54035568237305, 168.4956512451172, 399.1041259765625, -169.11517333984375, 430.43359375, 661.594482421875, 354.4752502441406, 40.50445556640625, -44.9156494140625, 488.6717224121094, 11.83648681640625, -76.68707275390625, 425.06341552734375, -35.84600067138672, 11.937772750854492, 230.57614135742188, 865.4070434570312, 94.23566436767578, 44.69451904296875, -95.27197265625, 145.09361267089844, 249.9955291748047, 315.4736022949219, 199.94094848632812, 90.97516632080078, -430.2630615234375, 97.95072937011719, 1008.4512939453125, 37.7908935546875, 152.54214477539062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000396.npy"}
{"epoch": 0.5814977973568282, "step": 397, "batch_size": 64, "mean": 297.37628173828125, "std": 297.7142639160156, "min": -340.5064697265625, "p10": -1.9820903778076167, "median": 257.7900848388672, "p90": 691.567791748047, "max": 1057.0762939453125, "pos_frac": 0.875, "sample": [31.398624420166016, 556.0360107421875, 40.77435302734375, 380.58563232421875, 156.444091796875, 107.69215393066406, 237.80624389648438, 294.987060546875, 531.0197143554688, 199.43626403808594, 724.609375, 82.14091491699219, 550.168212890625, -121.32659912109375, 157.15997314453125, 449.3495178222656, 328.4827880859375, -187.40573120117188, 439.54150390625, 614.438720703125, 321.4323425292969, 722.4818115234375, 326.95098876953125, 445.9867248535156, 1052.9425048828125, -1.5591888427734375, 695.3099365234375, 453.2693786621094, -172.69662475585938, 125.20286560058594, 545.7703857421875, 244.32518005371094, 130.58831787109375, 883.5224609375, 443.325927734375, 162.32785034179688, 27.454334259033203, 402.13909912109375, 134.5528564453125, 176.49534606933594, -340.5064697265625, -213.21151733398438, 33.23723602294922, -246.64227294921875, 532.1572265625, 191.5532989501953, 104.10926055908203, 137.92279052734375, -2.1633338928222656, 603.2448120117188, 52.0213737487793, 682.8361206054688, 245.30551147460938, 1057.0762939453125, 183.76034545898438, 855.2652587890625, 128.5893096923828, 270.274658203125, 507.9532470703125, 174.9506072998047, 309.67034912109375, 291.40277099609375, 455.09820556640625, 325.01416015625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000397.npy"}
{"epoch": 0.5829662261380323, "step": 398, "batch_size": 64, "mean": 268.0833740234375, "std": 332.5855712890625, "min": -289.31201171875, "p10": -58.144466400146484, "median": 175.58631134033203, "p90": 649.4923461914062, "max": 1261.4759521484375, "pos_frac": 0.765625, "sample": [181.20654296875, 85.73393249511719, 480.4750671386719, 641.4508056640625, -195.63209533691406, 7.584114074707031, 348.21221923828125, -206.57711791992188, 93.48265075683594, -65.34764862060547, 559.4746704101562, -283.92156982421875, 652.938720703125, 412.8890380859375, 436.5561828613281, 1034.826416015625, 169.96607971191406, 68.44945526123047, 679.593505859375, 302.03338623046875, 512.7239990234375, 100.66537475585938, 499.98797607421875, 475.0380554199219, 1228.588623046875, -13.859039306640625, 146.90826416015625, 236.2991485595703, 633.4111938476562, 413.91278076171875, 27.79779052734375, 204.967529296875, -54.9249267578125, -17.752853393554688, 279.6376037597656, 117.08161926269531, 83.8980941772461, 378.46875, 41.07685089111328, 325.79718017578125, 118.62248229980469, 513.5687255859375, 139.09857177734375, 77.76129913330078, 137.61465454101562, -59.524269104003906, 484.16015625, -40.271270751953125, -42.69178771972656, 92.48995971679688, 320.3503112792969, 1261.4759521484375, 604.6229248046875, 832.8327026367188, -289.31201171875, 536.43408203125, 455.04071044921875, 456.7847595214844, 692.7481689453125, 129.7838134765625, -52.05622863769531, -38.40826416015625, -46.256752014160156, -150.651611328125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000398.npy"}
{"epoch": 0.5844346549192364, "step": 399, "batch_size": 64, "mean": 233.65927124023438, "std": 335.41021728515625, "min": -452.8985595703125, "p10": -166.02588806152343, "median": 193.15814208984375, "p90": 696.7158447265625, "max": 1058.63427734375, "pos_frac": 0.734375, "sample": [57.190513610839844, 304.65875244140625, 93.6841812133789, -202.96417236328125, 690.990234375, -72.45189666748047, 705.3861083984375, 30.476757049560547, 147.95785522460938, -49.330657958984375, 292.7695617675781, -169.15377807617188, 699.169677734375, -59.56475830078125, 660.07275390625, 164.6831817626953, 369.2603759765625, -143.35794067382812, -96.77381134033203, 54.386234283447266, 704.9927978515625, 96.00715637207031, 400.4449462890625, 253.75096130371094, 544.6800537109375, 844.6754150390625, 158.80154418945312, 421.92889404296875, -322.2652587890625, 214.10028076171875, 16.132522583007812, 410.5880432128906, 182.79244995117188, 172.09872436523438, 403.7981262207031, -238.7525634765625, 1058.63427734375, 541.489501953125, 658.8911743164062, -188.6265411376953, 301.71148681640625, 486.5571594238281, 192.89845275878906, -452.8985595703125, -52.236812591552734, 459.46575927734375, -28.150693893432617, 867.23583984375, 32.73229217529297, -155.83486938476562, 437.4113464355469, 949.0631713867188, 502.49420166015625, 172.5856170654297, 563.3770141601562, 193.41783142089844, 210.91366577148438, 253.4183349609375, -395.4169006347656, 262.81280517578125, 433.5114440917969, -95.93995666503906, 162.5401611328125, -158.72747802734375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000399.npy"}
{"epoch": 0.5859030837004405, "step": 400, "batch_size": 64, "mean": 345.60205078125, "std": 317.1341857910156, "min": -491.4385986328125, "p10": -36.726124572753875, "median": 322.3320617675781, "p90": 719.3317199707031, "max": 1111.6796875, "pos_frac": 0.84375, "sample": [240.62460327148438, 566.5947875976562, 505.45367431640625, 250.962158203125, 680.169921875, -130.93441772460938, 557.148193359375, 216.97140502929688, -58.54137420654297, 660.5148315429688, 189.21542358398438, 4.900579452514648, 518.766357421875, 303.25555419921875, -491.4385986328125, 714.9457397460938, 741.485107421875, 134.75274658203125, 440.227294921875, 413.1639404296875, 865.3900146484375, 568.4857788085938, 101.41653442382812, 478.9262390136719, 66.96501922607422, 265.896484375, 126.26164245605469, 572.159423828125, 299.97454833984375, 362.0564880371094, 377.6036071777344, 197.3182373046875, -94.25629425048828, 427.36517333984375, 103.9709243774414, -255.0922393798828, -8.351966857910156, 790.5318603515625, 315.9345397949219, 575.4705810546875, -9.204116821289062, 215.3809814453125, 292.06280517578125, 975.9232177734375, 73.085205078125, 622.7200927734375, 250.91098022460938, 150.28758239746094, 18.258991241455078, 1111.6796875, 721.21142578125, -7.235809326171875, 319.71954345703125, 492.045166015625, 394.90411376953125, 658.0849609375, -257.09320068359375, 700.967529296875, 665.5487670898438, 422.77362060546875, 324.944580078125, 697.4090576171875, 736.404052734375, -48.521270751953125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000400.npy"}
{"epoch": 0.5873715124816447, "step": 401, "batch_size": 64, "mean": 199.28952026367188, "std": 409.0959777832031, "min": -891.3428344726562, "p10": -238.3463012695312, "median": 153.03118896484375, "p90": 789.4362121582034, "max": 1214.486083984375, "pos_frac": 0.6875, "sample": [208.5518035888672, 696.1639404296875, -27.104961395263672, 195.11859130859375, 491.53826904296875, 51.02696990966797, -296.5257873535156, -386.0986633300781, -891.3428344726562, -73.10374450683594, 136.31607055664062, 285.7539978027344, -197.1312255859375, 55.140953063964844, -29.361907958984375, -32.647438049316406, 343.80877685546875, 313.5745544433594, 146.14431762695312, 53.96498107910156, -197.44436645507812, -149.73045349121094, -83.52417755126953, 1021.36865234375, 42.289222717285156, 739.7061767578125, 1013.1810913085938, 495.95703125, 810.7490844726562, 204.85865783691406, 100.57979583740234, 299.95135498046875, -43.52865219116211, 61.14765167236328, -100.73287200927734, -514.4571533203125, 1037.1656494140625, -91.54176330566406, 246.6318817138672, 501.8798522949219, 188.4286651611328, 566.3746337890625, 366.1604919433594, -68.95783996582031, 267.274169921875, 159.91806030273438, 176.4726104736328, -291.31610107421875, 1214.486083984375, -553.4802856445312, 694.9420166015625, 470.0474853515625, 381.5145568847656, 111.64476776123047, 853.0399169921875, 896.7769775390625, 17.85132598876953, 306.3852233886719, 29.25384521484375, 57.80949020385742, 196.68743896484375, 703.3899536132812, -255.87570190429688, -172.59312438964844], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000401.npy"}
{"epoch": 0.5888399412628488, "step": 402, "batch_size": 64, "mean": 363.5229187011719, "std": 318.2452697753906, "min": -147.6934814453125, "p10": 14.432830047607425, "median": 313.2002716064453, "p90": 862.8498657226564, "max": 1421.703369140625, "pos_frac": 0.90625, "sample": [495.296875, 29.350317001342773, 13.235496520996094, 448.22784423828125, 253.58792114257812, 413.63995361328125, -7.2524261474609375, 313.9927673339844, -8.192083358764648, 736.0907592773438, 363.45281982421875, 598.3827514648438, 759.8286743164062, 217.2450714111328, -69.76226806640625, 645.4559936523438, 64.59449768066406, 295.24017333984375, 204.88211059570312, 409.1319580078125, 578.4328002929688, 218.80990600585938, 582.7914428710938, -147.6934814453125, 389.54931640625, 578.8779296875, 105.10104370117188, 935.5198974609375, 39.42759704589844, 899.2965087890625, 159.04342651367188, 287.0477294921875, 150.9936981201172, 487.1842041015625, 100.43848419189453, 185.23764038085938, 972.3494873046875, 130.33438110351562, 142.8557586669922, 689.1098022460938, 17.226608276367188, -95.41322326660156, 881.3341064453125, 983.3077392578125, 319.7127990722656, 414.11688232421875, 819.719970703125, 92.11276245117188, 341.0488586425781, 1421.703369140625, 457.19232177734375, 589.4800415039062, 64.41679382324219, 898.1011962890625, 242.20904541015625, 312.40777587890625, 339.242431640625, 308.81024169921875, 178.04977416992188, -132.04196166992188, 389.1895751953125, 219.22671508789062, 211.53726196289062, 331.6394348144531], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000402.npy"}
{"epoch": 0.5903083700440529, "step": 403, "batch_size": 64, "mean": 264.1478271484375, "std": 298.76361083984375, "min": -270.63726806640625, "p10": -81.19561691284179, "median": 237.39065551757812, "p90": 645.9346740722657, "max": 1140.186279296875, "pos_frac": 0.8125, "sample": [237.52059936523438, 22.343116760253906, 88.10828399658203, 216.63128662109375, 309.0835266113281, 100.46366882324219, 237.26071166992188, 390.2674560546875, -69.0173568725586, -257.1969909667969, 166.03762817382812, -16.795135498046875, -270.63726806640625, 515.0762939453125, 103.71903991699219, 837.6748046875, 253.59141540527344, 62.597434997558594, 442.048583984375, 255.7411346435547, 394.0560302734375, 108.37014770507812, 744.227294921875, -54.55426788330078, -101.14971160888672, 593.6909790039062, 527.1264038085938, 677.3942260742188, 243.1254425048828, 903.1851806640625, -86.41487121582031, 447.5376281738281, 612.2783813476562, 46.399898529052734, 48.23900604248047, 461.0378723144531, -150.866455078125, 234.18992614746094, 624.4219360351562, 434.3045654296875, -191.47752380371094, 321.2416076660156, 1140.186279296875, 26.163869857788086, 586.4053344726562, 135.50538635253906, 310.0516052246094, 655.1544189453125, 388.2400817871094, 935.95849609375, 267.096435546875, 458.3214111328125, 294.3352355957031, 137.36817932128906, -8.208894729614258, 33.71437072753906, 81.78658294677734, 354.5846252441406, -9.785598754882812, 101.51716613769531, 346.23907470703125, 214.69158935546875, -186.40731811523438, 181.66012573242188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000403.npy"}
{"epoch": 0.591776798825257, "step": 404, "batch_size": 64, "mean": 259.7740173339844, "std": 373.9214172363281, "min": -603.4934692382812, "p10": -251.50473480224608, "median": 268.85416412353516, "p90": 795.938232421875, "max": 1268.6878662109375, "pos_frac": 0.765625, "sample": [-346.14044189453125, 293.3376770019531, 60.982872009277344, 549.049560546875, 600.1516723632812, 169.4093017578125, 175.9711456298828, 962.27734375, 148.18865966796875, 794.22412109375, 1268.6878662109375, 219.77345275878906, 283.7813720703125, 899.3521728515625, -138.9876708984375, -369.2841491699219, 889.5872192382812, 93.99523162841797, -305.5545959472656, 116.21453094482422, -89.67276000976562, -67.4076156616211, -42.06144714355469, 453.31475830078125, 355.72113037109375, 350.52520751953125, -162.5831298828125, 361.4775085449219, 253.9269561767578, 357.1037292480469, 543.3679809570312, -236.5376739501953, 412.13494873046875, 160.56832885742188, 634.6387939453125, 120.55269622802734, -198.70091247558594, 163.29246520996094, 294.3092041015625, 486.18634033203125, -603.4934692382812, 433.4669189453125, 81.39291381835938, 642.7196044921875, 170.98175048828125, -265.93505859375, -257.919189453125, 544.8492431640625, 651.09619140625, -105.87074279785156, 374.7143859863281, 46.299278259277344, 394.6554870605469, 342.00823974609375, 495.29266357421875, 816.3921508789062, 336.5513610839844, 59.51625061035156, 1.7723140716552734, -270.9058837890625, 885.0076293945312, 160.88446044921875, 796.6728515625, 380.2146911621094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000404.npy"}
{"epoch": 0.593245227606461, "step": 405, "batch_size": 64, "mean": 214.91424560546875, "std": 385.647216796875, "min": -454.5104675292969, "p10": -293.034065246582, "median": 170.03092193603516, "p90": 798.2381896972659, "max": 1194.79296875, "pos_frac": 0.6875, "sample": [858.2332153320312, 821.849365234375, 239.63253784179688, -213.5997772216797, -445.4364013671875, -46.236083984375, -430.947509765625, 88.94027709960938, 395.1977233886719, -203.9337158203125, 55.16393280029297, 1136.947998046875, 545.11328125, 113.17544555664062, 115.27802276611328, 1194.79296875, -4.254274368286133, -77.7057876586914, 69.95541381835938, -138.6785125732422, 25.55597686767578, 271.0909423828125, 165.6437530517578, 438.26055908203125, 174.4180908203125, 135.73846435546875, -91.44282531738281, 548.905029296875, -347.16314697265625, -55.27025604248047, 403.4849548339844, -20.384103775024414, 384.4381103515625, -9.583826065063477, 411.35186767578125, 91.87068939208984, 254.336669921875, -327.07733154296875, 29.567861557006836, 11.132434844970703, 554.0433349609375, 227.68521118164062, 723.2755126953125, 352.47259521484375, -13.11417007446289, 178.89923095703125, 619.330322265625, -171.51858520507812, 250.37734985351562, 296.85791015625, 411.63214111328125, -346.93798828125, 220.5030059814453, 848.4017333984375, 77.2101821899414, -454.5104675292969, 499.58367919921875, 212.23309326171875, -70.995361328125, 946.8983154296875, 919.3272094726562, -351.71929931640625, 513.070068359375, 743.1454467773438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000405.npy"}
{"epoch": 0.5947136563876652, "step": 406, "batch_size": 64, "mean": 274.11834716796875, "std": 324.7137451171875, "min": -324.9259338378906, "p10": -45.76034698486327, "median": 206.734130859375, "p90": 732.7014160156253, "max": 1158.044189453125, "pos_frac": 0.796875, "sample": [400.7078857421875, 155.854736328125, 134.53549194335938, 104.50810241699219, 203.23956298828125, 192.03070068359375, 176.63983154296875, -324.9259338378906, 1096.21630859375, 158.46807861328125, 435.5607604980469, 225.38902282714844, 1158.044189453125, 499.802734375, 295.83380126953125, -26.518360137939453, 510.8336181640625, 276.9024353027344, 483.31103515625, -61.07239532470703, 844.7079467773438, 368.0651550292969, 361.28680419921875, -115.29707336425781, 79.88590240478516, 176.6406707763672, -17.050628662109375, -50.062137603759766, 480.4140625, 19.761505126953125, -26.309341430664062, 214.3804168701172, 474.1443176269531, -315.41326904296875, -206.083740234375, 505.46270751953125, 825.5491943359375, 192.39312744140625, 136.1708221435547, 2.6577930450439453, 204.5914306640625, 208.8768310546875, 416.9352111816406, -35.722835540771484, -11.986846923828125, 91.7952651977539, 772.8018798828125, 342.86956787109375, 35.487762451171875, 226.57830810546875, 639.1336669921875, 631.3423461914062, 369.9373474121094, 78.84733581542969, -287.3701171875, 14.451873779296875, 860.4414672851562, 557.1459350585938, 1028.3502197265625, 393.2053527832031, 370.24957275390625, 507.29583740234375, 97.27159118652344, -15.620750427246094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000406.npy"}
{"epoch": 0.5961820851688693, "step": 407, "batch_size": 64, "mean": 177.56365966796875, "std": 302.35662841796875, "min": -800.078125, "p10": -234.55619506835933, "median": 157.4922637939453, "p90": 576.367529296875, "max": 781.433349609375, "pos_frac": 0.78125, "sample": [-278.1705627441406, 522.1063232421875, 492.9031677246094, -262.66156005859375, -250.7095489501953, -26.680198669433594, 284.2862548828125, -139.29281616210938, -168.53021240234375, 164.80990600585938, 424.30255126953125, 155.97584533691406, 312.56182861328125, 723.0850219726562, 634.272216796875, 45.12665557861328, 19.279972076416016, 476.982666015625, 99.25122833251953, 0.0492706298828125, 369.7161560058594, 139.668701171875, 109.36408996582031, 73.49075317382812, 97.04350280761719, 451.57391357421875, 257.0564880371094, 215.88949584960938, 84.44654846191406, 100.53001403808594, -248.62100219726562, 219.99298095703125, 12.777641296386719, -105.89002227783203, 46.479637145996094, 124.80625915527344, 419.6548156738281, 711.134521484375, 105.01231384277344, -800.078125, 678.3363037109375, 355.7084045410156, 362.56060791015625, 11.541481018066406, 484.36639404296875, 168.0129852294922, 190.3428955078125, 781.433349609375, 349.9701232910156, 273.8913879394531, 571.7742919921875, -201.73831176757812, -76.25370788574219, 578.3360595703125, -28.70661163330078, 87.03760528564453, 159.00868225097656, 536.22021484375, 15.870529174804688, -368.38482666015625, 601.4759521484375, 289.79132080078125, -249.95399475097656, 180.43707275390625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000407.npy"}
{"epoch": 0.5976505139500734, "step": 408, "batch_size": 64, "mean": 196.42950439453125, "std": 249.96670532226562, "min": -295.85491943359375, "p10": -110.58454284667968, "median": 194.82819366455078, "p90": 538.9203582763674, "max": 808.3211669921875, "pos_frac": 0.75, "sample": [76.66038513183594, 639.2188110351562, 808.3211669921875, 7.178955078125, 436.62811279296875, 353.6019287109375, -76.6492919921875, 171.98277282714844, -29.61358642578125, -125.19654846191406, -5.044782638549805, -295.85491943359375, 59.05420684814453, -4.75238037109375, -3.6714096069335938, -186.9872589111328, 116.48979187011719, 293.11029052734375, 265.5522155761719, 51.87107849121094, 216.15501403808594, 318.17144775390625, -69.7105712890625, 391.7881774902344, -125.04216003417969, 165.08230590820312, 8.312459945678711, 236.22335815429688, 80.74230194091797, 667.8402709960938, 230.43565368652344, 202.78921508789062, 469.7979431152344, 3.004547119140625, -275.36376953125, -104.20449829101562, 271.8911437988281, 270.57293701171875, 56.29728698730469, 174.8128662109375, 381.5176086425781, 469.5542907714844, 351.54931640625, 448.12005615234375, 186.86717224121094, -237.96835327148438, 622.8026733398438, -50.83246612548828, -3.499675750732422, 62.291107177734375, 113.43406677246094, 568.5442504882812, 248.09280395507812, -113.31884765625, 340.6292419433594, 294.7595520019531, 79.62995147705078, 451.910400390625, 274.92279052734375, 334.4019775390625, 686.6116333007812, 600.1066284179688, 448.0494384765625, 271.81707763671875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000408.npy"}
{"epoch": 0.5991189427312775, "step": 409, "batch_size": 64, "mean": 196.50201416015625, "std": 283.4256286621094, "min": -431.8522644042969, "p10": -97.62911834716797, "median": 171.47945404052734, "p90": 515.269580078125, "max": 902.524169921875, "pos_frac": 0.765625, "sample": [-45.661956787109375, -335.1167907714844, 290.74188232421875, 226.02273559570312, 196.95875549316406, 404.369140625, -45.680458068847656, -97.96588134765625, -0.45635986328125, 127.38499450683594, 516.4342041015625, 512.4047241210938, -204.1876220703125, 179.572998046875, 479.6507873535156, 155.54583740234375, 223.04458618164062, 269.4212951660156, 205.2255859375, 889.349609375, 329.05096435546875, -159.74380493164062, 334.14105224609375, 44.51523971557617, 690.7156372070312, -148.24664306640625, 10.641899108886719, 47.54656982421875, -46.376556396484375, 158.55967712402344, 290.52294921875, 689.3248901367188, 494.98248291015625, -431.8522644042969, 47.48035430908203, 902.524169921875, -41.221763610839844, 96.94000244140625, -380.748779296875, 171.77493286132812, 494.6185607910156, -75.6866683959961, 512.5521240234375, 140.6641387939453, 187.82968139648438, 230.0042724609375, 80.83682250976562, 148.7474365234375, 760.4439086914062, 448.66485595703125, -68.62052154541016, 180.58799743652344, -96.84333801269531, 192.6598663330078, 171.18397521972656, 438.3103942871094, 324.6158447265625, 322.7431640625, 37.327850341796875, 135.8051300048828, 751.619384765625, 9.651138305664062, 83.4232177734375, 117.40045928955078], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000409.npy"}
{"epoch": 0.6005873715124816, "step": 410, "batch_size": 64, "mean": 235.4065704345703, "std": 278.1690979003906, "min": -496.79534912109375, "p10": -71.02486724853514, "median": 233.8155288696289, "p90": 557.5462036132812, "max": 876.74462890625, "pos_frac": 0.84375, "sample": [230.72161865234375, 298.87017822265625, -89.84602355957031, 78.01969146728516, 558.1627197265625, 332.331787109375, 514.8358154296875, 527.855712890625, 402.6562194824219, 11.816926956176758, 141.0159149169922, 219.32611083984375, 11.892578125, 518.8738403320312, -102.68424224853516, -75.888427734375, 637.07763671875, -449.26763916015625, 19.018672943115234, 308.48431396484375, 420.08837890625, 737.6788330078125, -59.67655944824219, 61.20539855957031, 556.107666015625, 684.206298828125, 370.7436218261719, -32.58127212524414, 339.90399169921875, 876.74462890625, 225.85235595703125, 25.554454803466797, 290.87738037109375, 8.658294677734375, 378.47076416015625, 88.37240600585938, -41.96442413330078, 13.280939102172852, 133.6751708984375, -146.00048828125, 205.9123077392578, 350.89007568359375, 161.05282592773438, 468.9674072265625, 205.222900390625, -496.79534912109375, 277.9423522949219, 236.90943908691406, 717.9638061523438, 311.69610595703125, 108.747802734375, 111.9177017211914, 418.77642822265625, 466.8233337402344, 250.1785125732422, 790.3609008789062, 194.02381896972656, 402.2845458984375, -360.77984619140625, 115.30522155761719, 290.9286804199219, 33.60832977294922, 481.5929260253906, 298.0190124511719], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000410.npy"}
{"epoch": 0.6020558002936858, "step": 411, "batch_size": 64, "mean": 234.29080200195312, "std": 317.9166564941406, "min": -577.058837890625, "p10": -67.29966201782224, "median": 178.2654800415039, "p90": 672.8669738769531, "max": 1319.6495361328125, "pos_frac": 0.78125, "sample": [218.72938537597656, -8.368782043457031, 29.70435333251953, 74.3795394897461, 573.617431640625, 89.11458587646484, 65.392822265625, 158.4320831298828, 322.66552734375, 401.08349609375, 232.47122192382812, -45.84577178955078, 200.81961059570312, 7.988956451416016, 25.236318588256836, -19.153871536254883, 366.4195861816406, 332.4917907714844, 95.42337036132812, 448.8903503417969, 791.0072631835938, 305.7521667480469, -117.9552001953125, -22.651914596557617, 74.00858306884766, 147.93931579589844, 177.38258361816406, 222.4709930419922, 128.93753051757812, 1032.051513671875, 146.6387939453125, 384.6827697753906, 813.0459594726562, -79.01953887939453, 551.1650390625, 40.36339569091797, 329.7463073730469, 675.6079711914062, 666.4713134765625, 164.71347045898438, 69.2323989868164, 385.77301025390625, 179.438720703125, 244.41647338867188, 148.62664794921875, 319.1153259277344, -6.719219207763672, -577.058837890625, 1319.6495361328125, 287.4202880859375, 229.0210723876953, 325.8360900878906, 640.0054931640625, 728.4945068359375, 506.85797119140625, -179.82733154296875, -20.07050323486328, -275.446533203125, 179.14837646484375, -41.549591064453125, 764.8311767578125, -158.01821899414062, 0.0774383544921875, -76.49418640136719], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000411.npy"}
{"epoch": 0.6035242290748899, "step": 412, "batch_size": 64, "mean": 240.63748168945312, "std": 317.7256774902344, "min": -580.2222290039062, "p10": -99.05812225341796, "median": 212.33866119384766, "p90": 610.5163269042969, "max": 1143.5924072265625, "pos_frac": 0.75, "sample": [-176.29421997070312, 211.75579833984375, 465.7906494140625, 442.0162353515625, 370.62115478515625, -147.98992919921875, 610.3687133789062, -12.093635559082031, 175.2010498046875, 602.4049072265625, 59.223968505859375, 324.4361877441406, 96.5142822265625, 398.14752197265625, 97.78782653808594, 712.1167602539062, 364.7113342285156, 83.7006607055664, -580.2222290039062, 194.57583618164062, 85.44336700439453, 45.598548889160156, 529.5301513671875, 212.92152404785156, 108.8331298828125, 337.6686096191406, -36.395172119140625, 347.3154602050781, -161.28097534179688, 234.36636352539062, -64.09447479248047, 57.665992736816406, 211.25918579101562, 357.2542724609375, 146.5873565673828, -24.14000129699707, -290.96722412109375, -83.29560852050781, 116.25859832763672, 947.1597900390625, 68.37857055664062, -483.1307373046875, 468.19549560546875, 601.884033203125, 643.3689575195312, -104.42042541503906, 526.16259765625, 717.3610229492188, -11.77587890625, 473.0078430175781, 369.5813903808594, -67.0604248046875, 134.45326232910156, 610.57958984375, 588.2003173828125, 1143.5924072265625, -57.18637466430664, 355.40655517578125, 413.3966064453125, -86.54608154296875, 402.37274169921875, 226.61688232421875, 611.8199462890625, 486.0779724121094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000412.npy"}
{"epoch": 0.604992657856094, "step": 413, "batch_size": 64, "mean": 230.0081329345703, "std": 294.12408447265625, "min": -812.5547485351562, "p10": -139.653343963623, "median": 271.5061798095703, "p90": 546.4365478515625, "max": 897.3558349609375, "pos_frac": 0.859375, "sample": [8.558685302734375, 58.1890754699707, 125.63861083984375, 345.5924377441406, 84.745361328125, 38.45207214355469, 251.47463989257812, 176.90655517578125, 65.8213882446289, 208.09024047851562, 544.4019165039062, 870.653076171875, -105.77143096923828, 270.4227600097656, -15.720054626464844, 10.912736892700195, 419.63885498046875, 178.93759155273438, 83.60523986816406, 727.712158203125, 120.7098388671875, 549.2017822265625, 408.6091003417969, -812.5547485351562, 79.78536987304688, 287.6826171875, 313.41619873046875, 146.12896728515625, 547.3085327148438, 897.3558349609375, 291.62335205078125, 305.8528137207031, 540.4893798828125, 350.32958984375, -223.72665405273438, 494.28485107421875, 386.79510498046875, 161.089111328125, 298.98187255859375, 299.2759094238281, 87.67375183105469, 272.589599609375, 599.5135498046875, -262.4479675292969, 531.2555541992188, 411.6644592285156, 287.5884704589844, -300.8550720214844, -432.86053466796875, -257.6317138671875, 168.7845916748047, 416.40716552734375, 358.5279541015625, 513.4781494140625, 10.157756805419922, 399.37640380859375, 278.2373352050781, 563.040771484375, 233.9825439453125, 429.512939453125, 153.1114501953125, 183.83262634277344, 438.85406494140625, -154.17416381835938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000413.npy"}
{"epoch": 0.6064610866372981, "step": 414, "batch_size": 64, "mean": 266.597900390625, "std": 301.786376953125, "min": -692.3080444335938, "p10": -19.803091049194332, "median": 256.3713073730469, "p90": 712.1410644531253, "max": 962.745849609375, "pos_frac": 0.875, "sample": [331.02899169921875, 279.2057800292969, 292.87493896484375, 158.7661895751953, 20.74515151977539, 857.1737670898438, 153.4527587890625, -692.3080444335938, 119.38304138183594, 310.10968017578125, 373.552001953125, 262.8468017578125, 625.6559448242188, 908.3538818359375, 295.77008056640625, -164.62074279785156, 289.9036865234375, 75.98727416992188, -66.34603118896484, 83.22681427001953, 224.29855346679688, 493.2572021484375, -21.459308624267578, 229.71951293945312, 491.9703369140625, 337.37530517578125, -111.50057983398438, 2.1414623260498047, 438.1900634765625, 69.83546447753906, 748.1885986328125, 417.97503662109375, 4.107074737548828, 222.39588928222656, 781.72265625, 130.61109924316406, 425.62396240234375, -15.938583374023438, -220.78736877441406, 277.337890625, 286.31182861328125, 237.85882568359375, 623.335693359375, -226.48025512695312, 86.99778747558594, 510.96771240234375, 31.756057739257812, 628.0301513671875, 67.46012115478516, 15.641918182373047, 295.6058349609375, 188.91062927246094, 167.12176513671875, 130.8794708251953, 249.89581298828125, 294.60552978515625, 830.25341796875, 38.951324462890625, 499.965576171875, 869.7639770507812, 962.745849609375, 363.0721435546875, 116.13677978515625, 352.6507568359375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000414.npy"}
{"epoch": 0.6079295154185022, "step": 415, "batch_size": 64, "mean": 203.53948974609375, "std": 292.0166015625, "min": -769.774169921875, "p10": -119.43657684326172, "median": 192.7205047607422, "p90": 618.1854003906253, "max": 972.5262451171875, "pos_frac": 0.765625, "sample": [360.1033020019531, -138.00286865234375, 333.5429382324219, 699.8404541015625, -60.66654968261719, 725.0850830078125, 548.3533935546875, 189.58216857910156, -7.955863952636719, 170.01759338378906, 456.4841613769531, 347.23272705078125, 393.96612548828125, 350.92974853515625, -122.93379974365234, -64.25639343261719, 205.96295166015625, -214.55734252929688, 377.68267822265625, 208.2017059326172, 6.920551300048828, 75.97695922851562, 4.745361328125, 696.425048828125, -121.01741790771484, 143.33880615234375, 719.804443359375, -62.355018615722656, 99.46790313720703, 541.4127807617188, -68.26370239257812, 381.6819152832031, 289.1183776855469, 63.93743896484375, 972.5262451171875, -250.3043670654297, -68.64024353027344, 115.20268249511719, 462.0443115234375, 648.1134033203125, 79.96586608886719, 320.1867980957031, 155.55975341796875, 36.03192138671875, 71.18270874023438, 195.8588409423828, -21.713964462280273, 279.7569580078125, 471.1879577636719, -161.58551025390625, 268.354248046875, 395.6773376464844, 271.3680725097656, 10.32525634765625, 339.7872314453125, 5.71452522277832, 245.71383666992188, 43.385345458984375, 420.5045166015625, -115.7479476928711, 672.853759765625, 122.44291687011719, -769.774169921875, 280.7422180175781], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000415.npy"}
{"epoch": 0.6093979441997063, "step": 416, "batch_size": 64, "mean": 308.8902893066406, "std": 274.02001953125, "min": -176.76724243164062, "p10": -28.499741363525356, "median": 273.4059295654297, "p90": 670.0021789550782, "max": 1031.5556640625, "pos_frac": 0.890625, "sample": [812.4059448242188, 185.41168212890625, 179.28009033203125, 437.44000244140625, -147.11712646484375, 555.2899780273438, 227.94122314453125, 74.46111297607422, 508.60455322265625, 665.1007690429688, 108.68477630615234, 96.60440063476562, 73.81754302978516, 321.3972473144531, 279.2807922363281, 464.75091552734375, 282.01483154296875, 39.164085388183594, 537.1194458007812, 216.7019805908203, 465.1329040527344, 484.5539245605469, 105.94486999511719, 239.03640747070312, 571.5701293945312, 158.26849365234375, 334.292236328125, 808.8831176757812, 65.41635131835938, 751.5960083007812, 273.30584716796875, 580.64794921875, 937.1724243164062, -176.76724243164062, -147.89434814453125, 51.41338348388672, 531.5630493164062, 486.834228515625, 5.138053894042969, 672.102783203125, 10.026359558105469, 1031.5556640625, 273.5060119628906, 180.8639373779297, 499.21929931640625, 362.8968811035156, 312.35675048828125, 525.1671752929688, 226.40357971191406, 88.03705596923828, 51.63047409057617, 250.81019592285156, -80.3658447265625, 621.5512084960938, 340.36846923828125, -58.18450927734375, 347.54620361328125, -60.67366409301758, 204.05963134765625, -42.91593933105469, 124.8242416381836, 255.96844482421875, 483.27178955078125, 704.4910888671875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000416.npy"}
{"epoch": 0.6108663729809104, "step": 417, "batch_size": 64, "mean": 247.69085693359375, "std": 249.3746337890625, "min": -245.95916748046875, "p10": -80.532608795166, "median": 211.9332733154297, "p90": 594.0760498046875, "max": 895.9957275390625, "pos_frac": 0.859375, "sample": [48.82239532470703, 21.191940307617188, -4.897237777709961, 277.21429443359375, 8.431529998779297, 554.1646728515625, 25.152114868164062, 199.93695068359375, -145.12197875976562, 682.8595581054688, 29.91348648071289, 581.8397216796875, 436.46282958984375, 120.86476135253906, 194.98863220214844, -85.51347351074219, -91.95201110839844, 692.173095703125, 649.2223510742188, 40.962432861328125, 599.3201904296875, 286.1275939941406, 670.9470825195312, 132.18780517578125, 543.8099975585938, 895.9957275390625, 365.20758056640625, 241.51168823242188, 242.81069946289062, 343.5627136230469, 102.70221710205078, 107.38304138183594, -68.91059112548828, 125.69080352783203, 299.924560546875, 494.5093994140625, 89.76126098632812, 350.38702392578125, 414.9774169921875, 188.45655822753906, 73.30940246582031, 357.9234924316406, 253.67103576660156, 315.4993591308594, -86.88955688476562, 216.798583984375, 160.46035766601562, 129.47805786132812, 380.314453125, 324.84088134765625, 471.90740966796875, -234.397705078125, 128.79318237304688, 357.588623046875, 16.379188537597656, -245.95916748046875, 207.06796264648438, 133.86248779296875, 620.17919921875, 446.9591064453125, 543.18798828125, 158.52647399902344, 546.8006591796875, -87.16619110107422], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000417.npy"}
{"epoch": 0.6123348017621145, "step": 418, "batch_size": 64, "mean": 267.42822265625, "std": 358.40191650390625, "min": -434.45941162109375, "p10": -160.05644226074216, "median": 212.7575454711914, "p90": 653.6790954589844, "max": 1224.0572509765625, "pos_frac": 0.765625, "sample": [227.8624267578125, -142.74197387695312, 299.2685241699219, -0.22638320922851562, 132.9922332763672, 945.8648071289062, 68.66033172607422, 32.659427642822266, 575.9642944335938, 557.3964233398438, -17.47967529296875, -46.595458984375, 379.94915771484375, 29.088172912597656, 1173.4691162109375, 527.5775146484375, 352.0, -167.85011291503906, 482.2568054199219, 91.1605224609375, -188.305908203125, 633.3897705078125, -167.4769287109375, 541.05517578125, 233.84805297851562, 10.395797729492188, 295.8766174316406, 597.3853759765625, 585.8707275390625, 206.61306762695312, 509.4859619140625, 67.37052917480469, 124.04015350341797, 848.7216796875, -434.45941162109375, -93.43090057373047, -170.99740600585938, 329.10101318359375, 218.9020233154297, 463.1092529296875, 1224.0572509765625, -55.98065948486328, 48.3990478515625, 131.83273315429688, 273.9534912109375, 161.51145935058594, 190.2333984375, 186.64097595214844, -302.10302734375, -279.2994384765625, 378.99505615234375, 102.05384826660156, 586.588134765625, -131.8953094482422, 658.239501953125, 304.31494140625, 643.0381469726562, 1052.27001953125, 150.10577392578125, 16.222393035888672, 560.9111328125, 894.15234375, 272.2955017089844, -62.90016174316406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000418.npy"}
{"epoch": 0.6138032305433186, "step": 419, "batch_size": 64, "mean": 285.44049072265625, "std": 299.9181213378906, "min": -446.49298095703125, "p10": -95.69550781249995, "median": 276.1434326171875, "p90": 661.4927734375001, "max": 934.590087890625, "pos_frac": 0.859375, "sample": [733.2576904296875, -446.49298095703125, 376.8783874511719, 533.2670288085938, 144.26426696777344, 619.8594970703125, -115.52951049804688, -9.535541534423828, 26.96533203125, 621.819091796875, 769.1083984375, 228.68661499023438, 279.61334228515625, 511.9980773925781, 429.5507507324219, 500.3647155761719, 226.15847778320312, 177.39453125, -404.85687255859375, 83.10932159423828, 14.115425109863281, 739.5535888671875, -187.1375274658203, 272.67352294921875, 543.1216430664062, -320.59552001953125, 166.06991577148438, 185.9702606201172, 437.35076904296875, 797.536376953125, 355.0627746582031, 108.78582763671875, 415.5274963378906, 289.2532958984375, 449.88134765625, 111.53597259521484, 140.04685974121094, 82.3052978515625, 449.8978576660156, 289.33563232421875, 560.353759765625, 190.39190673828125, 647.015625, 43.91694259643555, 85.03651428222656, 144.9460906982422, 58.574432373046875, -143.83221435546875, 667.697265625, 633.9197998046875, 586.0377197265625, 373.90576171875, 99.7598876953125, 934.590087890625, 398.2355651855469, 400.39959716796875, 452.80291748046875, 263.9573669433594, 112.27981567382812, 806.374267578125, -140.51426696777344, -49.416168212890625, 170.7069091796875, 344.8799743652344], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000419.npy"}
{"epoch": 0.6152716593245228, "step": 420, "batch_size": 64, "mean": 172.54815673828125, "std": 368.50567626953125, "min": -758.49072265625, "p10": -250.86905822753903, "median": 132.55504608154297, "p90": 582.9258850097657, "max": 1154.662841796875, "pos_frac": 0.65625, "sample": [-186.9365234375, 161.73666381835938, 35.223350524902344, 19.69552230834961, 547.6087646484375, -125.95977020263672, -231.34713745117188, -20.657135009765625, 220.37001037597656, 592.275634765625, -24.00398063659668, -7.286674499511719, -54.8021125793457, 178.82130432128906, -22.5050048828125, 100.33870697021484, 530.310546875, 296.19915771484375, 73.00663757324219, 223.15194702148438, -50.25299072265625, 258.6008605957031, -259.235595703125, 346.5578308105469, -65.42971801757812, 135.25552368164062, -32.32728576660156, 467.6534118652344, 72.5479965209961, -85.32089233398438, 535.6183471679688, 440.71600341796875, 408.38262939453125, -294.88970947265625, 129.8545684814453, 241.3881072998047, -269.61578369140625, 49.39405059814453, 664.2862548828125, 719.40478515625, -37.387657165527344, -758.49072265625, 617.1165161132812, 537.9158325195312, 561.1098022460938, 152.3524932861328, 450.2457580566406, 443.90362548828125, 379.08447265625, 319.43572998046875, 9.797159194946289, -508.3009948730469, 240.8752899169922, 29.08316421508789, -32.8447151184082, 1096.20263671875, -649.8571166992188, 97.31123352050781, 175.5377655029297, 248.3035888671875, 1121.52978515625, -58.525352478027344, -263.8071594238281, 1154.662841796875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000420.npy"}
{"epoch": 0.6167400881057269, "step": 421, "batch_size": 64, "mean": 202.6715087890625, "std": 296.26580810546875, "min": -762.3468017578125, "p10": -187.41248474121087, "median": 247.33277130126953, "p90": 572.1623413085938, "max": 796.1580810546875, "pos_frac": 0.78125, "sample": [321.6288146972656, 103.10301208496094, 160.5620574951172, -132.12750244140625, 276.0800476074219, 459.0135803222656, 763.4953002929688, -51.923492431640625, 83.96870422363281, 85.88551330566406, 312.48590087890625, -1.5307350158691406, 559.5673828125, 655.5697631835938, 448.67315673828125, 615.6925048828125, 312.54510498046875, -220.58656311035156, 128.4650115966797, 140.21121215820312, 750.1749267578125, 473.79058837890625, 141.80975341796875, -211.10604858398438, 251.25343322753906, 241.40794372558594, 404.15966796875, 280.4561767578125, 35.144996643066406, 236.76515197753906, 577.5601806640625, 373.4852600097656, 591.0670776367188, 265.2338562011719, 328.1632385253906, -496.43817138671875, 363.139892578125, 331.2681884765625, -400.0478820800781, 526.3925170898438, 243.412109375, -19.15529441833496, 335.3463134765625, 796.1580810546875, 253.00216674804688, 323.548095703125, 183.5933380126953, 2.5985240936279297, 40.685768127441406, 468.30047607421875, 107.98088073730469, -3.6523056030273438, 252.3280487060547, -762.3468017578125, -217.2119903564453, 109.79985046386719, 90.00152587890625, -115.13681030273438, -100.43084716796875, 270.174072265625, 228.8002471923828, -271.59063720703125, 327.8758850097656, 342.4356689453125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000421.npy"}
{"epoch": 0.618208516886931, "step": 422, "batch_size": 64, "mean": 233.17080688476562, "std": 316.6456298828125, "min": -680.056396484375, "p10": -101.36484909057616, "median": 208.27392578125, "p90": 655.4627868652344, "max": 1156.4371337890625, "pos_frac": 0.78125, "sample": [278.97235107421875, 688.2593383789062, 778.8301391601562, 528.5460205078125, 305.64276123046875, -215.3097686767578, -399.80657958984375, 651.0399780273438, -39.64708709716797, 1156.4371337890625, -72.74732971191406, 168.2473602294922, 107.56382751464844, 50.84613037109375, 393.2046203613281, 192.66641235351562, 63.437583923339844, -104.73043060302734, -42.797115325927734, 195.82203674316406, 55.25146484375, 699.1776123046875, 325.986328125, 85.384765625, 462.2351989746094, 162.3980712890625, 121.377197265625, 657.3582763671875, 180.48812866210938, 582.9044799804688, 754.536865234375, 359.501708984375, 403.2983703613281, 211.98582458496094, 321.19384765625, 333.1412048339844, 347.3043212890625, -402.52911376953125, 231.46682739257812, 29.116851806640625, -25.221393585205078, 354.54595947265625, 194.77005004882812, -680.056396484375, -66.21923828125, 565.913330078125, 476.203125, 96.80381774902344, 665.5548095703125, -38.11045837402344, 378.70263671875, 204.56202697753906, 51.41595458984375, 68.89704895019531, 484.64324951171875, 435.3002014160156, 329.66162109375, -93.51182556152344, 454.0907897949219, 494.3140563964844, -204.19393920898438, 155.18614196777344, 227.17877197265625, -213.5542449951172], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000422.npy"}
{"epoch": 0.6196769456681351, "step": 423, "batch_size": 64, "mean": 258.01800537109375, "std": 284.5015869140625, "min": -455.45379638671875, "p10": -53.29781188964843, "median": 214.84877014160156, "p90": 687.3826477050782, "max": 975.906982421875, "pos_frac": 0.828125, "sample": [200.67074584960938, 96.37377166748047, 147.3848419189453, 16.9969482421875, -99.27151489257812, -274.8385314941406, 763.7866821289062, 251.95388793945312, 595.6616821289062, 412.1700134277344, 697.05078125, -63.144309997558594, 726.28515625, 238.11492919921875, 480.0004577636719, -56.277896881103516, 728.1175537109375, 736.5775756835938, 805.3341064453125, 664.8236694335938, -22.39080810546875, 39.9744873046875, 349.84185791015625, 205.00827026367188, 510.07635498046875, 217.38751220703125, 101.02027130126953, 48.783721923828125, -455.45379638671875, -98.74520874023438, 356.57232666015625, -46.34428024291992, 228.44131469726562, 326.5346984863281, 604.25634765625, 222.04449462890625, 30.887617111206055, 465.88623046875, -159.3215789794922, 19.379241943359375, 288.40875244140625, 541.1177978515625, 212.31002807617188, 314.8697204589844, 975.906982421875, 194.97434997558594, 336.3689270019531, 10.898269653320312, 459.12725830078125, 29.937881469726562, 22.603099822998047, 242.1881561279297, 548.9401245117188, 385.7174987792969, 192.5767822265625, 142.0305633544922, -30.265838623046875, 208.83670043945312, 188.31332397460938, -23.611907958984375, 561.1036987304688, 203.73391723632812, 125.71047973632812, 369.7465515136719], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000423.npy"}
{"epoch": 0.6211453744493393, "step": 424, "batch_size": 64, "mean": 259.2139892578125, "std": 318.65765380859375, "min": -296.0882568359375, "p10": -138.28556747436522, "median": 226.84656524658203, "p90": 617.9068115234378, "max": 1146.7628173828125, "pos_frac": 0.796875, "sample": [-264.2662048339844, 358.17626953125, 32.92070007324219, -68.18829345703125, 16.366943359375, 541.7427368164062, 434.5497131347656, 309.9811096191406, 107.13248443603516, -271.3953857421875, 273.9694519042969, 289.0204162597656, -47.35865783691406, -221.4570770263672, 343.45306396484375, 694.075927734375, 473.9041748046875, 740.9634399414062, 140.3115692138672, 408.4992370605469, 554.8450317382812, 223.3862762451172, 529.946044921875, 550.3560180664062, 352.1409606933594, -125.9101333618164, 225.80589294433594, 132.40866088867188, -196.9119873046875, -45.90622329711914, 152.06056213378906, 1086.140869140625, -143.58932495117188, 275.1537170410156, 513.5346069335938, -53.58184814453125, 359.77349853515625, -279.37738037109375, 227.88723754882812, 139.77713012695312, -26.383142471313477, 126.38495635986328, -296.0882568359375, 59.96011734008789, 278.9367980957031, 365.65625, 212.93045043945312, 81.12012481689453, 382.2590637207031, 1101.66357421875, 357.0063171386719, 644.9332885742188, 438.7634582519531, 214.47332763671875, 381.8048095703125, 100.26739501953125, 840.6771850585938, 164.30300903320312, 191.11436462402344, 1146.7628173828125, 501.7662048339844, 287.74664306640625, 168.4935760498047, 94.80194091796875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000424.npy"}
{"epoch": 0.6226138032305433, "step": 425, "batch_size": 64, "mean": 180.99131774902344, "std": 321.116943359375, "min": -785.3352661132812, "p10": -171.45906219482418, "median": 202.31488037109375, "p90": 544.6693725585938, "max": 890.82666015625, "pos_frac": 0.765625, "sample": [-118.32807922363281, 42.06813049316406, 499.8468933105469, -409.939453125, -687.2981567382812, 186.90936279296875, 193.15480041503906, 291.2779846191406, 83.48269653320312, 256.16339111328125, -104.72374725341797, 508.69647216796875, -785.3352661132812, -62.96110916137695, 36.87642288208008, 59.98974609375, 90.53433990478516, 96.7183837890625, 521.722412109375, 127.81227111816406, 273.9260559082031, 179.87139892578125, -27.867961883544922, 324.99761962890625, 595.4852294921875, -1.9916915893554688, 452.40948486328125, 53.16175842285156, 10.397350311279297, 771.2918701171875, 129.32888793945312, 543.241943359375, 611.5675659179688, -187.1064910888672, 197.4154815673828, 262.4902648925781, 549.0028686523438, 207.2142791748047, -4.8607177734375, 663.29736328125, -530.874267578125, 394.5643310546875, 116.63829040527344, -317.2215881347656, 545.2811279296875, 363.5887451171875, 307.0353698730469, 227.24468994140625, 7.1712646484375, 357.36651611328125, 255.46893310546875, 248.33859252929688, -126.23158264160156, -244.49383544921875, 274.06890869140625, 344.154296875, 517.1741943359375, 890.82666015625, 499.2139587402344, 303.33880615234375, 466.7587890625, -134.94839477539062, 263.75048828125, 125.2900619506836], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000425.npy"}
{"epoch": 0.6240822320117474, "step": 426, "batch_size": 64, "mean": 245.62155151367188, "std": 296.7994384765625, "min": -636.2515869140625, "p10": -71.47893905639646, "median": 212.54100036621094, "p90": 656.1594848632815, "max": 851.6878051757812, "pos_frac": 0.78125, "sample": [299.5901184082031, 101.3248062133789, 91.71552276611328, 184.19406127929688, 229.71585083007812, 75.81697082519531, 62.41278076171875, 214.92178344726562, 743.3412475585938, 92.38316345214844, 176.5232391357422, 212.31488037109375, -6.858768463134766, -2.035940170288086, 612.433837890625, -191.01678466796875, 49.12150573730469, 241.40567016601562, 538.2197875976562, 212.76712036132812, -87.14493560791016, 690.1888427734375, -130.42507934570312, 228.28530883789062, 124.93605041503906, 154.26312255859375, -82.19361877441406, 373.6987609863281, 851.6878051757812, 502.0281677246094, 232.90834045410156, 581.6856689453125, 68.33460235595703, 262.6427917480469, 162.7843780517578, 808.884765625, -30.17608642578125, -279.9493408203125, 165.28125, 580.432373046875, -39.82177734375, 330.97265625, 785.3946533203125, 212.19122314453125, 408.56317138671875, 23.010032653808594, 365.7904357910156, 518.8253173828125, -25.3843994140625, 129.46119689941406, -217.73382568359375, 341.8423156738281, -8.46890640258789, 486.5341491699219, 520.3041381835938, 357.9345703125, -46.47801971435547, 481.2427062988281, 783.45703125, -636.2515869140625, 574.4541625976562, 674.8990478515625, 38.137176513671875, 544.4589233398438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000426.npy"}
{"epoch": 0.6255506607929515, "step": 427, "batch_size": 64, "mean": 188.287353515625, "std": 271.7843017578125, "min": -342.47528076171875, "p10": -151.17285614013673, "median": 177.32465362548828, "p90": 450.51180419921883, "max": 1034.0205078125, "pos_frac": 0.765625, "sample": [361.7773742675781, 31.345378875732422, 211.34268188476562, 419.70269775390625, 759.7626953125, 114.71398162841797, -27.326242446899414, 17.43689727783203, 370.51226806640625, 242.54879760742188, -22.043609619140625, 51.85945129394531, 147.45523071289062, 151.4436798095703, 177.5572967529297, 337.62646484375, 385.1550598144531, 718.9813232421875, 46.67334747314453, -299.32928466796875, 31.70146942138672, 178.81100463867188, 271.50018310546875, 330.77081298828125, 154.80288696289062, 354.13262939453125, 337.8778991699219, 869.4186401367188, 41.39079284667969, 165.72244262695312, -185.74197387695312, 461.5231628417969, 350.66571044921875, -149.48385620117188, -58.34352493286133, 379.9239196777344, 392.42840576171875, -57.58065414428711, 521.1896362304688, 197.02374267578125, 300.8699951171875, 145.11940002441406, 52.82311248779297, 177.09201049804688, -342.47528076171875, 198.85794067382812, 396.869873046875, 384.75738525390625, -269.295166015625, -283.6005554199219, 492.9736328125, 221.31109619140625, 424.8186340332031, 16.31969451904297, 1034.0205078125, -52.462982177734375, 59.63426208496094, 330.2740783691406, 336.54632568359375, 103.41368103027344, -38.611106872558594, -256.0090026855469, -15.888938903808594, -151.89671325683594], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000427.npy"}
{"epoch": 0.6270190895741556, "step": 428, "batch_size": 64, "mean": 245.55508422851562, "std": 372.8175964355469, "min": -780.3504638671875, "p10": -182.9501953125, "median": 276.37872314453125, "p90": 706.1985595703126, "max": 1106.099609375, "pos_frac": 0.734375, "sample": [-54.35526657104492, -185.25430297851562, -56.115577697753906, 434.8267517089844, 844.3485107421875, 404.6231384277344, 121.97496032714844, 591.1363525390625, 419.6312255859375, 1106.099609375, -191.25149536132812, 75.78713989257812, 511.1657409667969, 229.28555297851562, 427.12274169921875, -21.58282470703125, -547.6883544921875, 598.750732421875, 365.82843017578125, 657.774169921875, 563.7103881835938, 415.678955078125, -100.66018676757812, 17.58883285522461, 406.3406677246094, 694.681884765625, 356.06292724609375, 274.2847900390625, -143.57901000976562, 346.2818298339844, -46.019737243652344, 325.91717529296875, 715.4337158203125, -56.32775115966797, 278.47265625, 33.73002624511719, 603.86376953125, 524.4010620117188, 43.01030349731445, 99.49029541015625, 97.6121826171875, 157.14309692382812, -2.3193435668945312, -177.57394409179688, -780.3504638671875, -157.79640197753906, 65.49049377441406, 151.09909057617188, 1022.7527465820312, 197.95904541015625, 468.932861328125, 433.48052978515625, 556.106201171875, -674.6962280273438, 571.0064086914062, 311.27398681640625, 715.4454956054688, 749.9222412109375, 359.7528076171875, 206.5255889892578, 711.13427734375, -212.73890686035156, 106.13658142089844, -245.24368286132812], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000428.npy"}
{"epoch": 0.6284875183553598, "step": 429, "batch_size": 64, "mean": 192.90322875976562, "std": 330.3487854003906, "min": -637.6224975585938, "p10": -216.41407928466796, "median": 206.9274673461914, "p90": 617.3160827636721, "max": 855.4842529296875, "pos_frac": 0.78125, "sample": [33.280174255371094, 108.048583984375, -406.8106689453125, 308.689208984375, -91.63343048095703, 672.5068969726562, -114.12407684326172, 157.7129669189453, 127.76568603515625, 361.1304931640625, -484.98272705078125, 203.30078125, 149.14825439453125, 123.4953842163086, -200.37789916992188, 291.4942932128906, 135.95272827148438, -300.64031982421875, 674.3009033203125, -187.8747100830078, 823.765625, 855.4842529296875, 33.62091064453125, 429.9200744628906, 267.357177734375, 56.00229263305664, 323.0132751464844, 22.578533172607422, 630.9913330078125, 510.75189208984375, 268.5678405761719, 455.0941162109375, 373.6883544921875, -637.6224975585938, 19.947349548339844, -222.16639709472656, -60.78099822998047, 202.1973114013672, 218.54824829101562, -45.397178649902344, 585.4071655273438, 325.5736083984375, 261.26556396484375, 48.095672607421875, 788.4310302734375, 338.9970397949219, 520.1541748046875, -202.99200439453125, 210.5541534423828, -408.9898986816406, 216.72396850585938, 192.26197814941406, 537.409912109375, 224.5889892578125, 77.38072967529297, -460.8328552246094, 13.991744995117188, 283.67071533203125, 55.021339416503906, 584.2171630859375, 395.8238830566406, 771.9091796875, 460.6744384765625, 440.52484130859375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000429.npy"}
{"epoch": 0.6299559471365639, "step": 430, "batch_size": 64, "mean": 195.6875, "std": 264.0730895996094, "min": -389.0096435546875, "p10": -105.67976150512695, "median": 177.05226135253906, "p90": 563.7300842285158, "max": 869.1809692382812, "pos_frac": 0.78125, "sample": [236.44451904296875, 169.76319885253906, 21.814552307128906, -107.00839233398438, 522.6131591796875, 412.9977722167969, 24.369354248046875, 246.99496459960938, 189.28668212890625, 462.44036865234375, 31.116962432861328, -34.565025329589844, 17.399147033691406, -105.93704986572266, 210.21058654785156, 581.3516235351562, -344.1520080566406, 361.8331604003906, 73.14424133300781, 106.09182739257812, 204.75807189941406, 151.35806274414062, 124.17573547363281, 184.23733520507812, -40.246192932128906, 177.6997528076172, 338.4905090332031, 325.4143371582031, -84.11717224121094, 199.4656524658203, 80.14088439941406, -119.74080657958984, 869.1809692382812, 51.62895202636719, 117.39099884033203, 642.7415161132812, -46.80146026611328, -389.0096435546875, 23.23347282409668, 235.6315155029297, 457.01318359375, 683.923583984375, 433.3076171875, 171.56304931640625, 143.72848510742188, 691.999755859375, -105.07942199707031, -170.0640869140625, 435.0411071777344, 127.61357116699219, 176.40476989746094, -46.7896842956543, 429.2565002441406, 203.57997131347656, -46.26889419555664, 365.42236328125, 337.38055419921875, -249.0264892578125, 337.0695495605469, 802.7178955078125, 678.4096069335938, 308.71417236328125, 19.44312286376953, 216.7976531982422], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000430.npy"}
{"epoch": 0.631424375917768, "step": 431, "batch_size": 64, "mean": 286.8751525878906, "std": 306.9805908203125, "min": -411.384765625, "p10": -138.5743385314941, "median": 262.3029327392578, "p90": 615.7204711914063, "max": 981.966552734375, "pos_frac": 0.828125, "sample": [-157.7078094482422, 126.03096008300781, -188.54319763183594, 504.0898742675781, 246.23812866210938, 127.36136627197266, 821.956298828125, 115.12220764160156, 405.8253173828125, 570.46142578125, 226.37948608398438, 426.1794738769531, 54.08807373046875, -159.48388671875, -93.92957305908203, 178.4877166748047, 52.345062255859375, -411.384765625, 495.29095458984375, -207.2718963623047, 558.653076171875, 563.4266967773438, 981.966552734375, 877.6243286132812, 111.76011657714844, 605.4329833984375, 414.59674072265625, 393.5860595703125, 290.7249755859375, 404.5341796875, 637.3927001953125, 253.35281372070312, 241.76324462890625, 61.72578048706055, 473.97259521484375, -405.7287902832031, 908.3026123046875, -12.582351684570312, 254.26922607421875, -59.788780212402344, 509.6631164550781, 261.63946533203125, 956.4180908203125, -32.62702178955078, 98.74964904785156, 283.72125244140625, 370.39031982421875, 8.156534194946289, 251.85171508789062, 324.51898193359375, 362.7906799316406, 531.75830078125, 620.12939453125, 405.97552490234375, 502.5180969238281, 480.2835388183594, 229.27066040039062, 130.04550170898438, -177.20977783203125, 262.9664001464844, 572.6763305664062, 145.94500732421875, 136.5880584716797, 437.26922607421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000431.npy"}
{"epoch": 0.6328928046989721, "step": 432, "batch_size": 64, "mean": 287.7674560546875, "std": 298.71331787109375, "min": -248.5356903076172, "p10": -11.864456939697247, "median": 243.34576416015625, "p90": 722.4133728027346, "max": 1306.46923828125, "pos_frac": 0.890625, "sample": [73.82279968261719, 225.92874145507812, 328.11456298828125, 254.95468139648438, 214.4099578857422, 247.38243103027344, 85.3335952758789, 61.0040168762207, 243.92715454101562, 57.33088684082031, 14.161640167236328, 25.828968048095703, 581.986083984375, 474.26312255859375, -61.37971878051758, 1306.46923828125, 756.4039916992188, 242.76437377929688, 6.5604400634765625, 146.59039306640625, 90.66061401367188, 110.35104370117188, 245.98300170898438, 13.922821044921875, 485.1533203125, 293.6556701660156, 767.4140625, 164.3634490966797, 446.1683349609375, 44.31451416015625, 242.7310791015625, -19.760841369628906, -248.5356903076172, 236.50099182128906, -162.11322021484375, 574.5410766601562, 187.19395446777344, -142.0178985595703, 391.9998474121094, 100.62269592285156, 231.83541870117188, -144.09298706054688, 249.002197265625, 410.2076416015625, 286.4044494628906, 410.6187438964844, 596.4872436523438, 429.99603271484375, 426.3865051269531, 274.19140625, 152.9957733154297, 499.5775451660156, 24.681167602539062, 933.333984375, 297.79541015625, 138.03439331054688, 691.7614135742188, 466.0787353515625, 634.72216796875, 928.27783203125, -177.68878173828125, 746.9341430664062, 65.02042388916016, 735.5499267578125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000432.npy"}
{"epoch": 0.6343612334801763, "step": 433, "batch_size": 64, "mean": 188.17587280273438, "std": 347.1685791015625, "min": -849.5542602539062, "p10": -161.5725311279297, "median": 184.10688018798828, "p90": 586.1453796386719, "max": 788.6044921875, "pos_frac": 0.734375, "sample": [-19.731536865234375, 484.9010009765625, 529.2117919921875, 395.79547119140625, 348.9434509277344, -123.11366271972656, 328.09521484375, -845.3508911132812, 471.8123474121094, 633.6661987304688, -817.5742797851562, 1.1351165771484375, 613.6693115234375, 470.0145568847656, 142.7537841796875, -45.50883865356445, 103.19532775878906, 330.36419677734375, 590.9060668945312, -849.5542602539062, -18.932109832763672, 476.9374694824219, 677.0433959960938, 294.49993896484375, 124.71340942382812, 190.34593200683594, 51.12553405761719, 177.86782836914062, 1.3042755126953125, 453.1539306640625, 173.29876708984375, 788.6044921875, -310.3423767089844, 503.503173828125, 300.7626953125, 459.7047119140625, 448.63958740234375, -248.76605224609375, 285.3880615234375, -52.66167449951172, 706.60400390625, 575.037109375, 174.39453125, -130.38201904296875, 418.75848388671875, 44.330101013183594, 419.2281188964844, -6.220664978027344, -162.93406677246094, -257.67041015625, 605.2554321289062, 405.2827453613281, 106.697998046875, 28.531574249267578, 125.94143676757812, -158.39561462402344, 8.075569152832031, 437.9502258300781, -40.387725830078125, -3.6290664672851562, 357.7696838378906, 458.41253662109375, 146.9651336669922, 263.81903076171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000433.npy"}
{"epoch": 0.6358296622613803, "step": 434, "batch_size": 64, "mean": 228.21133422851562, "std": 386.4264221191406, "min": -432.5175476074219, "p10": -151.63961029052734, "median": 167.02930450439453, "p90": 719.0665466308598, "max": 2123.761474609375, "pos_frac": 0.75, "sample": [502.2204284667969, 117.3937759399414, 2123.761474609375, 239.6907501220703, -163.4490203857422, 165.18283081054688, 188.42059326171875, 494.4738464355469, 51.646949768066406, -63.40855407714844, 280.8225402832031, -85.53819274902344, 102.06145477294922, 759.337646484375, 10.029438018798828, 84.11767578125, -99.86022186279297, -432.5175476074219, 432.4465026855469, 625.1006469726562, 791.6495361328125, -172.24929809570312, 168.8757781982422, 217.80398559570312, 277.35906982421875, 176.4913787841797, -48.80937957763672, -61.766693115234375, 211.42994689941406, 268.3934631347656, 11.05975341796875, 245.0245361328125, 410.6844787597656, -145.34056091308594, -1.5910491943359375, 524.5447387695312, 92.78494262695312, 20.109529495239258, 219.75048828125, 228.63522338867188, -43.028656005859375, 477.8157653808594, 617.10009765625, 64.32652282714844, 127.6517105102539, 215.84249877929688, 402.5285339355469, 94.52964782714844, 136.79257202148438, 792.4314575195312, 308.54296875, -104.5445556640625, 869.3529052734375, -190.55409240722656, 78.66886901855469, 46.652984619140625, 1025.5164794921875, 185.8485107421875, -154.33920288085938, -305.79296875, 860.416015625, 523.864013671875, 13.075698852539062, -203.94509887695312], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000434.npy"}
{"epoch": 0.6372980910425844, "step": 435, "batch_size": 64, "mean": 258.73687744140625, "std": 335.90203857421875, "min": -709.6453857421875, "p10": -86.82174072265623, "median": 229.52776336669922, "p90": 687.7577514648439, "max": 1230.86767578125, "pos_frac": 0.828125, "sample": [1230.86767578125, 454.59783935546875, -709.6453857421875, 415.4363098144531, 584.4021606445312, 198.1173095703125, 177.25454711914062, 487.38848876953125, 321.9963073730469, 113.57151794433594, 592.0233764648438, 64.46797180175781, 34.98591613769531, 252.08737182617188, -95.00032043457031, 579.5613403320312, 1003.4270629882812, 792.17333984375, 282.5561218261719, -143.89756774902344, 11.570831298828125, 6.33453369140625, -9.086868286132812, 242.1119842529297, -136.99212646484375, 137.73875427246094, 216.2109832763672, 143.09365844726562, -109.09467315673828, 213.74771118164062, 307.6166687011719, 108.62303161621094, 243.30596923828125, 473.8744201660156, 580.4844360351562, 366.3222961425781, 360.8288269042969, -483.8232421875, 423.1675109863281, -312.4588623046875, 238.58047485351562, 264.7698669433594, 757.681640625, 14.762115478515625, 65.66046905517578, -31.899883270263672, 618.2861328125, 648.77197265625, 14.810306549072266, -33.2674560546875, 913.5736083984375, 45.09453582763672, 226.55087280273438, 99.35934448242188, 147.39425659179688, 704.4659423828125, 781.197265625, 303.420654296875, 232.50465393066406, 504.4642028808594, 141.67955017089844, -67.73838806152344, 324.89459228515625, 224.19570922851562], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000435.npy"}
{"epoch": 0.6387665198237885, "step": 436, "batch_size": 64, "mean": 221.29774475097656, "std": 294.2098693847656, "min": -614.6396484375, "p10": -145.04007110595703, "median": 198.9014434814453, "p90": 554.9984924316407, "max": 993.009765625, "pos_frac": 0.796875, "sample": [-80.96473693847656, 132.1140899658203, -146.75225830078125, 67.7030258178711, 551.9945678710938, 221.140625, 393.21875, 754.286376953125, 583.8489990234375, -217.60610961914062, 123.93650817871094, 788.9290771484375, -53.31072998046875, -158.40525817871094, 270.26422119140625, 195.2249755859375, -164.50218200683594, 44.424285888671875, 860.5827026367188, -333.77752685546875, 442.87823486328125, 445.3470153808594, 369.57501220703125, 359.0174255371094, 484.7290954589844, 90.1763916015625, 427.4622802734375, 488.0049133300781, 91.19126892089844, 146.32228088378906, 993.009765625, 60.94806671142578, 556.285888671875, 234.57472229003906, 276.3913879394531, 162.3397216796875, -25.561782836914062, 89.27457427978516, 244.20535278320312, 130.51287841796875, 422.6556396484375, 71.867919921875, 279.79296875, 202.57791137695312, 208.75169372558594, 341.1258850097656, 248.10105895996094, -174.87838745117188, 192.80587768554688, 183.76617431640625, 246.05764770507812, 172.92779541015625, 407.8893737792969, 547.8518676757812, 839.3258056640625, -4.8792266845703125, 65.48087310791016, 20.597305297851562, -614.6396484375, 253.01756286621094, -141.0449676513672, 170.764404296875, 452.72369384765625, -128.61764526367188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000436.npy"}
{"epoch": 0.6402349486049926, "step": 437, "batch_size": 64, "mean": 220.18910217285156, "std": 298.1448059082031, "min": -618.61669921875, "p10": -92.99727630615234, "median": 212.41947174072266, "p90": 568.6462646484376, "max": 988.9445190429688, "pos_frac": 0.75, "sample": [115.68495178222656, -268.2190246582031, 273.05560302734375, 588.3118286132812, 409.8710021972656, -12.780405044555664, 578.587646484375, 79.78250122070312, 327.616455078125, -34.557098388671875, -618.61669921875, 177.00296020507812, -333.3518981933594, 473.6750793457031, -40.181427001953125, 108.61692810058594, 211.51614379882812, -297.906982421875, 101.64389038085938, 528.731689453125, 46.249046325683594, 544.19580078125, 213.3227996826172, -61.59722137451172, -180.69815063476562, 671.5218505859375, 146.98080444335938, 137.3541259765625, 389.1320495605469, 468.7191162109375, 788.4085693359375, 379.8939208984375, 778.5614013671875, -83.79985046386719, 988.9445190429688, -117.43441772460938, -63.734703063964844, 9.98298454284668, 261.1307067871094, 86.88101196289062, 59.549522399902344, 44.509002685546875, 327.17010498046875, 270.3091125488281, 451.61431884765625, 87.73492431640625, 427.7915954589844, 458.20855712890625, 456.80859375, 545.44970703125, 144.4489288330078, 439.5642395019531, 426.5479736328125, -95.9501953125, -56.576499938964844, 685.810546875, 312.2295227050781, 388.46978759765625, -86.10713195800781, 284.1407470703125, 49.064273834228516, 469.7265930175781, 231.15744018554688, -2.0674667358398438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000437.npy"}
{"epoch": 0.6417033773861968, "step": 438, "batch_size": 64, "mean": 250.0226287841797, "std": 294.630859375, "min": -465.121337890625, "p10": -71.9025505065918, "median": 246.57339477539062, "p90": 647.1927612304688, "max": 950.8098754882812, "pos_frac": 0.796875, "sample": [56.092926025390625, 0.23322486877441406, 340.37103271484375, -167.2519073486328, 160.79837036132812, 132.28041076660156, 849.218994140625, -24.303407669067383, 209.59231567382812, -70.93829345703125, 316.1139831542969, 401.6560363769531, 27.056453704833984, 111.47759246826172, 378.11370849609375, 513.5552368164062, 298.36846923828125, 950.8098754882812, 111.62080383300781, 246.611328125, 196.69522094726562, 577.4909057617188, 376.2947082519531, 601.4804077148438, 162.96630859375, 277.4012145996094, 60.115699768066406, 377.9793701171875, -417.9169006347656, -47.68768310546875, -82.27288818359375, -14.354047775268555, 340.6857604980469, 689.385986328125, 623.399169921875, 346.1139831542969, 253.42730712890625, -465.121337890625, 505.61956787109375, 328.65380859375, 710.025390625, 165.27243041992188, -26.991130828857422, 797.6942749023438, -57.62725830078125, 246.53546142578125, 458.1368408203125, 369.5512390136719, 510.05694580078125, 117.15386199951172, 143.55624389648438, 264.2060852050781, 782.374267578125, 92.79002380371094, -97.31173706054688, 129.891845703125, 112.59915161132812, -332.6568603515625, -72.31580352783203, 552.2296142578125, 170.9698028564453, 657.3900146484375, 324.90789794921875, 451.1756286621094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000438.npy"}
{"epoch": 0.6431718061674009, "step": 439, "batch_size": 64, "mean": 209.06298828125, "std": 332.6714172363281, "min": -481.136474609375, "p10": -117.88628158569335, "median": 159.81808471679688, "p90": 684.8416931152345, "max": 1168.750732421875, "pos_frac": 0.703125, "sample": [245.34213256835938, 725.6661376953125, 60.51814270019531, 270.57513427734375, -2.54791259765625, -7.226898193359375, 704.9532470703125, -65.57490539550781, 161.78753662109375, 53.47972869873047, -261.7188720703125, 637.9147338867188, 88.08665466308594, -112.99810028076172, -69.4371566772461, 233.90097045898438, 346.35345458984375, 633.14404296875, 636.6763916015625, 250.34005737304688, 152.59620666503906, 927.1761474609375, 237.09182739257812, -55.702911376953125, 226.5416259765625, -214.93804931640625, 157.8486328125, -19.869770050048828, -3.3876571655273438, -5.798318862915039, 369.7663879394531, -304.7308654785156, 296.6398010253906, 439.2037048339844, 1168.750732421875, 293.5964660644531, 583.87548828125, 193.49649047851562, 60.83418273925781, 243.8028564453125, 134.4580078125, -90.5874252319336, -481.136474609375, -30.597137451171875, 9.851234436035156, 259.04461669921875, 147.4804229736328, 238.9714813232422, -388.15997314453125, 579.4683227539062, 754.1495971679688, 5.792858123779297, 225.77500915527344, 706.8971557617188, 87.13738250732422, -119.98121643066406, 444.18560791015625, 399.9183654785156, 56.91744613647461, 89.94322204589844, 1037.8060302734375, 330.4913330078125, -291.5830078125, -2.239025115966797], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000439.npy"}
{"epoch": 0.644640234948605, "step": 440, "batch_size": 64, "mean": 238.82818603515625, "std": 346.4654846191406, "min": -306.84271240234375, "p10": -93.46393814086915, "median": 145.41861724853516, "p90": 720.7781616210941, "max": 1490.10595703125, "pos_frac": 0.796875, "sample": [286.2559509277344, 310.34478759765625, 476.7035217285156, 423.8564453125, -219.727294921875, 49.51262664794922, 622.3307495117188, 115.58478546142578, 69.54969787597656, 815.0092163085938, 23.771718978881836, 111.46865844726562, 295.0548095703125, 119.46026611328125, 811.643798828125, -182.82974243164062, 103.24185180664062, -306.84271240234375, 21.684768676757812, 5.791297912597656, 175.84512329101562, 578.6569213867188, 69.58445739746094, 41.58917999267578, 340.52691650390625, 231.3776397705078, -111.00511932373047, 179.1829833984375, 441.88616943359375, 163.9364471435547, 179.01602172851562, 568.7578735351562, -75.31777954101562, 265.89971923828125, 17.132112503051758, -93.02761840820312, 1189.5408935546875, -268.4411315917969, 210.72467041015625, 28.790014266967773, 790.7611083984375, 238.09115600585938, 1009.5806884765625, 482.4281005859375, 221.00067138671875, 100.3475112915039, -170.22201538085938, 82.11564636230469, 481.4847106933594, 593.3396606445312, 91.92071533203125, 762.9699096679688, -45.30109405517578, 136.1853485107422, 20.525585174560547, 493.7452392578125, -1.6745719909667969, -40.807472229003906, -64.44547271728516, 154.65188598632812, 1490.10595703125, 455.0198059082031, -93.65093231201172, 10.3104248046875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000440.npy"}
{"epoch": 0.6461086637298091, "step": 441, "batch_size": 64, "mean": 231.1967010498047, "std": 397.0838623046875, "min": -792.4142456054688, "p10": -202.74330291748046, "median": 195.8362579345703, "p90": 732.5480773925781, "max": 1603.480224609375, "pos_frac": 0.734375, "sample": [259.291748046875, 126.61109924316406, 214.16839599609375, 533.8648071289062, 424.2799072265625, 803.0263671875, -60.7149658203125, 28.675338745117188, 177.50411987304688, -442.8071594238281, -187.5422821044922, -56.95690155029297, 733.012451171875, 885.0213012695312, 281.71734619140625, -305.35247802734375, 86.77662658691406, 791.6638793945312, 51.08837890625, 8.394317626953125, 41.215728759765625, 131.3987274169922, 12.476104736328125, 700.3309326171875, 757.2633666992188, 93.85708618164062, 227.33558654785156, 421.71282958984375, 372.8618469238281, 1603.480224609375, 48.00006866455078, -511.3036804199219, 262.60662841796875, 498.5243225097656, -363.81494140625, -88.6209716796875, 113.73249053955078, 268.2913818359375, 572.005615234375, -209.9283447265625, -92.3526611328125, 512.2606201171875, -91.29393005371094, -792.4142456054688, 465.22064208984375, 397.727783203125, 349.24420166015625, 378.90484619140625, 246.81204223632812, -57.840572357177734, 152.47950744628906, -209.25802612304688, -118.2442626953125, 483.0980224609375, 331.3980712890625, 87.3837890625, 731.4645385742188, 339.87054443359375, 485.58636474609375, 1091.840087890625, -28.55234146118164, -30.575340270996094, 139.3994140625, 721.283203125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000441.npy"}
{"epoch": 0.6475770925110133, "step": 442, "batch_size": 64, "mean": 264.99951171875, "std": 354.3354187011719, "min": -462.764892578125, "p10": -82.8918632507324, "median": 164.09744262695312, "p90": 719.9660156250002, "max": 1199.894287109375, "pos_frac": 0.796875, "sample": [-131.02053833007812, 180.74630737304688, 505.75726318359375, 49.57525634765625, 301.37750244140625, 151.72836303710938, -63.39128875732422, 203.21026611328125, 64.44508361816406, 154.85223388671875, 12.847545623779297, 368.2351989746094, -4.216072082519531, 411.7071533203125, 371.3395080566406, 620.1615600585938, 892.5679931640625, 392.1717224121094, -7.928253173828125, 149.1605682373047, 86.72809600830078, 524.8283081054688, 82.88424682617188, 239.59576416015625, -122.84178924560547, 155.22406005859375, 687.7449951171875, 297.7933349609375, 874.371826171875, 172.9708251953125, 93.93630981445312, -462.764892578125, 105.30075073242188, -91.24925231933594, -7.252998352050781, 483.36297607421875, -49.18387222290039, 478.04559326171875, 68.38838958740234, 1199.894287109375, -325.3052978515625, 358.16204833984375, 733.7750244140625, 28.819997787475586, 31.1409912109375, 1164.912841796875, 633.9107666015625, 289.27294921875, 490.3728942871094, 176.52310180664062, -9.4349365234375, -257.7464904785156, 74.77005004882812, 140.9330596923828, 5.2031097412109375, -252.72923278808594, 352.5006103515625, 376.4874267578125, 988.5922241210938, 569.136962890625, 1131.2703857421875, 660.3099975585938, 80.56953430175781, 77.41487121582031], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000442.npy"}
{"epoch": 0.6490455212922174, "step": 443, "batch_size": 64, "mean": 174.056640625, "std": 385.990966796875, "min": -667.430419921875, "p10": -224.7462646484375, "median": 152.12987518310547, "p90": 614.7157470703129, "max": 1235.621826171875, "pos_frac": 0.765625, "sample": [1212.2041015625, 62.208106994628906, 799.8092041015625, 19.472396850585938, 323.11004638671875, 149.40020751953125, -76.33979034423828, -68.88252258300781, -330.0074462890625, 114.48029327392578, 65.06028747558594, 46.48399353027344, 192.22216796875, 198.05072021484375, 22.701019287109375, -667.430419921875, 126.48744201660156, -221.84413146972656, -83.28878021240234, 12.640073776245117, 1226.9517822265625, 392.40374755859375, 508.31549072265625, 294.3724060058594, -633.3047485351562, -642.5902709960938, 272.42138671875, 154.8595428466797, 660.3158569335938, -667.3319091796875, 361.0950012207031, 393.720947265625, 220.8799285888672, 108.21238708496094, 36.54816818237305, -109.63257598876953, 181.5379638671875, 438.967529296875, 448.42449951171875, 143.4722137451172, -240.6736297607422, -27.474899291992188, 455.9685363769531, -175.03292846679688, 331.568603515625, 77.17036437988281, 302.654296875, 1235.621826171875, 193.02577209472656, 126.96598052978516, 462.9971618652344, 105.59417724609375, 203.20896911621094, 47.57313537597656, -175.44432067871094, 278.1896667480469, -225.9900360107422, 229.76515197753906, 175.2430419921875, 86.06265258789062, 709.380859375, 829.9080810546875, 216.2952880859375, 230.87060546875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000443.npy"}
{"epoch": 0.6505139500734214, "step": 444, "batch_size": 64, "mean": 293.02239990234375, "std": 355.6820373535156, "min": -493.4337463378906, "p10": -123.8894119262695, "median": 261.16259765625, "p90": 739.0816528320316, "max": 1386.5120849609375, "pos_frac": 0.796875, "sample": [308.41180419921875, -171.38262939453125, 31.335182189941406, 597.62451171875, 52.02989959716797, 337.356201171875, -82.68293762207031, 1262.2861328125, 822.7650146484375, -33.40730667114258, -98.20278930664062, 233.99319458007812, 248.72735595703125, 534.9611206054688, 583.2832641601562, 121.23077392578125, 821.806640625, 453.6131591796875, -201.13339233398438, 97.22700500488281, 34.45972442626953, -9.385370254516602, 419.6446228027344, 777.6966552734375, 68.37480163574219, 389.16522216796875, 316.4505310058594, -36.092742919921875, 297.2331848144531, -309.161865234375, -36.19286346435547, 1386.5120849609375, 248.89418029785156, 337.5810546875, 321.5615234375, 537.2194213867188, 1026.377197265625, 414.58245849609375, 519.046630859375, 243.65716552734375, -144.2999267578125, 922.7343139648438, 645.4354858398438, 112.6498031616211, 507.69293212890625, 486.1766052246094, 500.999267578125, 60.736976623535156, 261.61309814453125, 150.7090606689453, 299.22613525390625, 209.38526916503906, 260.71209716796875, 206.953125, 493.1457214355469, 274.79296875, 37.53602600097656, -134.89796447753906, -206.11087036132812, 39.173797607421875, 648.97998046875, -493.4337463378906, 239.6141357421875, 506.443115234375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000444.npy"}
{"epoch": 0.6519823788546255, "step": 445, "batch_size": 64, "mean": 264.8874816894531, "std": 329.8348083496094, "min": -426.85186767578125, "p10": -133.3806076049804, "median": 232.18791961669922, "p90": 666.1269226074219, "max": 1143.599609375, "pos_frac": 0.765625, "sample": [227.69284057617188, 61.37251281738281, 2.7645492553710938, 878.365478515625, -15.720626831054688, 534.9248046875, 424.9085388183594, 184.646240234375, 57.838226318359375, 178.8843231201172, 45.258785247802734, 599.1622314453125, -16.064376831054688, 407.08148193359375, 114.4019546508789, -426.85186767578125, 48.69244384765625, 456.8462219238281, 325.7558898925781, 185.23631286621094, 456.5784912109375, 343.8550720214844, 479.4690246582031, 705.9086303710938, -31.450847625732422, 146.21536254882812, 1105.7681884765625, 450.4570007324219, 443.91888427734375, 655.0802612304688, 722.9595336914062, 634.7359619140625, 140.7626190185547, 670.8612060546875, 92.86820983886719, -196.7940673828125, 189.03817749023438, 289.1333312988281, -183.96165466308594, 525.9669799804688, 8.696800231933594, 983.3221435546875, 535.7430419921875, -177.8850860595703, 536.921142578125, 378.84661865234375, 236.68299865722656, -205.4593505859375, -55.781333923339844, 4.337324142456055, 255.08474731445312, -15.324470520019531, -59.42006301879883, 443.8871154785156, 249.400390625, 217.15443420410156, -32.591949462890625, -158.73007202148438, 1143.599609375, -74.23185729980469, 534.2948608398438, 261.2110595703125, 264.70806884765625, -238.23385620117188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000445.npy"}
{"epoch": 0.6534508076358296, "step": 446, "batch_size": 64, "mean": 288.05413818359375, "std": 301.55267333984375, "min": -446.3147888183594, "p10": -42.93692779541014, "median": 269.0550231933594, "p90": 708.1209228515626, "max": 1057.8251953125, "pos_frac": 0.84375, "sample": [50.93195343017578, 314.41180419921875, 60.70143127441406, 716.500244140625, 859.1613159179688, 377.9327392578125, 51.76298904418945, -446.3147888183594, 187.9835968017578, 605.4970092773438, 879.0662231445312, 410.22235107421875, -32.396873474121094, 676.2456665039062, 370.4405212402344, 599.7733764648438, 592.881591796875, 357.6169738769531, 166.68191528320312, 378.3641052246094, -80.36431884765625, 129.79165649414062, 137.0504608154297, 166.64048767089844, 275.1922607421875, -349.55914306640625, 127.93162536621094, 114.08596801757812, 463.9624938964844, 222.58013916015625, 485.3908386230469, -116.09498596191406, 84.10177612304688, 195.01719665527344, 22.119543075561523, 150.75332641601562, -15.255416870117188, 462.322021484375, 314.2943420410156, 690.0925903320312, 369.67095947265625, 208.45216369628906, 675.5057373046875, 1057.8251953125, 366.92877197265625, -47.45409393310547, 843.1724853515625, 475.9718933105469, 262.91778564453125, 332.9030456542969, 302.972412109375, -1.2115631103515625, -81.577880859375, 164.76467895507812, 87.44356536865234, 219.92669677734375, 131.09228515625, 399.62677001953125, 795.611083984375, 212.62255859375, -292.8509826660156, 715.8473510742188, 300.0551452636719, 277.7317810058594], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000446.npy"}
{"epoch": 0.6549192364170338, "step": 447, "batch_size": 64, "mean": 297.126708984375, "std": 384.181884765625, "min": -527.57470703125, "p10": -124.78732604980469, "median": 250.30726623535156, "p90": 761.7721801757813, "max": 1974.347900390625, "pos_frac": 0.828125, "sample": [249.83953857421875, 50.848731994628906, 317.4757080078125, 183.0229949951172, -36.422149658203125, -527.57470703125, 526.6011962890625, 223.87429809570312, 337.66864013671875, -125.75936889648438, 49.13905715942383, 799.6812744140625, 262.656982421875, 208.9688262939453, 743.871337890625, 601.2085571289062, 773.2572021484375, 48.06134033203125, 1974.347900390625, 769.4439697265625, 142.00436401367188, 1108.90087890625, 425.9571838378906, 417.998291015625, -224.73011779785156, 227.423583984375, 453.8798828125, -143.11691284179688, 250.77499389648438, 149.49375915527344, -188.73782348632812, 40.79594421386719, 49.467933654785156, -122.51922607421875, 386.8585205078125, -13.578813552856445, 265.5052185058594, 565.5435791015625, 101.11260986328125, 32.093345642089844, 567.2691040039062, 8.899528503417969, 111.06306457519531, -375.944580078125, 774.79833984375, 499.1861572265625, 450.6176452636719, -149.57662963867188, 232.8900146484375, 28.5804386138916, 450.07440185546875, 323.35443115234375, 343.2969970703125, 269.3199462890625, 180.64663696289062, 594.6234130859375, 349.8995056152344, 214.8848419189453, 1027.27392578125, 584.6171875, 138.800048828125, 670.0074462890625, 486.5370178222656, -120.34738159179688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000447.npy"}
{"epoch": 0.6563876651982379, "step": 448, "batch_size": 64, "mean": 289.0185546875, "std": 359.1437072753906, "min": -750.3828735351562, "p10": -114.15853652954098, "median": 243.0093536376953, "p90": 678.7206298828125, "max": 1383.0032958984375, "pos_frac": 0.828125, "sample": [10.067039489746094, 67.61343383789062, -339.4863586425781, 167.06314086914062, 546.760986328125, 7.8565673828125, 109.74708557128906, -75.27454376220703, 288.3916015625, -750.3828735351562, 457.78472900390625, 157.61318969726562, 658.6826171875, 114.73129272460938, 252.83261108398438, 554.0422973632812, 55.43230438232422, 39.87806701660156, 608.8872680664062, -138.95529174804688, -130.82310485839844, 130.71331787109375, 884.167724609375, 437.6683349609375, 238.86239624023438, 245.54769897460938, 264.62286376953125, 212.08770751953125, 475.0349426269531, 866.857177734375, 339.926513671875, 592.45703125, 68.0163803100586, 375.3340759277344, 1383.0032958984375, -20.80088996887207, 116.74470520019531, 205.71290588378906, -5.673545837402344, 315.00360107421875, 687.308349609375, 531.1372680664062, 233.8253631591797, -185.7197265625, 1099.011474609375, 297.0899658203125, 219.189208984375, -58.76251220703125, 423.3454895019531, 137.86746215820312, 336.2077331542969, -159.24729919433594, 977.2293701171875, -201.37103271484375, 619.6317749023438, 928.1318969726562, 8.957916259765625, 378.0754699707031, 96.87158966064453, 315.7864074707031, 600.7174072265625, 645.1519775390625, 538.5333862304688, 240.47100830078125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000448.npy"}
{"epoch": 0.657856093979442, "step": 449, "batch_size": 64, "mean": 283.281494140625, "std": 435.1645202636719, "min": -842.6619873046875, "p10": -217.39022216796874, "median": 233.8327178955078, "p90": 720.7878234863282, "max": 2091.3984375, "pos_frac": 0.78125, "sample": [30.497177124023438, 45.465431213378906, 337.59320068359375, 20.797592163085938, 716.00244140625, 665.2269287109375, 264.67138671875, 7.403127670288086, 158.5396728515625, 261.9611511230469, 882.5713500976562, -203.8168487548828, 13.506248474121094, 202.83847045898438, 634.8145751953125, 25.230205535888672, 444.87078857421875, 180.20535278320312, 117.08157348632812, -311.013671875, 864.4105224609375, 330.3033142089844, 592.6787719726562, 692.0439453125, 362.832275390625, 716.7118530273438, -358.3143615722656, 32.10829162597656, 357.4208984375, 439.9538879394531, 192.98667907714844, 211.12850952148438, -26.50189208984375, 647.0994873046875, 599.3372802734375, 688.5379028320312, -96.57869720458984, 177.70887756347656, -244.945068359375, -223.20738220214844, 2091.3984375, 722.53466796875, -249.3983917236328, 295.89642333984375, 354.35614013671875, 132.65609741210938, -52.73382568359375, -842.6619873046875, -85.1988525390625, 1026.992919921875, 256.53692626953125, 858.294189453125, 711.86083984375, -240.48655700683594, 424.49566650390625, 178.08624267578125, 359.99700927734375, 1054.152587890625, 279.685791015625, -61.34574890136719, -194.18605041503906, 38.44818115234375, 155.79383850097656, 464.67987060546875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000449.npy"}
{"epoch": 0.6593245227606461, "step": 450, "batch_size": 64, "mean": 327.5614013671875, "std": 321.4566650390625, "min": -729.8837280273438, "p10": 1.924383926391621, "median": 305.47361755371094, "p90": 746.6911743164063, "max": 1190.27490234375, "pos_frac": 0.890625, "sample": [473.4295349121094, 742.2702026367188, 823.3526611328125, 325.6950378417969, 107.496337890625, 748.5858764648438, 1190.27490234375, -78.51402282714844, 373.0162658691406, -6.071464538574219, 680.8740844726562, 495.04345703125, 711.199462890625, 469.9356994628906, 386.26593017578125, 800.7183227539062, 99.37579345703125, 53.601318359375, 181.0474853515625, 108.6114501953125, 347.97320556640625, 641.432373046875, 441.205810546875, 739.6126708984375, 175.9320526123047, -12.50857162475586, 113.46446228027344, 854.7459716796875, 385.19024658203125, 652.9445190429688, 43.10054016113281, 386.0601806640625, 317.711181640625, 227.8456573486328, 295.04949951171875, 20.581363677978516, 315.8977355957031, 408.4594421386719, 218.363525390625, 126.03076171875, 861.5505981445312, -170.6688690185547, 165.10919189453125, 392.892822265625, 116.27861022949219, 352.04254150390625, 58.04700469970703, 621.718017578125, 257.03741455078125, 60.29349899291992, 488.15362548828125, -66.28274536132812, 619.8607177734375, 272.8431396484375, 34.87982940673828, -729.8837280273438, 278.8135986328125, 316.7005920410156, 51.41315460205078, 254.6033477783203, 117.70008087158203, 1036.887939453125, -9.805778503417969, 198.44386291503906], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000450.npy"}
{"epoch": 0.6607929515418502, "step": 451, "batch_size": 64, "mean": 236.28216552734375, "std": 448.90203857421875, "min": -1299.7734375, "p10": -206.06430816650388, "median": 264.4001007080078, "p90": 748.37919921875, "max": 1201.405517578125, "pos_frac": 0.796875, "sample": [43.127044677734375, -99.07330322265625, 49.77264404296875, 369.7818298339844, 554.8637084960938, 230.77273559570312, 612.603515625, 659.4845581054688, -52.630027770996094, 330.1975402832031, 414.6138000488281, 1201.405517578125, 408.8402099609375, 872.393310546875, 134.5692138671875, 190.91763305664062, 126.9170150756836, -1299.7734375, -937.1862182617188, -152.49920654296875, 92.58910369873047, 87.0958023071289, 428.590087890625, 611.5927734375, 850.549072265625, 271.0055236816406, 423.1824951171875, 257.794677734375, 348.2194519042969, 564.8065185546875, 136.78558349609375, 1076.0450439453125, 1005.096435546875, 367.242431640625, 518.1869506835938, 737.5047607421875, 435.4083557128906, -374.75634765625, 296.4294128417969, 402.90771484375, -39.64360809326172, -158.78610229492188, 445.7819519042969, -1062.4779052734375, 100.38719177246094, 205.33010864257812, 217.3855743408203, 121.2692642211914, 271.97607421875, 139.72567749023438, 797.0324096679688, -268.9290771484375, -167.30459594726562, 566.3316650390625, 515.1958618164062, 140.06265258789062, 372.7633056640625, -489.4859924316406, 178.058837890625, 10.603269577026367, 753.0396728515625, -222.6756134033203, 294.53167724609375, 206.51177978515625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000451.npy"}
{"epoch": 0.6622613803230544, "step": 452, "batch_size": 64, "mean": 287.4557189941406, "std": 288.8720703125, "min": -244.79238891601562, "p10": -108.5353530883789, "median": 271.3389129638672, "p90": 661.4002014160156, "max": 949.5687866210938, "pos_frac": 0.8125, "sample": [757.4608154296875, -244.79238891601562, 572.2893676757812, 655.2666015625, -169.17788696289062, 520.9814453125, 413.38446044921875, 629.1159057617188, 627.3832397460938, -63.47838592529297, 448.5047607421875, 315.17864990234375, 144.49868774414062, 174.30628967285156, 392.06585693359375, 299.6856994628906, 662.0067749023438, 275.8518981933594, 464.66986083984375, 14.771589279174805, 391.103759765625, -9.229301452636719, 449.5103759765625, 91.71546173095703, 382.5665283203125, 700.3385620117188, 176.4433135986328, -243.21929931640625, 580.0498046875, 235.76992797851562, -77.26608276367188, 407.26129150390625, 122.10066223144531, 164.83889770507812, 749.3809814453125, 143.0037841796875, 240.5200653076172, 659.98486328125, 66.17345428466797, 202.63650512695312, 245.70596313476562, -189.455322265625, 192.30001831054688, 484.5981750488281, 266.825927734375, 457.45404052734375, 131.67095947265625, 780.545654296875, 949.5687866210938, 82.89146423339844, 155.45616149902344, 319.33978271484375, 375.11041259765625, -8.53433609008789, -234.07589721679688, 331.14947509765625, -113.65753173828125, 515.783447265625, 504.91259765625, 774.2916259765625, -121.10735321044922, 239.73001098632812, -96.58360290527344, 29.58880043029785], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000452.npy"}
{"epoch": 0.6637298091042585, "step": 453, "batch_size": 64, "mean": 283.8158264160156, "std": 371.39581298828125, "min": -324.1893310546875, "p10": -120.45666961669922, "median": 186.5876007080078, "p90": 727.8721313476564, "max": 1714.81494140625, "pos_frac": 0.75, "sample": [695.5411376953125, 590.9326782226562, -79.07433319091797, 54.53978729248047, 136.21484375, 115.59980773925781, 365.1727600097656, 103.24317932128906, 235.15850830078125, 78.172607421875, 260.7127990722656, 618.3216552734375, 542.7840576171875, 622.4091796875, -14.260250091552734, 109.09315490722656, 326.42266845703125, 129.76971435546875, -142.7307891845703, 150.3506622314453, 541.9634399414062, -200.69715881347656, 251.82443237304688, 464.1405029296875, 659.106201171875, -257.5796813964844, -55.67375183105469, 756.7700805664062, -324.1893310546875, 409.7679138183594, -120.85047912597656, 209.79312133789062, 507.8157958984375, 952.348388671875, 870.1259765625, 14.033313751220703, -5.9975433349609375, 141.83383178710938, 1714.81494140625, 741.728271484375, 527.759765625, -111.53426361083984, 928.0668334960938, 25.61675262451172, 269.1014404296875, 103.48493957519531, 578.00048828125, 289.71905517578125, 477.25311279296875, 483.02459716796875, -204.63002014160156, 514.4192504882812, -119.53778076171875, 560.2879638671875, 64.71826171875, -95.18341827392578, -184.62828063964844, 125.90303802490234, 1018.9381103515625, 490.1586608886719, -22.103778839111328, -26.26856231689453, 138.81097412109375, 163.382080078125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000453.npy"}
{"epoch": 0.6651982378854625, "step": 454, "batch_size": 64, "mean": 306.5694885253906, "std": 419.3467102050781, "min": -812.966064453125, "p10": -98.49201812744138, "median": 246.46866607666016, "p90": 757.994677734375, "max": 1778.1998291015625, "pos_frac": 0.8125, "sample": [247.05276489257812, 122.2174072265625, 993.4624633789062, 221.37135314941406, 521.6815795898438, 41.94172668457031, 791.3194580078125, 410.49847412109375, -110.90982055664062, 208.1171112060547, -723.3555908203125, 568.2002563476562, 379.0928039550781, 367.76788330078125, 240.6337890625, 252.61212158203125, 350.21136474609375, 401.29486083984375, 65.44617462158203, -812.966064453125, 18.549028396606445, 716.2095947265625, -78.70083618164062, 780.1383666992188, -452.6559143066406, -105.062744140625, 686.0252075195312, 80.73758697509766, -16.116134643554688, 231.64364624023438, 604.0497436523438, 105.74632263183594, 305.8082275390625, 103.1450424194336, 448.2269287109375, 142.76370239257812, 231.41641235351562, 227.89462280273438, 391.75146484375, 245.8845672607422, 510.6483154296875, 468.5468444824219, 73.47383117675781, 436.244873046875, -3.7126922607421875, 768.2120361328125, 1406.621826171875, 452.742431640625, -211.8153839111328, 171.0325927734375, 1778.1998291015625, 1096.1041259765625, 306.9880065917969, 562.3529052734375, 72.00875091552734, 450.6631164550781, 693.4274291992188, -136.7601318359375, 588.07275390625, -83.16032409667969, 187.3654022216797, 96.15339660644531, 734.1541748046875, -0.2616691589355469], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000454.npy"}
{"epoch": 0.6666666666666666, "step": 455, "batch_size": 64, "mean": 361.14434814453125, "std": 439.1692199707031, "min": -842.593017578125, "p10": -101.72677459716795, "median": 297.4573669433594, "p90": 1110.8136718750002, "max": 1384.37353515625, "pos_frac": 0.859375, "sample": [76.27543640136719, 110.617919921875, 828.3587036132812, 1066.5263671875, -486.66278076171875, 350.6343688964844, 678.1951904296875, -205.20118713378906, 1203.2210693359375, -213.52964782714844, 1129.7939453125, 301.61883544921875, 1136.09765625, -313.9021301269531, 1161.1328125, -111.41960906982422, 237.36080932617188, 524.2503662109375, 297.7477111816406, 234.8330841064453, 148.10360717773438, 541.636962890625, 250.77476501464844, -13.350210189819336, 427.29278564453125, 1199.462158203125, 812.4097290039062, -79.11016082763672, 309.5989990234375, 491.43682861328125, 200.1888427734375, 403.85595703125, 260.9002685546875, 297.1670227050781, 249.15713500976562, 219.06329345703125, 285.44805908203125, 312.9372863769531, 59.5260009765625, -842.593017578125, 404.9211120605469, 62.13558578491211, 263.357177734375, 617.3523559570312, 45.18754577636719, 28.05268096923828, 82.68841552734375, 491.13165283203125, 859.1347045898438, 261.9697570800781, 691.8159790039062, 96.290771484375, 472.7386474609375, 468.6210021972656, 97.00452423095703, 110.14285278320312, 46.09108352661133, 847.9359130859375, 464.99432373046875, 358.07647705078125, 1299.360107421875, 399.0149230957031, -279.00738525390625, 1384.37353515625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000455.npy"}
{"epoch": 0.6681350954478708, "step": 456, "batch_size": 64, "mean": 346.8193359375, "std": 368.3343505859375, "min": -194.33169555664062, "p10": -18.67543907165527, "median": 255.58399963378906, "p90": 916.7721801757814, "max": 1575.185546875, "pos_frac": 0.859375, "sample": [164.65472412109375, 196.6240692138672, 741.2428588867188, 32.781646728515625, -35.53839111328125, 328.9984130859375, 93.98861694335938, -14.477508544921875, 239.884521484375, 767.8186645507812, 379.02716064453125, 17.17125129699707, 52.61848449707031, 1575.185546875, 293.9453125, -31.58972930908203, -66.21324157714844, 271.2834777832031, 328.4107971191406, 421.7021789550781, 334.9757080078125, 65.50105285644531, -20.51055908203125, 1107.6002197265625, -20.474552154541016, 443.32440185546875, 1343.244140625, 430.0496520996094, 590.3082275390625, 27.94000244140625, 315.96026611328125, 515.0499267578125, 730.6571044921875, 159.8291473388672, 194.05966186523438, 198.9834442138672, 97.40110778808594, 462.06683349609375, 1027.3634033203125, 102.364013671875, 1011.4950561523438, 395.130126953125, 301.61474609375, 59.671531677246094, -194.33169555664062, 534.6990966796875, 73.25150299072266, -49.521148681640625, 888.8302001953125, 47.97783660888672, 928.747314453125, 1041.0489501953125, 374.7516174316406, -9.853958129882812, 205.54721069335938, 79.61125183105469, 345.3157653808594, 85.45999145507812, 391.5238037109375, 233.88226318359375, 162.96664428710938, 380.4387512207031, 810.04638671875, 234.92050170898438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000456.npy"}
{"epoch": 0.6696035242290749, "step": 457, "batch_size": 64, "mean": 340.1048583984375, "std": 358.640869140625, "min": -412.6234130859375, "p10": -123.77299652099606, "median": 392.29290771484375, "p90": 792.7764648437502, "max": 1325.341064453125, "pos_frac": 0.828125, "sample": [-34.865318298339844, 455.1698913574219, 440.10687255859375, 402.5920104980469, 667.1610107421875, 678.2783813476562, 850.29345703125, 884.3511352539062, 70.6845932006836, 52.47552490234375, 1125.30419921875, 410.64886474609375, 28.05789566040039, 251.93373107910156, 72.67579650878906, 238.56666564941406, 561.2808837890625, 141.81954956054688, -148.01724243164062, 578.1470947265625, 720.9071044921875, -12.909954071044922, 160.7995147705078, 535.9003295898438, 488.62847900390625, 1035.849365234375, -412.6234130859375, -217.82130432128906, 550.8157958984375, 569.56396484375, 82.66654968261719, 492.5223083496094, 468.1974792480469, 528.4254150390625, 306.94207763671875, -92.26185607910156, 741.766357421875, 468.1183166503906, 66.94815063476562, 35.98297119140625, 450.567626953125, 824.9122924804688, 512.9849243164062, 141.0579376220703, 67.88343811035156, 126.18299865722656, 578.0702514648438, -137.27777099609375, 481.99200439453125, 389.9212341308594, -194.55615234375, 700.7192993164062, 259.2464904785156, 814.637939453125, -348.2337646484375, 391.7628173828125, 92.77711486816406, -149.13595581054688, 392.822998046875, -9.979320526123047, 229.587890625, 575.0632934570312, 5.279838562011719, 1325.341064453125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000457.npy"}
{"epoch": 0.671071953010279, "step": 458, "batch_size": 64, "mean": 297.208984375, "std": 372.3271484375, "min": -546.070068359375, "p10": -168.9265182495117, "median": 307.7273254394531, "p90": 652.1668395996094, "max": 1612.2095947265625, "pos_frac": 0.828125, "sample": [-546.070068359375, 651.4270629882812, 477.98504638671875, 264.0276794433594, 678.5006103515625, -180.7056121826172, 1593.47509765625, -140.82969665527344, 306.5947265625, 587.1573486328125, 104.3745346069336, 246.75274658203125, 479.81597900390625, 336.5071105957031, 717.4502563476562, -81.03564453125, -151.40260314941406, 15.492574691772461, 215.26150512695312, 649.7554321289062, 340.7549133300781, 497.2878723144531, 53.355674743652344, 552.8779907226562, 375.96209716796875, 363.3265380859375, 331.69659423828125, 74.78253173828125, 245.4208526611328, 486.7610168457031, 263.15740966796875, 625.3359985351562, 495.4018859863281, 900.2042236328125, 49.22674560546875, 13.86224365234375, 349.3200988769531, 76.23905944824219, 404.8987121582031, 314.9107360839844, 730.922119140625, 394.14056396484375, -124.5476303100586, -213.64987182617188, 157.19781494140625, -272.63427734375, 136.30999755859375, -176.436767578125, 155.39083862304688, 1612.2095947265625, 549.7073364257812, 214.04391479492188, 260.2234191894531, 88.31671142578125, 195.10647583007812, 652.48388671875, 308.85992431640625, 446.2832946777344, 255.4346466064453, 315.248779296875, -221.1107940673828, 454.2825012207031, -277.1689758300781, 341.44287109375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000458.npy"}
{"epoch": 0.6725403817914831, "step": 459, "batch_size": 64, "mean": 326.720947265625, "std": 404.5224914550781, "min": -646.2001342773438, "p10": -175.6685760498047, "median": 284.6866912841797, "p90": 861.9357849121095, "max": 1356.5352783203125, "pos_frac": 0.796875, "sample": [71.68223571777344, 118.79202270507812, 420.962158203125, 215.05181884765625, 189.70516967773438, 364.15234375, -266.8520202636719, 124.80912780761719, 591.9712524414062, -356.4721374511719, 595.8916015625, 461.9376220703125, 462.336669921875, 466.50714111328125, -10.71306037902832, 433.9310302734375, -74.84869384765625, 622.533447265625, 364.2340087890625, 905.5353393554688, 41.40911865234375, 841.4097290039062, 563.203369140625, -167.54115295410156, 309.6423034667969, -176.49266052246094, 946.6826171875, -173.74571228027344, 26.874340057373047, -0.35256385803222656, 226.36647033691406, -44.588584899902344, 870.732666015625, -646.2001342773438, 1356.5352783203125, 283.4451599121094, 1133.02197265625, 520.7676391601562, 248.6499481201172, -233.989501953125, 177.12130737304688, 207.99563598632812, 488.46209716796875, 202.62811279296875, 112.29989624023438, 285.92822265625, -430.3447570800781, 197.85366821289062, 438.6554870605469, 709.3429565429688, 724.1319580078125, 101.38966369628906, 401.670166015625, 1184.25390625, 263.4512939453125, 140.04908752441406, 978.6046752929688, -185.49913024902344, 824.4317016601562, 711.024658203125, 810.0149536132812, 317.642578125, 10.209720611572266, 611.845458984375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000459.npy"}
{"epoch": 0.6740088105726872, "step": 460, "batch_size": 64, "mean": 344.1812438964844, "std": 387.73492431640625, "min": -349.292724609375, "p10": -127.56275329589843, "median": 317.6717224121094, "p90": 856.3496520996094, "max": 1380.1815185546875, "pos_frac": 0.828125, "sample": [777.353759765625, 613.0508422851562, 367.8330993652344, -144.42361450195312, 111.20814514160156, 834.76416015625, 989.655029296875, 319.4673156738281, 22.402976989746094, 266.67193603515625, -98.95613861083984, 644.8419189453125, 121.52256774902344, 544.2412109375, 133.14476013183594, 192.1937255859375, 315.8761291503906, -349.292724609375, 152.1480255126953, 1252.39208984375, 1327.8331298828125, 190.8795166015625, 369.7757873535156, 60.606689453125, 967.092041015625, 315.7236022949219, 1380.1815185546875, 376.2943115234375, -111.3430404663086, 26.14398193359375, 267.5684509277344, -151.45114135742188, 511.935791015625, -117.02699279785156, 562.8726806640625, 590.9016723632812, -132.0780792236328, 402.843017578125, 277.1705017089844, -42.687095642089844, 862.760009765625, 590.607421875, 412.3568420410156, 406.4518127441406, 520.5999755859375, 127.84857177734375, 93.62492370605469, 135.6269989013672, 333.59515380859375, 236.3835906982422, -311.9651794433594, 105.00838470458984, 118.68700408935547, 431.7085876464844, -308.084716796875, 596.9041748046875, 980.2623291015625, 398.7094421386719, -205.88546752929688, 841.3921508789062, 636.036865234375, 395.4507141113281, 88.03221893310547, 402.15594482421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000460.npy"}
{"epoch": 0.6754772393538914, "step": 461, "batch_size": 64, "mean": 279.996337890625, "std": 369.794677734375, "min": -482.05645751953125, "p10": -94.72923507690427, "median": 225.90211486816406, "p90": 699.0469543457032, "max": 1373.791748046875, "pos_frac": 0.796875, "sample": [115.10821533203125, 23.124801635742188, 707.65966796875, 376.4239196777344, -67.64643096923828, 661.0206298828125, -482.05645751953125, 1195.1827392578125, 88.97598266601562, 551.28955078125, 194.48373413085938, -79.58155059814453, 194.6925048828125, 537.1337280273438, 988.7484130859375, 442.2388916015625, 650.3887329101562, -59.8316650390625, 253.60789489746094, 346.1378173828125, 278.22412109375, 678.9506225585938, 629.4979858398438, 947.795654296875, -197.46498107910156, 11.863250732421875, 1045.424072265625, 1373.791748046875, 260.865966796875, 396.539306640625, 452.1630554199219, 219.73985290527344, 201.17547607421875, 35.75518798828125, 91.202880859375, 313.118408203125, -299.73931884765625, 11.881317138671875, -101.22109985351562, 71.48709106445312, 765.5661010742188, -65.87958526611328, 587.9647216796875, 583.8822631835938, -248.25257873535156, 309.9917907714844, 232.0643768310547, 547.2578125, 18.550506591796875, -62.4385986328125, 538.1009521484375, 92.68224334716797, 115.08963012695312, 168.2828369140625, 530.4259643554688, -74.53455352783203, -161.57296752929688, 129.75588989257812, 264.87469482421875, 219.05270385742188, 109.84477996826172, -391.985107421875, 297.437255859375, 355.45416259765625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000461.npy"}
{"epoch": 0.6769456681350955, "step": 462, "batch_size": 64, "mean": 225.1715087890625, "std": 378.9349670410156, "min": -546.557373046875, "p10": -218.11126403808595, "median": 213.30335235595703, "p90": 586.3021911621095, "max": 1358.1025390625, "pos_frac": 0.671875, "sample": [361.533935546875, 32.419593811035156, -6.613395690917969, -187.22657775878906, 232.12991333007812, 542.174072265625, 40.07305908203125, 422.9681091308594, 553.4188842773438, 721.90869140625, 492.0970458984375, -217.24658203125, -326.97943115234375, 88.83059692382812, 162.43338012695312, 456.63037109375, 543.3824462890625, 265.2832336425781, -226.92315673828125, 628.4231567382812, -201.55398559570312, 99.24057006835938, 357.7628479003906, -546.557373046875, -117.11918640136719, -8.659393310546875, 568.7935180664062, -350.48028564453125, 32.2299690246582, -31.130504608154297, 1139.086669921875, 433.1135559082031, -227.3615264892578, 231.7492218017578, 526.6107177734375, 1108.7523193359375, -189.55787658691406, -54.44373321533203, 1122.9996337890625, 258.993408203125, 147.64805603027344, -73.34233093261719, -65.73115539550781, 213.04653930664062, -218.48184204101562, -37.12104797363281, 369.8454895019531, 488.90289306640625, -289.3385009765625, 225.03887939453125, -39.592926025390625, 392.4669189453125, 1358.1025390625, 373.851318359375, 453.7849426269531, -144.12298583984375, 272.9710693359375, 528.2711181640625, 593.805908203125, 483.4439697265625, 80.97645568847656, 196.40658569335938, 155.3981475830078, 213.56016540527344], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000462.npy"}
{"epoch": 0.6784140969162996, "step": 463, "batch_size": 64, "mean": 303.56036376953125, "std": 412.2686767578125, "min": -517.50341796875, "p10": -207.66333923339843, "median": 276.26629638671875, "p90": 817.5169006347656, "max": 1421.976318359375, "pos_frac": 0.734375, "sample": [61.19389343261719, 552.17919921875, 1.845632553100586, 206.92063903808594, -58.76372528076172, 435.8160095214844, -248.59130859375, -19.41309356689453, 432.38800048828125, -62.27586364746094, 580.03515625, -200.74282836914062, 534.6524047851562, 619.9998779296875, 171.38238525390625, 452.5044250488281, 409.0091247558594, 501.7671203613281, 651.363525390625, 30.149852752685547, 62.61116027832031, 1421.976318359375, 474.0494689941406, 251.0662841796875, 975.0245361328125, 631.219970703125, 169.35317993164062, 289.9985046386719, -102.84109497070312, 313.0480651855469, -265.7278137207031, -51.604530334472656, -517.50341796875, 46.67407989501953, 823.2603149414062, -349.3584899902344, -127.73750305175781, -142.171142578125, 462.9397277832031, -187.53456115722656, 667.2281494140625, 352.2709655761719, 642.9465942382812, 253.0360107421875, 744.3299560546875, 1021.8823852539062, -19.75701904296875, 320.79071044921875, 947.4407348632812, -238.55519104003906, 245.7887420654297, 1268.0633544921875, 234.86720275878906, 804.1156005859375, -210.6292724609375, 107.70443725585938, 369.3574523925781, 262.5340881347656, -353.419677734375, 458.8365173339844, 472.1253356933594, 504.1235656738281, 1166.3095703125, 178.31016540527344], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000463.npy"}
{"epoch": 0.6798825256975036, "step": 464, "batch_size": 64, "mean": 443.506591796875, "std": 432.139892578125, "min": -417.7161865234375, "p10": -57.70039520263671, "median": 454.6034851074219, "p90": 948.8960693359377, "max": 1965.025390625, "pos_frac": 0.859375, "sample": [-296.39801025390625, 964.0142822265625, 357.7552795410156, 174.6883544921875, -50.09869384765625, 1002.4000244140625, 601.4951782226562, 159.2360382080078, 27.992584228515625, 511.798583984375, 676.0103149414062, 622.325927734375, -402.3678894042969, 835.250244140625, 369.5851135253906, 460.96722412109375, 638.775634765625, 187.17579650878906, 180.66845703125, 734.2635498046875, 769.04638671875, 463.17877197265625, 698.6443481445312, -153.8670196533203, 995.6204223632812, 212.73922729492188, 512.3052978515625, 517.1036376953125, -265.8857116699219, 190.58001708984375, 776.4993896484375, 228.3162841796875, 447.7413330078125, 56.06869888305664, 309.0773010253906, 603.3992919921875, 732.4530029296875, 249.50050354003906, 371.73175048828125, 172.8404541015625, 225.9024658203125, -20.488887786865234, 876.77392578125, 588.69384765625, 114.95222473144531, 496.6293029785156, -111.9732437133789, 270.19183349609375, 14.421272277832031, 620.7689208984375, 573.0886840820312, 1353.2938232421875, 1965.025390625, 1031.8343505859375, 448.23974609375, 913.6202392578125, 807.2706298828125, -60.95826721191406, 1326.99560546875, 676.6361083984375, 587.244384765625, -417.7161865234375, 379.6556396484375, 81.6867446899414], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000464.npy"}
{"epoch": 0.6813509544787077, "step": 465, "batch_size": 64, "mean": 364.3895568847656, "std": 412.2130432128906, "min": -949.7523803710938, "p10": -136.11190032958982, "median": 359.4424285888672, "p90": 794.0848571777344, "max": 1668.8031005859375, "pos_frac": 0.859375, "sample": [642.5626220703125, 1167.483642578125, 560.2324829101562, 234.78724670410156, -196.2410888671875, 124.33035278320312, 492.03753662109375, 328.3045959472656, 139.50341796875, 185.19198608398438, 723.2885131835938, 275.75384521484375, 306.02032470703125, 143.4849853515625, 559.8013916015625, 800.6760864257812, 405.9640808105469, 344.6813659667969, 214.1524200439453, 75.12895965576172, -496.3341979980469, 871.7943115234375, 582.5086059570312, 219.82986450195312, 1668.8031005859375, -66.93125915527344, 416.44158935546875, 146.8348388671875, 273.41119384765625, 438.52545166015625, 612.3203125, 466.37347412109375, -147.32553100585938, 131.7540740966797, 241.39686584472656, 513.0166625976562, 21.548439025878906, 318.4043884277344, -223.00259399414062, 566.5896606445312, 762.7625122070312, 330.4195251464844, 395.0607604980469, -557.2291870117188, 613.6697998046875, 580.6624755859375, 202.37379455566406, 626.833740234375, 95.69120788574219, 778.705322265625, 802.7491455078125, -189.4827880859375, 1082.6544189453125, -949.7523803710938, 755.8397216796875, 752.581787109375, -109.94676208496094, 332.58770751953125, 580.0623779296875, 474.6880187988281, 64.43479919433594, 374.2034912109375, 820.09619140625, 618.1611328125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000465.npy"}
{"epoch": 0.6828193832599119, "step": 466, "batch_size": 64, "mean": 298.36883544921875, "std": 387.90032958984375, "min": -719.7628173828125, "p10": -215.58688507080078, "median": 274.1829528808594, "p90": 834.4644409179689, "max": 1164.0826416015625, "pos_frac": 0.78125, "sample": [450.3188781738281, -225.8410186767578, 275.3231201171875, -218.08688354492188, 600.4949951171875, 410.7998046875, 556.992919921875, 244.80174255371094, 555.560791015625, -16.613296508789062, 214.65618896484375, 647.971435546875, -330.26470947265625, 758.2105712890625, -59.276119232177734, 198.46438598632812, 301.6513671875, 50.94954299926758, 275.7569885253906, 92.62865447998047, 464.2862243652344, 205.98654174804688, 357.51531982421875, 362.04486083984375, 812.4052124023438, 159.479736328125, 4.456703186035156, 753.65185546875, 269.75518798828125, 349.3479919433594, 351.519775390625, 168.21188354492188, 556.0028076171875, 1164.0826416015625, -719.7628173828125, 1002.9257202148438, 304.6484069824219, 762.912109375, 871.292236328125, 273.04278564453125, 843.9183959960938, 5.063800811767578, 261.328125, 166.69558715820312, 556.69287109375, 1018.3191528320312, -163.33673095703125, -370.8186950683594, -6.563728332519531, 159.322265625, 980.1884765625, 254.8272247314453, 611.8787231445312, -23.469398498535156, 703.319091796875, 382.78515625, -388.39508056640625, 11.758193969726562, 847.3136596679688, -209.75355529785156, 371.4681701660156, -125.32866668701172, -251.43594360351562, 201.52322387695312], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000466.npy"}
{"epoch": 0.684287812041116, "step": 467, "batch_size": 64, "mean": 321.11962890625, "std": 564.7387084960938, "min": -929.2738037109375, "p10": -389.83014221191405, "median": 300.0533447265625, "p90": 1135.6630981445314, "max": 1594.8626708984375, "pos_frac": 0.75, "sample": [739.5943603515625, 1277.0447998046875, 688.4678344726562, 1210.72021484375, -538.863525390625, -929.2738037109375, 484.84735107421875, 386.3428039550781, 412.4985656738281, 306.70343017578125, 232.92068481445312, 463.5491027832031, 1429.24267578125, 80.55500793457031, 516.172607421875, 411.58526611328125, 98.6830825805664, 778.15380859375, 1161.3800048828125, 412.8834228515625, 391.24755859375, -13.292377471923828, 158.89312744140625, 825.1856689453125, -281.96722412109375, -299.8189697265625, -458.1522216796875, 285.195068359375, 1075.656982421875, 336.08563232421875, 903.7210693359375, 154.40589904785156, 475.7510986328125, -371.1352844238281, 1180.8609619140625, -38.120758056640625, 1182.2777099609375, 7.552669525146484, -50.540340423583984, 709.9349975585938, 293.40325927734375, -909.6853637695312, 195.06719970703125, -204.6221923828125, 106.86631774902344, 315.8570251464844, 5.19970703125, -397.84222412109375, -425.16119384765625, 717.4741821289062, 797.1687622070312, 683.1524658203125, 1051.6951904296875, 165.44839477539062, 1021.30859375, -52.00720977783203, 608.7794799804688, 34.091156005859375, -324.321044921875, 1594.8626708984375, -879.1196899414062, 32.20266342163086, 62.36846160888672, 262.5196533203125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000467.npy"}
{"epoch": 0.6857562408223201, "step": 468, "batch_size": 64, "mean": 391.59246826171875, "std": 475.5113220214844, "min": -802.3114013671875, "p10": -85.13971710205072, "median": 332.28411865234375, "p90": 1115.8807861328128, "max": 1680.3765869140625, "pos_frac": 0.875, "sample": [410.9141845703125, 3.5803470611572266, 1370.279541015625, 500.6090087890625, 1253.703857421875, 853.7113037109375, 553.9924926757812, 726.433349609375, 451.3245544433594, 586.559326171875, 248.88258361816406, 494.69970703125, 142.82794189453125, 359.1310729980469, 358.2272644042969, 150.38409423828125, 157.0666961669922, -173.2015380859375, 457.2174987792969, 288.421875, 205.499755859375, 23.549602508544922, 84.98668670654297, 370.3706359863281, 398.9022216796875, 1052.400390625, 260.98223876953125, 1298.329833984375, 987.5509033203125, 70.59872436523438, -376.0169982910156, 6.7157745361328125, 611.7136840820312, 347.92724609375, 579.7498779296875, 547.4676513671875, 85.17871856689453, 126.04586791992188, 610.8350830078125, -802.3114013671875, 1401.9764404296875, 82.9466781616211, 457.81109619140625, 274.9371032714844, 316.6409912109375, 76.46231079101562, -222.2330322265625, 282.0666198730469, 6.6141510009765625, 1143.086669921875, 313.6658630371094, 1680.3765869140625, 203.01882934570312, 660.1016845703125, 4.020820617675781, 401.70684814453125, 763.072509765625, 1529.720703125, 449.15899658203125, -106.70196533203125, 303.222412109375, -500.4501953125, -34.82780456542969, -109.71961212158203], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000468.npy"}
{"epoch": 0.6872246696035242, "step": 469, "batch_size": 64, "mean": 387.48480224609375, "std": 438.544189453125, "min": -598.7075805664062, "p10": -16.73579177856445, "median": 341.77516174316406, "p90": 1022.056695556641, "max": 1486.928955078125, "pos_frac": 0.875, "sample": [760.4590454101562, 438.4844665527344, 117.48238372802734, 343.7583923339844, 1486.928955078125, 344.5460510253906, 870.60498046875, 50.01982879638672, -17.885604858398438, 254.8944091796875, 563.4395751953125, 18.830915451049805, -196.3782501220703, 45.827178955078125, 1063.88330078125, 43.66301727294922, 434.2538146972656, 483.5314025878906, 448.737060546875, 452.3335266113281, 921.0654296875, 267.2193298339844, 134.0713348388672, 35.99314880371094, 1159.387939453125, 2.84625244140625, 766.8047485351562, 338.5710754394531, 1216.183837890625, -209.0980224609375, 555.7725219726562, 8.425216674804688, 770.2400512695312, 312.3936462402344, 561.728759765625, 211.86170959472656, 61.425987243652344, -232.92642211914062, 12.027679443359375, -422.53851318359375, 450.1901550292969, 729.2402954101562, 1083.2786865234375, 35.77412033081055, -423.3238525390625, 174.497802734375, 290.7952575683594, 941.8319702148438, 395.7291259765625, 69.95484161376953, 867.0447998046875, 930.9771118164062, 470.8176574707031, 1056.438720703125, 1270.5474853515625, 571.167724609375, 436.7566833496094, 339.79193115234375, 83.36466979980469, -14.052894592285156, 152.15545654296875, 297.9727783203125, 707.9152221679688, -598.7075805664062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000469.npy"}
{"epoch": 0.6886930983847284, "step": 470, "batch_size": 64, "mean": 355.32403564453125, "std": 448.1801452636719, "min": -931.7890014648438, "p10": -170.99363250732418, "median": 377.931640625, "p90": 841.5333740234375, "max": 1554.569580078125, "pos_frac": 0.765625, "sample": [307.2956237792969, 531.9246215820312, -16.16534423828125, 99.84516143798828, 266.0439147949219, -78.12010955810547, 251.08819580078125, 109.8019027709961, -22.00079345703125, -563.9777221679688, 594.748046875, 532.9050903320312, 473.2204284667969, 326.2951354980469, 551.47802734375, 850.222900390625, -138.18614196777344, 624.0949096679688, -931.7890014648438, 665.0215454101562, -106.14151000976562, 505.96990966796875, 156.91546630859375, -267.5331726074219, 209.39495849609375, 163.91867065429688, -185.05398559570312, 367.24127197265625, 451.62054443359375, 317.55841064453125, 616.2570190429688, -334.4911193847656, 1069.556640625, 229.2891082763672, 34.55144500732422, 663.2910766601562, 447.687255859375, 591.3677368164062, 1554.569580078125, 588.25146484375, 1208.873046875, 144.79171752929688, 662.6956787109375, 806.7568359375, -99.89781188964844, 1255.0208740234375, 527.0225830078125, 434.6638488769531, 679.973876953125, 780.1399536132812, 388.62200927734375, -242.0162353515625, -95.4976806640625, 247.90484619140625, 534.2559814453125, 821.2578125, -334.54071044921875, 1176.1937255859375, 738.6827392578125, -79.99690246582031, 871.7852172851562, 100.31222534179688, 277.3969421386719, 428.3711242675781], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000470.npy"}
{"epoch": 0.6901615271659325, "step": 471, "batch_size": 64, "mean": 320.2552490234375, "std": 484.7325744628906, "min": -800.5458984375, "p10": -185.271337890625, "median": 232.4853973388672, "p90": 758.3551025390625, "max": 1851.5595703125, "pos_frac": 0.8125, "sample": [1188.8924560546875, -21.787643432617188, 93.22930908203125, 764.2265625, 354.8162841796875, 744.655029296875, 617.3284912109375, 1851.5595703125, 740.4246826171875, 273.7704162597656, 28.187362670898438, 241.525634765625, 8.534833908081055, 153.18887329101562, 387.82830810546875, 42.555477142333984, -243.2222900390625, -79.69456481933594, 139.239990234375, 29.45266342163086, 1730.630615234375, -181.0255584716797, 221.23309326171875, 272.7047424316406, 612.1005859375, -16.778478622436523, -308.227783203125, 47.335113525390625, -388.4013671875, 531.2185668945312, -187.09095764160156, 1041.23876953125, 452.0777893066406, 861.6209106445312, 401.2183532714844, 645.6123657226562, 547.2171630859375, 120.80372619628906, 443.31158447265625, 559.5341796875, 131.04135131835938, 223.44515991210938, 65.92003631591797, 29.790416717529297, 692.9942626953125, 1691.99072265625, 122.39544677734375, -800.5458984375, 683.9755859375, 30.753128051757812, -351.0723876953125, 140.82305908203125, 25.055978775024414, 308.05035400390625, 75.63128662109375, 591.5126342773438, 24.618173599243164, -170.43345642089844, 425.4718017578125, 307.9256286621094, 633.4973754882812, 700.8410034179688, -283.14068603515625, 474.75067138671875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000471.npy"}
{"epoch": 0.6916299559471366, "step": 472, "batch_size": 64, "mean": 376.7626647949219, "std": 607.9761962890625, "min": -1135.1539306640625, "p10": -156.8596206665039, "median": 238.22737884521484, "p90": 1213.4762695312502, "max": 2195.75439453125, "pos_frac": 0.796875, "sample": [-1135.1539306640625, -354.54034423828125, 763.728759765625, -408.0802917480469, 171.5469207763672, 64.72787475585938, 555.9385986328125, -163.83599853515625, 47.272918701171875, -539.6314086914062, 280.2065734863281, 20.220462799072266, 142.05532836914062, 404.6336975097656, 57.711212158203125, 818.2322998046875, -77.07054138183594, 509.91595458984375, 383.6392822265625, 50.904541015625, 1140.291748046875, 83.56071472167969, 125.36572265625, 15.377304077148438, 254.893310546875, 444.2420654296875, 208.87005615234375, -110.7509994506836, 240.06549072265625, 713.0642700195312, 34.9149169921875, 860.913818359375, 215.49234008789062, 997.3970947265625, 2195.75439453125, 1714.509765625, 94.28641510009766, 1448.2774658203125, 195.3567352294922, 437.4480285644531, 406.5234375, 936.990966796875, 95.86386108398438, 236.38926696777344, -293.5028076171875, 1307.15625, 1629.17333984375, 172.1703643798828, 1244.841064453125, -43.38999938964844, 2155.780517578125, -140.58140563964844, -80.01387023925781, -83.64470672607422, 932.2830810546875, 726.1455078125, 771.3486938476562, 277.0557556152344, 282.6283264160156, -420.58306884765625, 242.1911163330078, 406.7557678222656, 279.5400695800781, 169.93699645996094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000472.npy"}
{"epoch": 0.6930983847283406, "step": 473, "batch_size": 64, "mean": 201.38873291015625, "std": 770.7557373046875, "min": -1956.13232421875, "p10": -428.0656341552734, "median": 219.28704833984375, "p90": 968.3811828613282, "max": 3606.635986328125, "pos_frac": 0.6875, "sample": [-170.58444213867188, 778.5971069335938, 297.7432861328125, -230.60414123535156, -416.4699401855469, 975.7105102539062, -151.00225830078125, 169.41952514648438, 250.33265686035156, 376.5861511230469, 1180.0833740234375, 576.4428100585938, -58.73707580566406, 238.90447998046875, -414.4423828125, -433.03521728515625, 138.11190795898438, 4.7773590087890625, 173.7878875732422, 1283.7181396484375, -306.13507080078125, -71.70606994628906, 560.5940551757812, 268.2314758300781, 192.41424560546875, 166.09716796875, 436.13006591796875, 396.71673583984375, 185.66201782226562, 102.34677124023438, -1956.13232421875, 166.042724609375, -694.592041015625, 280.2816162109375, -1064.599365234375, -1709.7994384765625, 3606.635986328125, 317.79010009765625, 471.2685546875, 1186.728271484375, 253.27854919433594, 252.54510498046875, 839.6671752929688, 734.3275146484375, 227.54212951660156, 676.056640625, -546.31396484375, -411.7555847167969, 206.18943786621094, 951.2794189453125, 539.3509521484375, 1125.26025390625, -295.55316162109375, -161.03445434570312, 68.02493286132812, -171.73342895507812, 808.5908813476562, -1628.9779052734375, 500.8877868652344, 218.01223754882812, -232.74818420410156, 330.1619873046875, 220.56185913085938, 1281.943115234375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000473.npy"}
{"epoch": 0.6945668135095447, "step": 474, "batch_size": 64, "mean": 323.47344970703125, "std": 600.9982299804688, "min": -1444.4500732421875, "p10": -246.9366271972656, "median": 350.76292419433594, "p90": 910.5594238281251, "max": 2059.6181640625, "pos_frac": 0.765625, "sample": [-1353.55810546875, 262.8993835449219, 200.27316284179688, 872.8084106445312, 789.5057373046875, -502.6192321777344, 920.905029296875, 115.2407455444336, 225.09725952148438, 568.19873046875, 943.4595947265625, 2059.6181640625, 30.35014533996582, 356.0875549316406, 44.926231384277344, 886.419677734375, 380.5189514160156, -12.57952880859375, -430.37115478515625, 430.6631164550781, 22.486526489257812, 458.5567932128906, -227.88653564453125, 73.27401733398438, 274.7817077636719, 706.6463623046875, 767.4580078125, 247.73370361328125, 558.7650146484375, 345.43829345703125, 571.3491821289062, -74.2633056640625, 38.66002655029297, 369.337646484375, 544.4700927734375, 285.31390380859375, 570.19921875, 171.5069580078125, -212.8155975341797, -255.1009521484375, 717.0286254882812, -104.35028076171875, 672.8106689453125, 383.63287353515625, 179.7576904296875, 269.5882873535156, 1211.82861328125, 1172.6966552734375, -89.05696105957031, -581.0559692382812, -80.88770294189453, 800.885986328125, 401.1213073730469, 671.1209106445312, 542.28369140625, 207.2744903564453, 360.5030517578125, -889.2448120117188, 682.5238647460938, -68.12165832519531, 393.976318359375, 1311.6563720703125, -1444.4500732421875, 1957.023681640625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000474.npy"}
{"epoch": 0.6960352422907489, "step": 475, "batch_size": 64, "mean": 411.9920959472656, "std": 438.4285583496094, "min": -293.8107604980469, "p10": -137.60540924072262, "median": 401.4526672363281, "p90": 1027.8492980957037, "max": 1571.1500244140625, "pos_frac": 0.78125, "sample": [78.84353637695312, -40.5706901550293, 397.2183532714844, 142.57028198242188, -76.590576171875, 642.0053100585938, 1262.0462646484375, 389.6092224121094, 1133.8258056640625, 1098.156005859375, 306.4611511230469, 309.1103515625, 459.31182861328125, 27.2158203125, 450.8887939453125, -293.8107604980469, 713.498046875, -10.807905197143555, 697.5374145507812, -209.64724731445312, 765.645751953125, 730.0267944335938, 758.691162109375, 117.49789428710938, -29.56137466430664, 406.9277648925781, 405.6869812011719, 389.11505126953125, 57.6767692565918, 513.04248046875, 289.9073486328125, 508.2232971191406, -66.08030700683594, 355.1677551269531, 507.77178955078125, 873.4161987304688, 326.2846984863281, 753.6610107421875, -218.9199981689453, 1393.89208984375, 411.27593994140625, -19.87065887451172, 215.13046264648438, -279.6849365234375, 840.3925170898438, 83.22956848144531, 512.9210205078125, -104.92842102050781, 301.575439453125, 546.3074951171875, -151.60983276367188, 703.3736572265625, 553.3050537109375, 53.431976318359375, 1571.1500244140625, -232.08506774902344, 294.04986572265625, 1446.802978515625, 672.6951293945312, 1094.034912109375, 515.7738037109375, 807.233642578125, -197.4558563232422, 415.5025634765625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000475.npy"}
{"epoch": 0.697503671071953, "step": 476, "batch_size": 64, "mean": 343.2000732421875, "std": 466.4506530761719, "min": -538.7675170898438, "p10": -191.41544647216793, "median": 261.5976257324219, "p90": 909.4218627929689, "max": 2093.20166015625, "pos_frac": 0.8125, "sample": [683.2601318359375, 214.2095184326172, -154.77420043945312, 752.7714233398438, 287.759033203125, 522.1884765625, 1476.0384521484375, 199.61953735351562, 162.44223022460938, -51.33789825439453, 1104.8541259765625, 30.56507110595703, 216.8247528076172, 380.669189453125, 788.9795532226562, 629.6009521484375, 652.1546630859375, 121.4452133178711, 152.00344848632812, 143.17404174804688, 587.1062622070312, 330.0340576171875, -12.520584106445312, 1285.5323486328125, 528.599365234375, 532.8427734375, -506.62042236328125, -251.4366455078125, 618.41943359375, 84.30913543701172, 9.861175537109375, 216.27047729492188, -266.9750671386719, -234.04226684570312, 459.1485595703125, -94.89012145996094, -205.6888427734375, -538.7675170898438, 677.17333984375, 143.6790771484375, 162.07781982421875, 596.4381713867188, 241.316162109375, -158.11085510253906, 136.3968963623047, 174.60960388183594, 862.8424682617188, 929.3844604492188, 2093.20166015625, 1014.9354248046875, 298.25445556640625, 8.089309692382812, 291.3992004394531, 281.87908935546875, 134.3935089111328, 105.85244750976562, 517.5800170898438, 450.7108154296875, 144.6128387451172, 369.1105041503906, 436.43896484375, 1129.881103515625, 448.35125732421875, -379.3246765136719], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000476.npy"}
{"epoch": 0.6989720998531571, "step": 477, "batch_size": 64, "mean": 338.82806396484375, "std": 585.413818359375, "min": -1636.4080810546875, "p10": -253.14712829589843, "median": 260.27787017822266, "p90": 1014.8591308593751, "max": 2098.14599609375, "pos_frac": 0.75, "sample": [-544.5724487304688, 423.01727294921875, 995.0635375976562, -502.30816650390625, 1598.9390869140625, 220.0072479248047, 272.5883483886719, 861.7902221679688, 725.623291015625, 991.3684692382812, 14.152931213378906, 75.28458404541016, 227.24851989746094, 823.3677978515625, 1104.1348876953125, -155.0683135986328, 895.386962890625, 533.93603515625, 1105.770263671875, -243.92987060546875, 479.1009826660156, 4.02288818359375, -11.905715942382812, 251.14735412597656, -257.0973815917969, 745.0725708007812, 293.0080261230469, 215.97805786132812, -100.45382690429688, -30.590696334838867, -609.6688842773438, 550.7132568359375, -1636.4080810546875, -632.36328125, 1301.77099609375, 185.18118286132812, 155.09115600585938, 152.61959838867188, 664.2767944335938, 159.58474731445312, 674.6458740234375, 33.254642486572266, 581.5054321289062, 1099.4952392578125, 184.55764770507812, 370.6769714355469, 293.47161865234375, 877.329345703125, 435.6589660644531, 269.40838623046875, 1023.3429565429688, -240.21878051757812, -83.94734954833984, 166.927978515625, -5.9154510498046875, 870.65869140625, 718.9739990234375, -46.718841552734375, 41.663909912109375, -412.01190185546875, 59.540977478027344, 687.08056640625, 2098.14599609375, 686.588623046875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000477.npy"}
{"epoch": 0.7004405286343612, "step": 478, "batch_size": 64, "mean": 400.74652099609375, "std": 525.7381591796875, "min": -888.8919677734375, "p10": -268.4684188842773, "median": 343.3346405029297, "p90": 1199.9158325195317, "max": 1641.951171875, "pos_frac": 0.828125, "sample": [-244.26979064941406, 190.3456573486328, 727.1325073242188, 201.86907958984375, 41.793113708496094, 317.73681640625, 368.9324645996094, -377.048095703125, 384.42510986328125, 24.266204833984375, 481.1612548828125, 563.0808715820312, -406.3288269042969, -340.8460693359375, 403.2636413574219, 896.5388793945312, 742.3819580078125, 1641.951171875, 461.9136962890625, 140.4837646484375, -460.140869140625, -276.4047546386719, -415.83978271484375, 148.34579467773438, 425.38568115234375, 1080.171630859375, 442.79949951171875, 460.3567810058594, -72.13992309570312, 195.44960021972656, 1416.7928466796875, 1095.411865234375, 311.86151123046875, 256.09344482421875, 430.42041015625, 208.76815795898438, 1452.7940673828125, 1461.2601318359375, 964.0863037109375, -249.95030212402344, 663.6954345703125, 9.264076232910156, 124.80125427246094, -108.2208480834961, 742.8946533203125, 267.9277038574219, 135.8059844970703, 1292.218017578125, 262.66510009765625, 314.5234375, 609.8365478515625, 6.090171813964844, 258.4256591796875, 787.1043701171875, 518.5664672851562, 286.4227600097656, 1348.511962890625, 795.9951782226562, 1244.7032470703125, -888.8919677734375, 729.4985961914062, 426.2669982910156, 717.8984375, 7.47052001953125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000478.npy"}
{"epoch": 0.7019089574155654, "step": 479, "batch_size": 64, "mean": 389.60333251953125, "std": 642.262939453125, "min": -1244.4913330078125, "p10": -243.53586730957028, "median": 317.643798828125, "p90": 1107.4448486328126, "max": 3050.5986328125, "pos_frac": 0.703125, "sample": [-12.260398864746094, -16.105445861816406, 1008.70068359375, 619.003173828125, 1354.7015380859375, 1188.619140625, -506.79205322265625, -98.68209075927734, 157.4516143798828, 13.250577926635742, 689.7055053710938, -203.04898071289062, -165.32595825195312, 277.5531005859375, 771.0399780273438, 350.12078857421875, 261.41192626953125, -15.9996337890625, 1079.4913330078125, 448.0487365722656, 212.68777465820312, 518.9418334960938, -127.00001525878906, 462.6490783691406, 3050.5986328125, 971.0515747070312, 716.4567260742188, 709.8251953125, 687.202392578125, 549.7908935546875, 1139.9918212890625, 181.77874755859375, 243.88751220703125, 923.166259765625, -670.7354125976562, 1845.7578125, 106.05288696289062, 754.2330932617188, 749.317626953125, 91.95249938964844, 427.1033020019531, 1119.4249267578125, 663.6720581054688, 508.0296630859375, -36.29844665527344, -1244.4913330078125, 506.19622802734375, -322.30548095703125, -260.88739013671875, -106.57147216796875, 140.61549377441406, -64.99674224853516, -316.19232177734375, -564.2105102539062, 285.16680908203125, -174.42178344726562, 509.329345703125, -107.7113037109375, 35.8277587890625, 1347.011474609375, 372.57550048828125, 1038.282958984375, 104.51365661621094, 756.4613037109375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000479.npy"}
{"epoch": 0.7033773861967695, "step": 480, "batch_size": 64, "mean": 277.12738037109375, "std": 594.3228149414062, "min": -964.0910034179688, "p10": -573.5583557128905, "median": 215.83362579345703, "p90": 979.0830871582034, "max": 2295.4013671875, "pos_frac": 0.671875, "sample": [-655.4325561523438, 184.51817321777344, -74.77047729492188, 91.58917236328125, 606.8792724609375, 67.3175277709961, 270.35565185546875, 1003.862548828125, 631.270263671875, 716.7930908203125, 501.15081787109375, 1396.438720703125, 437.5869140625, -282.61456298828125, -680.9739379882812, -735.790771484375, 684.6829223632812, 816.5321655273438, -103.78451538085938, 425.9310302734375, 71.12093353271484, -85.76957702636719, 21.204574584960938, 349.4783935546875, 598.6464233398438, 217.34922790527344, 2295.4013671875, 281.3739318847656, 639.6836547851562, -307.576904296875, 800.9197998046875, 876.6725463867188, 921.2643432617188, -163.77085876464844, -638.9283447265625, 197.94467163085938, 506.71563720703125, -13.299301147460938, 431.95452880859375, -964.0910034179688, -807.5355224609375, -733.3273315429688, 529.5169067382812, 1146.0023193359375, 1475.0435791015625, -143.13290405273438, 214.31802368164062, 365.07659912109375, -70.29635620117188, 263.0704650878906, -9.060905456542969, -421.02838134765625, 145.4429931640625, 159.77029418945312, 190.34115600585938, -31.345623016357422, 128.5728759765625, -7.791294097900391, 604.8114013671875, 1009.4681396484375, 582.0823364257812, -191.79295349121094, 1226.0823974609375, 774.0289916992188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000480.npy"}
{"epoch": 0.7048458149779736, "step": 481, "batch_size": 64, "mean": 442.5881652832031, "std": 564.5189819335938, "min": -973.196533203125, "p10": -192.15142669677735, "median": 383.9327850341797, "p90": 1058.8685424804687, "max": 2120.46875, "pos_frac": 0.765625, "sample": [2120.46875, -973.196533203125, 569.2532958984375, 247.59234619140625, -189.6026153564453, 90.13679504394531, -110.08016967773438, 384.3630676269531, 447.5185852050781, 1465.357666015625, 172.84442138671875, -141.54251098632812, 327.594970703125, -395.5037841796875, 462.38555908203125, 881.1240234375, -158.646728515625, 764.9489135742188, 1523.5732421875, -80.95631408691406, -29.18341064453125, 758.0541381835938, 433.5244445800781, 706.8451538085938, 397.13104248046875, 201.37350463867188, 235.21435546875, 1001.7281494140625, 539.954345703125, -359.5912170410156, 65.01812744140625, 873.2889404296875, 383.50250244140625, -193.2437744140625, 714.380126953125, 810.796630859375, -331.15966796875, -77.99519348144531, 1053.285888671875, 527.6918334960938, -284.57989501953125, 136.0436248779297, 883.6026611328125, 958.4465942382812, 109.8951187133789, 180.179931640625, 1509.4476318359375, 248.32040405273438, 1672.6163330078125, 1061.2611083984375, 299.3818359375, -7.1315460205078125, 315.7139587402344, 549.1751098632812, 618.4841918945312, 177.43798828125, 262.6867980957031, 880.0812377929688, 775.0194091796875, 810.4039306640625, 1365.6475830078125, -375.26763916015625, 350.26385498046875, 710.2639770507812], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000481.npy"}
{"epoch": 0.7063142437591777, "step": 482, "batch_size": 64, "mean": 191.0445098876953, "std": 597.5621948242188, "min": -1988.06689453125, "p10": -491.04835510253906, "median": 219.32125091552734, "p90": 836.217108154297, "max": 1593.577392578125, "pos_frac": 0.734375, "sample": [237.85598754882812, -68.66148376464844, 281.7270812988281, -861.1927490234375, 424.3778076171875, 664.3792724609375, -1065.76513671875, 616.6558837890625, 617.119140625, 455.61248779296875, 280.2640075683594, 602.1553344726562, -13.38006591796875, 361.66241455078125, 597.2218627929688, 1003.5938110351562, 207.08746337890625, -291.24798583984375, 216.4104461669922, 400.38690185546875, 46.1421012878418, 358.6252746582031, 445.17022705078125, 847.6376342773438, -960.3091430664062, 222.2320556640625, 139.87835693359375, 916.755859375, 240.19236755371094, 461.23345947265625, 562.5018920898438, 591.868408203125, 651.3828735351562, 131.85263061523438, -214.92288208007812, 160.91751098632812, 1006.8640747070312, -491.4591979980469, 1326.5848388671875, -490.0897216796875, 513.2739868164062, 114.6127700805664, -156.48818969726562, 36.458656311035156, 123.38911437988281, 80.06906127929688, -1988.06689453125, 380.7239990234375, 1593.577392578125, 25.056053161621094, 14.829254150390625, -58.06501770019531, 12.564208984375, 397.6900634765625, -1180.4326171875, 263.8487243652344, 141.72781372070312, 1433.4061279296875, -183.8013916015625, 98.23777770996094, -146.355224609375, -152.78652954101562, 809.5692138671875, -565.5101928710938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000482.npy"}
{"epoch": 0.7077826725403817, "step": 483, "batch_size": 64, "mean": 449.8697509765625, "std": 573.0479736328125, "min": -645.8302001953125, "p10": -198.77911529541015, "median": 375.9490661621094, "p90": 1119.1322875976564, "max": 2901.0009765625, "pos_frac": 0.796875, "sample": [-247.13002014160156, 54.813140869140625, 1348.1846923828125, 775.7435913085938, -336.0465087890625, 855.3755493164062, 364.7274169921875, 387.17071533203125, 654.29541015625, 691.2610473632812, 435.97259521484375, 785.5770263671875, 473.104736328125, 239.49147033691406, 158.40609741210938, 144.73484802246094, 944.3619995117188, 872.3948974609375, -441.55859375, 527.2008056640625, 332.5020446777344, -12.720693588256836, 691.4434814453125, 1010.16943359375, 72.28961181640625, -203.46768188476562, 1297.5782470703125, 620.7941284179688, 39.32267761230469, 1077.9542236328125, -230.1660614013672, 1232.43505859375, 49.254722595214844, 872.73779296875, -73.63233947753906, 1136.780029296875, 277.9220886230469, 817.1742553710938, 201.79708862304688, 1052.48486328125, 753.3327026367188, 576.4550170898438, 597.4105834960938, -484.0347900390625, 76.82428741455078, 340.54937744140625, 472.2555236816406, 211.87474060058594, 1058.3048095703125, -187.83912658691406, 269.793701171875, -645.8302001953125, 130.51809692382812, 2901.0009765625, -18.762802124023438, 88.04075622558594, 1140.5830078125, 209.07281494140625, 474.86431884765625, -179.37149047851562, 697.5452880859375, 1278.0562744140625, 244.88751220703125, -164.60150146484375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000483.npy"}
{"epoch": 0.7092511013215859, "step": 484, "batch_size": 64, "mean": 398.82061767578125, "std": 493.5233459472656, "min": -662.2799072265625, "p10": -107.24089584350583, "median": 338.14093017578125, "p90": 991.1386657714845, "max": 1755.146728515625, "pos_frac": 0.75, "sample": [385.20013427734375, -213.80430603027344, 1541.22314453125, 114.97099304199219, -70.61817932128906, -234.68910217285156, 497.3297119140625, -67.58826446533203, 1755.146728515625, 219.022216796875, 97.02516174316406, 446.2259826660156, 400.11370849609375, 552.9014282226562, 832.2614135742188, 602.6948852539062, 384.3935546875, -67.4317626953125, -205.19937133789062, -662.2799072265625, 413.9950256347656, -10.737010955810547, 927.60009765625, 949.2229614257812, 1658.32861328125, 60.281219482421875, 404.10406494140625, 737.4878540039062, 33.802852630615234, -72.44821166992188, -47.01754379272461, 46.9171142578125, 798.403564453125, 1114.6004638671875, -21.477134704589844, 253.8638916015625, 286.4278564453125, 177.0090789794922, 135.4048309326172, -246.22520446777344, -49.440635681152344, 355.2248840332031, 766.0994873046875, 452.0769348144531, 1222.617919921875, 9.2642822265625, 513.7424926757812, 866.4488525390625, 321.0569763183594, -181.64352416992188, -39.51600646972656, 897.3740844726562, 784.58935546875, 804.7919921875, 122.53691101074219, 1040.08740234375, 1009.1025390625, 313.14306640625, 541.7268676757812, 109.09757232666016, 887.7152099609375, 56.30986022949219, -122.15204620361328, 937.8208618164062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000484.npy"}
{"epoch": 0.71071953010279, "step": 485, "batch_size": 64, "mean": 368.166748046875, "std": 508.7798767089844, "min": -837.5756225585938, "p10": -188.50270385742184, "median": 305.6375274658203, "p90": 1015.3333984375, "max": 1701.035400390625, "pos_frac": 0.78125, "sample": [576.517822265625, 716.6304321289062, 102.26387023925781, 318.646728515625, 346.1991271972656, 42.855751037597656, 984.03369140625, 165.48635864257812, 383.9063720703125, 442.9714660644531, 202.27822875976562, -31.82398223876953, 1275.5084228515625, 1044.7034912109375, 217.34869384765625, 1015.4389038085938, -424.42633056640625, -38.050758361816406, 988.36669921875, 960.920166015625, 901.4976806640625, 1200.1981201171875, 162.96939086914062, 434.17840576171875, 1701.035400390625, 1115.519775390625, -33.820213317871094, 26.084863662719727, 753.4921264648438, -203.85305786132812, -61.276283264160156, -52.07940673828125, -405.2694091796875, 139.423828125, 355.6138916015625, 619.2290649414062, 408.421630859375, 125.98767852783203, 153.86660766601562, 592.278564453125, -70.77178955078125, 139.24923706054688, 1688.625732421875, 515.5518188476562, -152.68521118164062, 335.1554870605469, 351.1279602050781, 500.46197509765625, 292.6283264160156, -593.5748291015625, 1015.0872192382812, 902.17431640625, 1.5714454650878906, -241.52755737304688, 29.767959594726562, -289.24951171875, 830.6729736328125, 184.34385681152344, 282.174560546875, 228.80374145507812, 39.428138732910156, -837.5756225585938, 802.6492309570312, 385.3073425292969], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000485.npy"}
{"epoch": 0.7121879588839941, "step": 486, "batch_size": 64, "mean": 430.4989013671875, "std": 612.7023315429688, "min": -752.8858642578125, "p10": -385.74214477539056, "median": 405.03016662597656, "p90": 1191.3497802734375, "max": 2758.621826171875, "pos_frac": 0.796875, "sample": [834.853515625, 510.94976806640625, 76.16230773925781, 351.1955871582031, 407.2100524902344, -604.5079345703125, 75.44400024414062, 114.82889556884766, 870.8106689453125, -194.37631225585938, 511.73583984375, 1227.5213623046875, 643.6717529296875, -708.8399047851562, -412.5379638671875, 1246.4888916015625, 908.0777587890625, 762.3253173828125, -228.1800994873047, 1017.805908203125, 1149.068603515625, -323.21856689453125, 1331.515625, 1164.472412109375, 239.3659210205078, -175.30819702148438, 583.8932495117188, 407.6690368652344, 33.416229248046875, 1253.4935302734375, 1071.0634765625, 286.850341796875, -538.4451904296875, -599.669189453125, 317.2076416015625, 355.6563415527344, 147.9593505859375, 1281.9833984375, 868.9912109375, 689.436279296875, 402.85028076171875, 635.0706787109375, 470.59918212890625, -73.31185913085938, 572.8319702148438, 1004.517333984375, 814.5924072265625, 596.2721557617188, 818.5068359375, 1202.86865234375, 134.02194213867188, 327.2497253417969, 65.90731811523438, 380.9954528808594, 591.7886352539062, 93.88441467285156, 55.61524200439453, 584.634521484375, -752.8858642578125, 85.40455627441406, 361.4149475097656, -17.792816162109375, -517.7689208984375, 2758.621826171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000486.npy"}
{"epoch": 0.7136563876651982, "step": 487, "batch_size": 64, "mean": 322.0364990234375, "std": 502.3965759277344, "min": -1571.9105224609375, "p10": -164.0370223999023, "median": 303.6696319580078, "p90": 933.3265991210939, "max": 1736.2080078125, "pos_frac": 0.75, "sample": [380.69384765625, 1191.54296875, 207.53150939941406, -202.14126586914062, 833.3395385742188, 1105.605712890625, 663.4705810546875, 481.5263366699219, 1459.5042724609375, -14.425895690917969, 49.48419189453125, -483.5882568359375, 895.719970703125, -84.5656509399414, 116.87889862060547, 1736.2080078125, 261.79412841796875, -254.4844970703125, 755.4281005859375, 790.3369140625, 410.83685302734375, 548.8262939453125, -1571.9105224609375, 124.53812408447266, -114.89464569091797, 405.58819580078125, -10.649757385253906, 949.4437255859375, 1175.8232421875, -33.603065490722656, 275.6221923828125, 320.4898986816406, 368.08319091796875, 495.3311767578125, 192.24533081054688, -465.44427490234375, 203.79470825195312, -518.6046142578125, 561.5531616210938, 336.1715393066406, 309.24896240234375, 74.1585693359375, 470.81182861328125, -178.02870178222656, 472.00482177734375, 181.9296875, 29.897628784179688, 298.0903015136719, 386.7099914550781, 239.6240234375, 129.4394989013672, 491.8153076171875, 278.91436767578125, 416.7967529296875, -56.70640563964844, -131.3897705078125, 221.00375366210938, 994.5894165039062, 710.2728881835938, 601.462890625, -15.706451416015625, 664.2302856445312, -36.329742431640625, 514.3966064453125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000487.npy"}
{"epoch": 0.7151248164464024, "step": 488, "batch_size": 64, "mean": 331.89697265625, "std": 421.5869445800781, "min": -416.9698486328125, "p10": -124.04000167846678, "median": 304.6629333496094, "p90": 929.4729431152344, "max": 1631.230224609375, "pos_frac": 0.765625, "sample": [-2.4129714965820312, 43.06894302368164, 1476.745849609375, 292.54693603515625, 907.9743041992188, 225.44102478027344, 473.91595458984375, -54.509315490722656, -74.84060668945312, 316.7789306640625, 342.5240478515625, 234.7322235107422, 484.5654602050781, 938.6866455078125, 337.12091064453125, 805.939697265625, -137.94088745117188, 16.458593368530273, -416.9698486328125, -87.72381591796875, 594.7454833984375, 14.830177307128906, 399.36376953125, 396.9710388183594, 50.57460021972656, -199.461181640625, 83.014404296875, 618.6338500976562, 1631.230224609375, 120.11238098144531, 466.09564208984375, 591.3030395507812, -387.0303649902344, 995.918212890625, 665.3259887695312, 109.62980651855469, 1060.2122802734375, 180.87826538085938, 351.2740173339844, 151.51904296875, 557.7745361328125, 831.7161865234375, 200.93157958984375, 232.34434509277344, 179.14515686035156, 390.1323547363281, 516.0693359375, 1103.8349609375, -179.26177978515625, -100.87406158447266, -74.61675262451172, 450.45489501953125, 354.1103210449219, 71.1261978149414, 375.906005859375, 511.7129211425781, -25.364850997924805, 442.7936096191406, -2.610950469970703, -133.96826171875, -363.14581298828125, 528.10107421875, 258.7974853515625, 1099.0550537109375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000488.npy"}
{"epoch": 0.7165932452276065, "step": 489, "batch_size": 64, "mean": 336.3999938964844, "std": 472.2243957519531, "min": -513.57666015625, "p10": -95.17437820434569, "median": 196.19001007080078, "p90": 1013.6443420410158, "max": 1653.9525146484375, "pos_frac": 0.703125, "sample": [754.0897216796875, -11.561203002929688, 196.3926239013672, 114.00398254394531, -31.24009132385254, 141.55755615234375, 214.5146484375, -22.661598205566406, 100.78704833984375, 195.98739624023438, -172.56153869628906, 161.14352416992188, 559.0693359375, 656.9749145507812, 399.6341857910156, 125.05155944824219, -138.87290954589844, 228.72821044921875, -1.2133865356445312, 856.8245239257812, -24.302167892456055, -34.3194580078125, 312.765869140625, 230.57373046875, -71.73202514648438, 149.70018005371094, 144.78306579589844, 27.213260650634766, 127.04728698730469, 526.3011474609375, -103.2291259765625, 49.55558776855469, -26.589235305786133, -162.83782958984375, 1216.511474609375, 561.4077758789062, 20.045299530029297, 109.09779357910156, 257.26708984375, -513.57666015625, 625.37255859375, -231.31838989257812, 1234.930419921875, -406.3662109375, 1027.74267578125, 980.7482299804688, 621.8193359375, 1610.17822265625, 312.1785888671875, 694.3311767578125, 1653.9525146484375, 629.2447509765625, 444.4803466796875, 1452.419677734375, -76.37996673583984, -13.249395370483398, 705.41943359375, 449.67742919921875, 504.0396423339844, 353.2076416015625, -2.7435455322265625, 1354.796142578125, 503.0643615722656, -20.2779541015625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000489.npy"}
{"epoch": 0.7180616740088106, "step": 490, "batch_size": 64, "mean": 340.3176574707031, "std": 608.2664184570312, "min": -1072.3033447265625, "p10": -334.0047302246094, "median": 178.16615295410156, "p90": 1127.9788208007812, "max": 1965.629638671875, "pos_frac": 0.75, "sample": [71.27120971679688, -1072.3033447265625, 1210.027587890625, -179.53428649902344, 154.09915161132812, -55.698143005371094, 154.60598754882812, 660.9955444335938, 672.03173828125, -638.9057006835938, 106.41839599609375, -89.4927978515625, 1128.94580078125, 62.364227294921875, 405.4859619140625, -330.0843200683594, -201.06112670898438, 1064.56201171875, 583.7787475585938, -636.2178955078125, 8.391490936279297, 142.97238159179688, -602.1817626953125, 104.32997131347656, -957.96826171875, 1125.7225341796875, 165.5897216796875, 509.164306640625, 427.5260314941406, 772.2966918945312, 312.13995361328125, 1965.629638671875, 172.74578857421875, 516.3593139648438, 183.58651733398438, -191.51898193359375, 1554.167724609375, 861.0921630859375, 1564.2490234375, 94.40794372558594, 138.4141845703125, 289.0364074707031, 1166.044921875, 727.0957641601562, -124.29322814941406, 783.9660034179688, -335.6849060058594, 602.499267578125, 44.20269775390625, 460.21356201171875, 292.07305908203125, -477.87213134765625, 1625.802490234375, 5.94671630859375, 995.85888671875, 930.509765625, -26.94561767578125, -5.819339752197266, 602.656982421875, 826.3242797851562, 58.44789505004883, 754.9742431640625, 499.87933349609375, 147.0079345703125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000490.npy"}
{"epoch": 0.7195301027900147, "step": 491, "batch_size": 64, "mean": 246.24530029296875, "std": 448.6011657714844, "min": -791.8199462890625, "p10": -248.12343139648434, "median": 204.82103729248047, "p90": 846.4466491699222, "max": 1629.955078125, "pos_frac": 0.6875, "sample": [320.3563232421875, -171.97767639160156, 764.8876342773438, 449.896240234375, 1284.768310546875, 59.837013244628906, 315.3015441894531, -2.3706493377685547, 304.6163024902344, 91.52874755859375, 272.6340026855469, 137.4385528564453, 153.30892944335938, -458.30218505859375, 355.5240173339844, 200.80909729003906, 641.275634765625, 380.50616455078125, -122.66740417480469, 387.5106506347656, 215.15042114257812, -60.22905731201172, 884.3260498046875, -371.6637878417969, 256.96612548828125, 256.315673828125, 662.9338989257812, -176.29039001464844, 881.4005126953125, 61.008949279785156, 1130.3623046875, 383.732666015625, -791.8199462890625, 76.22347259521484, -260.828369140625, 208.83297729492188, -35.435707092285156, 1025.0579833984375, 264.5311279296875, 480.52923583984375, 573.9569702148438, -460.148681640625, 1629.955078125, 72.32833862304688, -161.78431701660156, 255.23226928710938, -218.47857666015625, 144.0876922607422, -54.460296630859375, 683.4026489257812, 38.35405349731445, -354.5484619140625, 747.7238159179688, -69.07914733886719, -14.245794296264648, 183.6566619873047, 315.1206970214844, 630.203125, 117.62001037597656, -46.78521728515625, 1245.65869140625, -51.01740264892578, 413.50994873046875, -316.5480041503906], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000491.npy"}
{"epoch": 0.7209985315712188, "step": 492, "batch_size": 64, "mean": 245.94187927246094, "std": 528.39599609375, "min": -742.2894287109375, "p10": -394.96666564941404, "median": 147.9176025390625, "p90": 1020.2057250976562, "max": 1369.9947509765625, "pos_frac": 0.6875, "sample": [195.83474731445312, 199.9445037841797, 572.5695190429688, -70.58924102783203, 259.76446533203125, 50.91392517089844, 5.357391357421875, 759.9677734375, -555.522216796875, -129.0465545654297, 90.9916000366211, 646.62548828125, 672.7587890625, -280.4087219238281, -732.9835205078125, 885.50732421875, -166.36221313476562, 1197.479736328125, -113.75955200195312, 1333.1441650390625, -443.67828369140625, 1369.9947509765625, 293.1321105957031, -369.6612854003906, 369.6898193359375, 330.0423583984375, 271.4642639160156, -358.891845703125, 78.25309753417969, 233.52174377441406, 39.28778076171875, 33.511314392089844, 95.85045623779297, -742.2894287109375, -57.65250015258789, 1090.6082763671875, 840.5394287109375, 316.372802734375, -666.6160278320312, 187.2864990234375, 1019.6734619140625, -715.82080078125, 1172.151123046875, -8.106681823730469, -78.96144104003906, 734.506591796875, 222.98719787597656, 80.273193359375, 923.4032592773438, 92.76876831054688, 432.5251770019531, -65.55564880371094, 85.02361297607422, -405.81182861328125, 1020.433837890625, 890.0045166015625, 475.82330322265625, 98.54435729980469, -8.300701141357422, 517.8541259765625, -266.2078552246094, 1283.3072509765625, 108.5487060546875, 398.263916015625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000492.npy"}
{"epoch": 0.7224669603524229, "step": 493, "batch_size": 64, "mean": 267.3201599121094, "std": 489.1380310058594, "min": -1702.092529296875, "p10": -170.25860443115232, "median": 224.43500518798828, "p90": 937.3139343261721, "max": 1368.2587890625, "pos_frac": 0.703125, "sample": [995.955810546875, 141.5377960205078, 413.3205261230469, 285.61126708984375, -97.61651611328125, 1147.12548828125, 758.8694458007812, -70.91506958007812, 217.05197143554688, -426.97601318359375, 431.00341796875, 21.557754516601562, 774.9363403320312, 417.7724609375, 486.86724853515625, 129.13064575195312, -314.5991516113281, 383.7652893066406, 169.37722778320312, 41.30284881591797, 231.8180389404297, 311.41741943359375, 461.283935546875, 423.83477783203125, -161.60711669921875, 404.1771240234375, 646.2207641601562, -156.94451904296875, -201.70379638671875, -111.45206451416016, 113.5902099609375, -1702.092529296875, 382.2923278808594, -8.912368774414062, 678.2496337890625, -52.761695861816406, 1268.907958984375, 1247.2607421875, 140.61614990234375, -187.54539489746094, -157.52590942382812, 388.79833984375, 1055.888671875, 681.7818603515625, 342.3623962402344, 959.052490234375, -501.1245422363281, 421.12725830078125, 1368.2587890625, 886.5906372070312, 186.8651123046875, 650.184814453125, -173.9663848876953, 207.63531494140625, 261.4281005859375, -117.0350112915039, -88.0557632446289, 238.3656768798828, 638.71484375, 177.68350219726562, 120.62943267822266, -61.02239990234375, -80.64834594726562, 70.77357482910156], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000493.npy"}
{"epoch": 0.723935389133627, "step": 494, "batch_size": 64, "mean": 419.6595458984375, "std": 458.2921447753906, "min": -601.1669921875, "p10": -11.381896781921382, "median": 369.4533386230469, "p90": 1028.8653076171875, "max": 1711.1170654296875, "pos_frac": 0.859375, "sample": [-2.3894710540771484, 290.4290771484375, 540.9431762695312, 365.9444274902344, 1711.1170654296875, 1181.6016845703125, -601.1669921875, 709.7158813476562, 592.4883422851562, -13.346742630004883, 449.1130676269531, 67.85111999511719, 772.39306640625, 337.57037353515625, -260.6495361328125, 1079.8143310546875, 634.7327270507812, -121.81378173828125, 392.5361022949219, 780.4102783203125, 141.21170043945312, -490.98382568359375, 253.46743774414062, 805.6478881835938, 791.0628662109375, 77.7879867553711, 810.2654418945312, 885.97509765625, 178.98155212402344, 258.9867858886719, -528.8455810546875, 49.13356018066406, 368.1454772949219, -156.6959228515625, 1563.9678955078125, 387.4590148925781, 312.80023193359375, 105.53978729248047, 451.15216064453125, 1243.58984375, 1030.134521484375, 390.41162109375, -6.7972564697265625, 4.383842468261719, 365.31304931640625, 655.3360595703125, 781.1293334960938, 1025.90380859375, 370.7611999511719, 667.64501953125, 192.67176818847656, 187.3190460205078, 534.3734741210938, 78.32144165039062, 151.3575897216797, 493.373046875, 770.3236694335938, 556.214599609375, 120.34020233154297, 433.7451171875, 1191.4051513671875, 78.29002380371094, 315.7569885253906, 54.553985595703125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000494.npy"}
{"epoch": 0.7254038179148311, "step": 495, "batch_size": 64, "mean": 369.12579345703125, "std": 537.7481079101562, "min": -988.8880004882812, "p10": -107.65750350952148, "median": 298.54505920410156, "p90": 1026.857147216797, "max": 2278.252685546875, "pos_frac": 0.796875, "sample": [427.38372802734375, -49.54966735839844, -28.77984619140625, -71.4403076171875, 196.35009765625, 283.2734375, -335.60009765625, 45.2359619140625, 300.3597412109375, 32.03028869628906, 148.49256896972656, 657.4391479492188, 793.4481201171875, -430.3457946777344, 384.13885498046875, 438.4700622558594, -189.72201538085938, 792.6858520507812, 81.87045288085938, 629.800048828125, 150.325439453125, 36.912933349609375, 302.3697814941406, 38.624691009521484, 291.7580261230469, 1092.04736328125, 973.5250854492188, 622.097900390625, 145.03713989257812, 439.817138671875, 268.1385498046875, 309.1304931640625, 165.17672729492188, 422.4719543457031, 1049.7137451171875, 912.1629028320312, 816.56298828125, -110.78096008300781, 2278.252685546875, 1347.9903564453125, 358.0668640136719, -46.00511169433594, 210.80703735351562, 590.166748046875, 387.67254638671875, -604.0201416015625, 14.066696166992188, 296.7303771972656, 78.74874114990234, -250.2719268798828, -100.36943817138672, -8.29156494140625, 209.0764617919922, 489.09576416015625, 1270.3804931640625, 755.8155517578125, 301.5026550292969, 713.7830200195312, -988.8880004882812, 1138.3934326171875, 897.6202392578125, 303.7808837890625, 1910.5697021484375, 38.743873596191406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000495.npy"}
{"epoch": 0.7268722466960352, "step": 496, "batch_size": 64, "mean": 203.5760498046875, "std": 415.4187927246094, "min": -774.1094360351562, "p10": -358.9038421630859, "median": 136.20486450195312, "p90": 789.9075683593751, "max": 1106.1915283203125, "pos_frac": 0.765625, "sample": [124.27536010742188, -423.0567321777344, 970.3070068359375, 833.1949462890625, -701.4539184570312, 729.9515991210938, 88.23404693603516, 310.2760009765625, 776.2806396484375, 66.06838989257812, 1106.1915283203125, 296.07940673828125, 392.5611267089844, 247.01385498046875, -188.4044189453125, 47.213871002197266, -197.7364501953125, 482.5494384765625, -589.5208740234375, 1060.44580078125, 602.2652587890625, 377.3738098144531, 795.7476806640625, 310.37322998046875, 838.994140625, 283.79632568359375, -47.97007751464844, 12.284286499023438, 72.12631225585938, -370.2125549316406, 95.99022674560547, 129.80296325683594, 155.66134643554688, 370.8634338378906, 63.88817596435547, 1.0173492431640625, 474.73028564453125, 103.7880630493164, -561.7517700195312, 245.0663604736328, -414.855712890625, -332.516845703125, 751.7179565429688, 3.2087173461914062, 142.6067657470703, 538.9681396484375, 77.30781555175781, 451.0563049316406, 593.3272094726562, 491.3097839355469, 93.40109252929688, -206.6882781982422, 470.242431640625, -7.174631118774414, 234.06790161132812, -774.1094360351562, 91.50135803222656, 852.3943481445312, 14.594535827636719, -26.235557556152344, 420.8683166503906, 223.6111602783203, -62.15424346923828, 18.11297607421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000496.npy"}
{"epoch": 0.7283406754772394, "step": 497, "batch_size": 64, "mean": 190.6160888671875, "std": 443.15484619140625, "min": -928.5642700195312, "p10": -258.6882873535156, "median": 166.38233184814453, "p90": 767.6532714843751, "max": 1226.3240966796875, "pos_frac": 0.65625, "sample": [-124.52754211425781, 792.581787109375, 606.3695678710938, 621.37548828125, 10.122049331665039, -242.04205322265625, 528.2659912109375, -227.4910430908203, -76.7987060546875, -122.20237731933594, 241.4204559326172, 831.530029296875, 1117.173095703125, 424.4825439453125, 49.14514923095703, 160.82794189453125, 389.125, 201.66160583496094, 610.1106567382812, 488.5777893066406, -94.55247497558594, 10.851387023925781, 458.2294006347656, 774.829833984375, -88.84007263183594, 56.22541809082031, 730.1942138671875, -125.28273010253906, -22.303421020507812, -646.2196044921875, 454.0146484375, 76.25096130371094, -5.800212860107422, 364.5321044921875, 66.394287109375, 1142.243408203125, -65.267578125, 1226.3240966796875, 50.31880187988281, 74.15017700195312, 165.67123413085938, 246.01478576660156, 239.71083068847656, 750.907958984375, -906.741455078125, 169.97567749023438, 544.5819702148438, -361.0117492675781, 298.4180603027344, 217.59115600585938, 287.4006652832031, 167.0934295654297, 965.7080688476562, -215.3321990966797, -184.50730895996094, 285.19647216796875, -277.45654296875, -265.8223876953125, -55.25727844238281, -928.5642700195312, 606.3548583984375, -519.0276489257812, 348.7966613769531, -96.27085876464844], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000497.npy"}
{"epoch": 0.7298091042584435, "step": 498, "batch_size": 64, "mean": 314.157958984375, "std": 542.7174682617188, "min": -1289.7606201171875, "p10": -273.43362121582027, "median": 337.1105194091797, "p90": 997.1360168457036, "max": 1767.2158203125, "pos_frac": 0.71875, "sample": [-733.3981323242188, 188.04971313476562, 681.4007568359375, 310.7026062011719, -162.04867553710938, 691.17724609375, 719.4762573242188, 775.5409545898438, 357.759521484375, 698.431640625, 412.55584716796875, 702.1153564453125, 281.64276123046875, 825.4904174804688, -1289.7606201171875, 381.28753662109375, 886.3474731445312, 1171.5028076171875, -78.26873016357422, 347.12506103515625, 160.16668701171875, 713.489013671875, -115.846435546875, 1078.21337890625, -256.26373291015625, 280.7310485839844, 385.846923828125, 111.46790313720703, -398.52972412109375, -10.832695007324219, 292.6127624511719, 364.15057373046875, -1024.3077392578125, 1258.092529296875, 496.3988952636719, 17.5601806640625, -152.33639526367188, 440.28253173828125, 653.9170532226562, 452.684326171875, 1242.814697265625, 9.922798156738281, -5.49176025390625, 67.53120422363281, 544.369873046875, 570.6759033203125, -480.8921203613281, 101.50729370117188, 1130.1119384765625, 847.6552124023438, 244.10874938964844, -227.9969482421875, 327.0959777832031, -95.38507843017578, -282.8983154296875, -48.674224853515625, 477.35302734375, 122.42670440673828, 852.9049682617188, 394.49005126953125, -280.7921447753906, 1044.6168212890625, 1767.2158203125, -131.18783569335938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000498.npy"}
{"epoch": 0.7312775330396476, "step": 499, "batch_size": 64, "mean": 411.035888671875, "std": 531.9590454101562, "min": -992.4893188476562, "p10": -147.66192169189452, "median": 490.4496307373047, "p90": 926.7003662109375, "max": 2024.513427734375, "pos_frac": 0.765625, "sample": [910.56396484375, 103.91548156738281, 246.65771484375, -992.4893188476562, -61.762474060058594, 494.4253845214844, 487.9927062988281, 573.0609130859375, -86.03144073486328, -114.94302368164062, 610.841552734375, 592.2289428710938, 159.16571044921875, 933.615966796875, 604.7229614257812, 266.2034606933594, 647.1278686523438, -545.540771484375, 2024.513427734375, 280.927001953125, 218.34384155273438, 878.6055908203125, 647.8040771484375, 43.28331756591797, -464.43963623046875, 512.34765625, 267.19464111328125, 791.558349609375, -307.54583740234375, 1175.6895751953125, 184.61590576171875, -15.15130615234375, 590.1417236328125, -116.0595474243164, 288.6006164550781, -13.995107650756836, 665.8157958984375, 876.8150634765625, 767.4547119140625, -187.1161346435547, 44.87928771972656, 492.90655517578125, 706.8089599609375, 672.5355834960938, 642.3965454101562, 200.37237548828125, 611.1758422851562, 1843.97998046875, 1196.1544189453125, 118.7853775024414, 511.2297668457031, 708.6298828125, 451.5613708496094, 783.0478515625, -95.50372314453125, 792.3284912109375, 1422.7576904296875, 357.63214111328125, -516.6348876953125, -126.52967834472656, -156.71859741210938, 700.3943481445312, 2.361227035522461, 1002.5857543945312], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000499.npy"}
{"epoch": 0.7327459618208517, "step": 500, "batch_size": 64, "mean": 339.9359130859375, "std": 523.6790771484375, "min": -1022.9015502929688, "p10": -327.6290130615233, "median": 243.02774047851562, "p90": 990.4715942382815, "max": 1747.6717529296875, "pos_frac": 0.78125, "sample": [-8.666292190551758, -221.5583953857422, -482.16510009765625, 1747.6717529296875, -425.1007385253906, 467.8394470214844, 1014.91064453125, 498.7278747558594, 818.2688598632812, 6.818717956542969, 859.026123046875, 683.1385498046875, 856.245849609375, -435.1292419433594, -20.624757766723633, 474.3302917480469, 36.9407958984375, 58.51953125, 220.1154022216797, 1144.639404296875, 731.6800537109375, 42.33290100097656, 684.8937377929688, 1365.26953125, 0.41048431396484375, 776.446533203125, -372.55645751953125, 757.695068359375, 834.1336059570312, -707.7335815429688, 767.9229736328125, 1025.8157958984375, -51.21748352050781, 124.47991180419922, 228.32269287109375, 738.7711181640625, -612.991943359375, 110.94669342041016, 933.4471435546875, -1022.9015502929688, -97.63434600830078, 158.82005310058594, 227.7745361328125, 637.5146484375, 73.7404556274414, -56.98560333251953, 598.7420654296875, 82.13744354248047, 102.5452880859375, 257.7327880859375, 1030.2811279296875, 634.0283813476562, 682.4146728515625, 195.90805053710938, 545.3219604492188, 503.1432800292969, 603.3966064453125, 372.4490966796875, 9.097707748413086, 520.297119140625, 10.323163986206055, 1142.873779296875, -222.79830932617188, 95.65748596191406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000500.npy"}
{"epoch": 0.7342143906020558, "step": 501, "batch_size": 64, "mean": 359.9496154785156, "std": 541.607666015625, "min": -828.0913696289062, "p10": -268.04286499023436, "median": 349.5670166015625, "p90": 1102.6743896484377, "max": 1699.7650146484375, "pos_frac": 0.78125, "sample": [32.069480895996094, 343.71710205078125, 318.04388427734375, -320.5192565917969, -59.324832916259766, 1164.1546630859375, 217.24610900878906, 958.6006469726562, 271.53692626953125, 65.44760131835938, 144.32720947265625, -167.04563903808594, 380.62274169921875, 221.30775451660156, 1699.7650146484375, 648.6322631835938, 1005.796630859375, 410.85150146484375, 355.41693115234375, 771.5275268554688, 1114.014404296875, -112.70323181152344, 314.60076904296875, 168.67684936523438, 781.1181030273438, -406.6401672363281, -612.7264404296875, 378.88055419921875, 1012.738037109375, 509.9770202636719, -732.6112060546875, 273.398193359375, 421.926513671875, 506.89013671875, 298.6051025390625, 1368.8077392578125, -195.0337371826172, 532.9203491210938, 389.3302917480469, 787.0499877929688, 1166.068115234375, 15.579185485839844, -15.259870529174805, -272.4817810058594, 570.7788696289062, 128.66278076171875, 2.1639137268066406, 590.21240234375, 29.01296615600586, -828.0913696289062, 359.2191162109375, 631.8134765625, -257.6853942871094, -99.28927612304688, 1.7787513732910156, -642.8485107421875, 1076.21435546875, 853.0036010742188, 1363.103759765625, 390.40411376953125, 1607.7772216796875, 530.201904296875, 467.067626953125, 107.97615051269531], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000501.npy"}
{"epoch": 0.73568281938326, "step": 502, "batch_size": 64, "mean": 366.5157470703125, "std": 445.8046569824219, "min": -404.61474609375, "p10": -36.04360427856445, "median": 333.7083740234375, "p90": 936.0600463867188, "max": 2492.7080078125, "pos_frac": 0.8125, "sample": [516.4361572265625, -236.66819763183594, 822.556396484375, 600.5897827148438, -26.43023681640625, 591.0978393554688, -78.5965805053711, -44.4454345703125, 1197.077392578125, 382.7444763183594, 407.3284912109375, 217.60287475585938, 351.89422607421875, 1088.481689453125, 860.42529296875, -1.5821533203125, 3.21209716796875, 245.6475067138672, -404.61474609375, 166.38131713867188, 460.8473205566406, 1075.4559326171875, 183.29541015625, 339.18499755859375, 35.17333984375, 195.53038024902344, 88.82881164550781, 324.8588562011719, 454.92156982421875, 401.54034423828125, -242.70677185058594, 1002.1669311523438, -9.277534484863281, 438.5242919921875, 479.3165588378906, 937.145751953125, -15.90418815612793, 2492.7080078125, 12.666976928710938, -37.35319519042969, 577.1378173828125, 96.30888366699219, 220.9295654296875, 390.48907470703125, 339.0421447753906, 543.6105346679688, 989.4373168945312, 28.527841567993164, 668.94287109375, 426.21917724609375, 710.232666015625, 17.162399291992188, 30.40363121032715, 115.5794677734375, 933.5267333984375, -91.51759338378906, 335.2364501953125, 63.623870849609375, 665.4658203125, 303.1859130859375, -32.987892150878906, 378.5799560546875, 139.62831115722656, 332.1802978515625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000502.npy"}
{"epoch": 0.737151248164464, "step": 503, "batch_size": 64, "mean": 338.29443359375, "std": 441.872802734375, "min": -360.7088317871094, "p10": -92.98617782592774, "median": 230.79865264892578, "p90": 838.5323669433594, "max": 1879.8602294921875, "pos_frac": 0.765625, "sample": [15.686717987060547, -88.39629364013672, 979.0127563476562, -61.55171203613281, -93.42951965332031, -177.74374389648438, 484.79034423828125, 239.8993682861328, 108.48674774169922, -91.95171356201172, -316.61358642578125, 299.27471923828125, -224.66964721679688, 224.15769958496094, 995.7691040039062, 656.2640380859375, 302.8711242675781, 650.953369140625, 139.8289794921875, 838.234619140625, -82.49757385253906, 144.4818878173828, 67.77427673339844, 55.65326690673828, 404.1585693359375, -18.304733276367188, -123.43419647216797, 469.70086669921875, 206.29116821289062, 199.4185791015625, 397.04412841796875, -16.57220458984375, 116.10902404785156, 582.052490234375, 15.153419494628906, -96.15044403076172, 1879.8602294921875, 1658.179931640625, 489.908935546875, 391.61517333984375, 139.30487060546875, 630.8819580078125, 670.6500854492188, 76.30242156982422, -90.30746459960938, 654.7234497070312, 214.0493621826172, 813.04345703125, 208.21231079101562, 385.20013427734375, 93.20218658447266, 622.5367431640625, 451.3477783203125, 172.89637756347656, 1018.8244018554688, 568.2867431640625, -82.8936538696289, 687.1170654296875, -360.7088317871094, 1426.62646484375, 237.43960571289062, 415.47265625, 838.6599731445312, 238.65835571289062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000503.npy"}
{"epoch": 0.7386196769456681, "step": 504, "batch_size": 64, "mean": 393.63238525390625, "std": 514.9931640625, "min": -570.77099609375, "p10": -291.05259704589844, "median": 311.44606018066406, "p90": 1030.9860961914064, "max": 1776.4219970703125, "pos_frac": 0.78125, "sample": [101.16989135742188, 138.775634765625, 1062.949462890625, 306.64044189453125, 109.20753479003906, 247.51480102539062, 626.621826171875, -221.91256713867188, 1776.4219970703125, 1167.0753173828125, 1626.84765625, -531.4334106445312, 496.1905822753906, 540.5995483398438, -38.14888000488281, -17.037246704101562, 559.020263671875, 211.33447265625, 20.41527557373047, -89.49710083007812, 116.62602996826172, -17.311500549316406, 351.13580322265625, -299.5611572265625, 292.2707824707031, 210.8517608642578, -389.29052734375, 512.78662109375, 223.29214477539062, 368.58026123046875, -329.6813049316406, 1005.950927734375, -314.5747985839844, 1367.2139892578125, 953.853515625, 629.3316650390625, 753.5069580078125, 575.3368530273438, 176.89743041992188, 316.2516784667969, 153.08038330078125, 1040.06494140625, 730.1524047851562, 490.0720520019531, 842.5322875976562, -148.40621948242188, -272.2726745605469, 616.0785522460938, 73.96748352050781, 928.4815673828125, 168.94024658203125, 678.8350219726562, 1009.8021240234375, 668.6143798828125, 335.7447509765625, 277.217529296875, -299.10113525390625, 909.5264892578125, 1527.031982421875, 292.05072021484375, -570.77099609375, 497.4107666015625, 634.9522094726562, 12.2440185546875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000504.npy"}
{"epoch": 0.7400881057268722, "step": 505, "batch_size": 64, "mean": 222.71705627441406, "std": 404.29095458984375, "min": -661.0847778320312, "p10": -213.6968002319336, "median": 168.23545837402344, "p90": 766.4131286621094, "max": 1151.777099609375, "pos_frac": 0.6875, "sample": [10.251365661621094, 586.0584716796875, -192.0301513671875, 501.85693359375, -48.34956359863281, 881.0201416015625, 59.67335510253906, -1.2915496826171875, 106.94964599609375, 3.7657928466796875, 797.8845825195312, 733.2698974609375, 1151.777099609375, -19.431884765625, 654.7842407226562, 432.31085205078125, -83.84880065917969, 311.37396240234375, 169.87176513671875, -204.72425842285156, 230.61289978027344, 649.513671875, 181.94866943359375, 81.440673828125, -217.54217529296875, -625.5555419921875, 153.8400421142578, 291.1675720214844, 744.3707885742188, 348.8056640625, -127.05439758300781, 388.01519775390625, 407.885986328125, -268.5124816894531, 170.33596801757812, 686.949951171875, 77.53360748291016, 746.8986206054688, -63.56078338623047, 413.4056396484375, -661.0847778320312, -584.5410766601562, 161.5800018310547, 312.31658935546875, 13.36871337890625, 110.45195770263672, 1004.9955444335938, 499.5495910644531, -141.84617614746094, 747.978271484375, 219.6543426513672, -134.4700927734375, -354.7200622558594, 774.3137817382812, 203.1315155029297, -66.85685729980469, 802.284423828125, -356.7249755859375, -127.37799072265625, 141.4389190673828, 872.4853515625, 166.59915161132812, 587.060791015625, -57.3663444519043], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000505.npy"}
{"epoch": 0.7415565345080763, "step": 506, "batch_size": 64, "mean": 342.12078857421875, "std": 455.77642822265625, "min": -645.5447387695312, "p10": -192.79783782958978, "median": 304.84352111816406, "p90": 966.5874267578126, "max": 1798.57421875, "pos_frac": 0.78125, "sample": [-332.23419189453125, 36.212860107421875, 593.4482421875, 22.901473999023438, 242.1779022216797, 303.259521484375, -222.03883361816406, 444.630126953125, 869.432373046875, 38.196468353271484, 569.77490234375, 200.7389373779297, 268.425537109375, 206.4532012939453, -500.78033447265625, 930.0189208984375, 464.8778076171875, -49.31196594238281, 306.4275207519531, 643.2722778320312, 365.88177490234375, 1121.15673828125, 387.9590148925781, 126.23546600341797, 1798.57421875, -78.03518676757812, -220.69488525390625, 465.14794921875, 574.6799926757812, 289.6513671875, 330.82958984375, 1243.5997314453125, 421.8997497558594, 106.71220397949219, 577.5289916992188, -645.5447387695312, 425.1348571777344, 162.4837646484375, 982.2596435546875, 821.98095703125, 523.337890625, 571.2869873046875, -127.70472717285156, 416.60284423828125, 141.30833435058594, -592.6203002929688, 178.09445190429688, 3.2242050170898438, 1006.5472412109375, 991.0516357421875, 1005.1873779296875, -3.7583751678466797, 278.838623046875, 718.2788696289062, 686.5490112304688, 879.9177856445312, -27.358779907226562, 753.36572265625, 192.05015563964844, -22.177406311035156, -336.9693603515625, 74.00857543945312, 389.9457092285156, -96.59940338134766], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000506.npy"}
{"epoch": 0.7430249632892805, "step": 507, "batch_size": 64, "mean": 301.274169921875, "std": 524.9344482421875, "min": -845.6919555664062, "p10": -215.46004180908199, "median": 243.70167541503906, "p90": 769.2303588867188, "max": 2205.810546875, "pos_frac": 0.75, "sample": [-14.974069595336914, 244.1865692138672, 13.742477416992188, -845.6919555664062, 1062.548095703125, 40.489898681640625, 69.77861022949219, 402.69677734375, 156.77188110351562, 179.78204345703125, -28.160110473632812, 328.127685546875, 42.56135559082031, 306.345947265625, 641.734130859375, 667.8543090820312, -304.2847900390625, 282.50152587890625, 75.12361145019531, 2012.7847900390625, 486.7161865234375, 629.7805786132812, 29.532604217529297, -25.375816345214844, 421.2876281738281, 336.48626708984375, 513.50146484375, 659.9912719726562, -381.2791442871094, 278.96044921875, 243.21678161621094, -10.95111083984375, -156.726318359375, 8.67750358581543, 106.0337905883789, 506.0451354980469, 483.257568359375, 1380.2930908203125, 552.6133422851562, 103.42113494873047, 256.96014404296875, 19.040054321289062, -10.57562255859375, 329.1948547363281, 2205.810546875, 821.05908203125, 197.95596313476562, 267.45355224609375, -276.0685729980469, 571.1726684570312, 738.734130859375, -476.2098388671875, -38.524261474609375, 705.0056762695312, 1514.96826171875, 531.6300048828125, 782.3001708984375, 390.5428771972656, 241.4715576171875, -54.08763885498047, 32.13317108154297, -95.65670776367188, -240.6316375732422, -631.5327758789062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000507.npy"}
{"epoch": 0.7444933920704846, "step": 508, "batch_size": 64, "mean": 250.27903747558594, "std": 547.1024169921875, "min": -1455.8714599609375, "p10": -336.32742919921867, "median": 230.16346740722656, "p90": 830.6053649902345, "max": 1950.5718994140625, "pos_frac": 0.671875, "sample": [196.75677490234375, 170.96234130859375, -695.99365234375, 636.6236572265625, 507.52313232421875, -32.55712890625, 104.96279907226562, 1794.201904296875, 411.7263488769531, 692.8263549804688, 305.9054260253906, 877.9810791015625, 289.1271667480469, 1950.5718994140625, 501.6846923828125, 302.0289306640625, 178.95761108398438, -372.47735595703125, 1141.6949462890625, 847.2384643554688, -427.8287353515625, -110.61314392089844, 283.272216796875, -778.5275268554688, 907.7413330078125, -233.30697631835938, 105.83685302734375, -251.97760009765625, -91.19453430175781, 299.71624755859375, -92.82765197753906, 386.64166259765625, -7.5280609130859375, -29.805187225341797, 1380.597412109375, -1455.8714599609375, 97.4024887084961, 144.8403778076172, 184.15097045898438, 791.7947998046875, 263.5701599121094, 454.01885986328125, 111.83572387695312, 325.91058349609375, 561.8753051757812, 186.02435302734375, 438.4393615722656, 413.933349609375, 709.0502319335938, 567.8461303710938, 423.6658020019531, -99.97543334960938, -227.4673614501953, 770.402587890625, 605.2950439453125, -140.26641845703125, -121.12741088867188, -49.45965576171875, -418.85150146484375, 573.78125, -403.43896484375, 324.33599853515625, 6.295995712280273, -170.09481811523438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000508.npy"}
{"epoch": 0.7459618208516887, "step": 509, "batch_size": 64, "mean": 281.5257568359375, "std": 412.03369140625, "min": -518.079833984375, "p10": -280.4738464355469, "median": 243.83392333984375, "p90": 843.5683898925782, "max": 1576.305419921875, "pos_frac": 0.8125, "sample": [-33.99591827392578, -78.66259765625, -518.079833984375, 749.77099609375, -135.54736328125, 423.1139831542969, 75.09101104736328, 163.85369873046875, 488.72515869140625, 292.8752136230469, 558.7320556640625, 153.3978271484375, 295.05389404296875, 135.69741821289062, 219.78065490722656, 140.06494140625, 850.256591796875, -436.017333984375, -365.1630859375, 353.0567321777344, 1024.969482421875, 475.2680969238281, 153.90492248535156, 271.15692138671875, 437.00518798828125, 653.7879028320312, 85.76129150390625, 157.12509155273438, 307.2948913574219, 423.5295104980469, 718.843505859375, -332.60888671875, 136.35987854003906, -288.1202392578125, 459.6612854003906, 883.6342163085938, 645.52392578125, 80.46430969238281, 251.45834350585938, -182.32041931152344, 6.055755615234375, 336.47528076171875, 276.254638671875, 116.76004028320312, 47.566619873046875, -507.27130126953125, 1019.1182861328125, 1164.2742919921875, 16.825912475585938, 1576.305419921875, -262.63226318359375, 138.675537109375, -317.3713073730469, 115.28945922851562, 236.20950317382812, 366.05621337890625, 215.59420776367188, 152.73684692382812, 299.364501953125, 453.13629150390625, 697.2509155273438, 945.4962158203125, 402.81158447265625, 827.9625854492188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000509.npy"}
{"epoch": 0.7474302496328928, "step": 510, "batch_size": 64, "mean": 364.1890563964844, "std": 373.8440856933594, "min": -459.8821105957031, "p10": -62.73488235473632, "median": 350.72314453125, "p90": 756.9151855468751, "max": 1245.247314453125, "pos_frac": 0.796875, "sample": [251.87905883789062, 708.5845947265625, 1118.157958984375, 418.6843566894531, -139.75119018554688, 642.3837890625, 702.3360595703125, 386.40423583984375, -30.496597290039062, 765.6372680664062, 331.819580078125, 457.0097961425781, -226.4007110595703, 438.541015625, 1004.56787109375, 695.1047973632812, 681.79345703125, -459.8821105957031, -53.23674774169922, 720.28759765625, 303.2348327636719, 280.5263671875, 406.69793701171875, 158.17233276367188, 276.0491638183594, 612.1134033203125, 546.0419311523438, 870.2955932617188, 1010.5299682617188, -256.68743896484375, 598.2376708984375, 1245.247314453125, -0.3725318908691406, 179.91253662109375, 532.38916015625, 199.05886840820312, 300.8174133300781, 1115.253173828125, -255.2654266357422, 288.957275390625, 261.98626708984375, 92.15933990478516, 637.93603515625, 309.5240478515625, 437.93231201171875, 21.172874450683594, 493.3797912597656, 369.626708984375, 74.27763366699219, 600.4645385742188, 728.0601806640625, -31.79758644104004, 125.69329833984375, 736.5636596679688, -57.08992004394531, 409.7568664550781, -271.7943420410156, 686.9344482421875, -45.42504119873047, -65.1541519165039, 24.66416358947754, 656.1664428710938, 249.50686645507812, 38.92153549194336], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000510.npy"}
{"epoch": 0.748898678414097, "step": 511, "batch_size": 64, "mean": 281.7838134765625, "std": 488.244140625, "min": -1201.8525390625, "p10": -181.68356628417956, "median": 196.12075805664062, "p90": 819.7486267089846, "max": 1520.1788330078125, "pos_frac": 0.796875, "sample": [-73.19823455810547, 1166.0321044921875, 117.81216430664062, 121.49885559082031, 347.9064636230469, 166.22691345214844, 846.9277954101562, 1370.282958984375, 328.4132080078125, 171.42051696777344, 112.03162384033203, 487.59002685546875, 701.4629516601562, 685.6072998046875, 54.42616271972656, 1067.9541015625, -1185.4510498046875, 438.4530029296875, -233.751220703125, 663.2268676757812, 1193.575927734375, 662.6655883789062, 83.52467346191406, 412.2733154296875, 211.16598510742188, 98.9441909790039, -554.2496337890625, 370.92938232421875, -43.410430908203125, 525.52880859375, 474.383056640625, 18.939651489257812, -26.1404972076416, 236.7703399658203, 285.8320007324219, 251.8683319091797, 728.6797485351562, -245.26791381835938, 7.195068359375, 108.78832244873047, 756.33056640625, 181.07553100585938, 156.9258575439453, -331.2925109863281, -58.607200622558594, 712.4578857421875, 607.242431640625, 93.60263061523438, -10.966390609741211, 90.57887268066406, -228.17727661132812, 392.3367004394531, 424.6418151855469, 30.93733787536621, 136.20303344726562, 416.37640380859375, 259.294921875, -13.363853454589844, 1520.1788330078125, 94.8482666015625, 488.1858825683594, -1201.8525390625, 152.48440551757812, 1207.8546142578125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000511.npy"}
{"epoch": 0.750367107195301, "step": 512, "batch_size": 64, "mean": 391.6944885253906, "std": 397.4191589355469, "min": -418.7760009765625, "p10": -48.14629592895506, "median": 368.021484375, "p90": 946.1700012207034, "max": 1622.933837890625, "pos_frac": 0.84375, "sample": [991.070068359375, 984.037109375, 751.0330810546875, -292.52740478515625, -33.88706970214844, -30.298248291015625, 1622.933837890625, 317.69122314453125, 67.0015869140625, -54.25739288330078, 92.25410461425781, 616.3602294921875, 4.18614387512207, 397.294921875, -239.43972778320312, 74.55424499511719, 488.32952880859375, 471.9377746582031, 304.4058837890625, -11.003290176391602, -177.1260986328125, 689.5576782226562, 81.25198364257812, 485.78521728515625, 1026.9884033203125, 677.9077758789062, 219.40252685546875, 354.05755615234375, 36.00304412841797, 731.6348876953125, 301.3514099121094, 309.27496337890625, 510.22406005859375, 489.97235107421875, 227.3004913330078, 147.3363800048828, 1032.481201171875, 381.98541259765625, -84.9081802368164, 824.426025390625, 130.81210327148438, 975.6100463867188, 410.5180969238281, 860.2275390625, 34.055938720703125, -418.7760009765625, 420.55810546875, 27.37326431274414, 233.05809020996094, 508.31671142578125, 980.1832275390625, 40.64727020263672, 422.01336669921875, 642.3817749023438, 217.94137573242188, 559.1041259765625, 696.5374145507812, 209.39376831054688, 303.3907470703125, 877.4765625, -219.3290252685547, 838.438232421875, 849.3098754882812, 682.6215209960938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000512.npy"}
{"epoch": 0.7518355359765051, "step": 513, "batch_size": 64, "mean": 276.1199645996094, "std": 431.35321044921875, "min": -619.0232543945312, "p10": -188.0719528198242, "median": 232.11022186279297, "p90": 915.4990600585937, "max": 1336.9537353515625, "pos_frac": 0.734375, "sample": [667.8865966796875, 283.883056640625, 276.08233642578125, 105.79467010498047, 269.05609130859375, 207.2098846435547, 324.599365234375, 186.6258087158203, -250.21202087402344, 1010.08251953125, 572.3776245117188, -16.23541259765625, 545.0419921875, 60.47466278076172, 42.055259704589844, 708.7614135742188, 33.68878173828125, -42.078125, 196.2002410888672, 271.64996337890625, 619.2737426757812, 916.9647827148438, 1104.8919677734375, 253.0789794921875, 520.3193969726562, 109.27413940429688, 232.65635681152344, 180.14569091796875, 353.69915771484375, -161.46343994140625, 134.46151733398438, 1160.80859375, 514.5048828125, 495.10211181640625, 334.13751220703125, -56.172943115234375, -45.79325866699219, -619.0232543945312, 385.14239501953125, 1336.9537353515625, -136.37464904785156, 128.68228149414062, 131.46234130859375, -596.9263916015625, 548.5712280273438, -100.22479248046875, -412.73681640625, -19.728219985961914, 973.2772827148438, 740.19677734375, -536.179931640625, 0.3799247741699219, 903.993896484375, 1080.72802734375, 231.5640869140625, 237.17254638671875, -199.47560119628906, 602.7637329101562, -67.17313385009766, -80.40397644042969, -350.4394226074219, 912.0790405273438, 182.60272216796875, 275.9598388671875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000513.npy"}
{"epoch": 0.7533039647577092, "step": 514, "batch_size": 64, "mean": 273.0328369140625, "std": 447.1219177246094, "min": -573.353759765625, "p10": -211.4595199584961, "median": 199.61526489257812, "p90": 872.3574523925784, "max": 1721.2196044921875, "pos_frac": 0.6875, "sample": [-347.4645080566406, -152.970703125, 109.27442932128906, -64.94232177734375, 1007.1639404296875, 139.52255249023438, 591.748291015625, 618.5265502929688, 1721.2196044921875, 900.6574096679688, 167.89759826660156, 582.6979370117188, -217.67138671875, -112.23065185546875, 377.90496826171875, -86.52043151855469, -264.3424377441406, 132.28652954101562, 1090.6016845703125, 1078.407470703125, -193.33802795410156, -56.85700225830078, 296.6160583496094, -98.70465850830078, 37.279884338378906, 426.5124206542969, 294.677734375, 130.1190185546875, 372.8619384765625, -77.25727081298828, 796.974853515625, 446.56915283203125, -120.84293365478516, -54.6806640625, 144.9249267578125, 315.55242919921875, 1037.9886474609375, -212.37249755859375, 58.69957733154297, -166.69415283203125, 343.65185546875, -410.1114501953125, -196.831787109375, 603.448486328125, 246.22445678710938, 84.25437927246094, -573.353759765625, -209.32923889160156, -231.11285400390625, 756.779052734375, 293.3940734863281, 495.43701171875, 349.16705322265625, 687.92626953125, 806.32421875, 156.07704162597656, 305.0994873046875, 418.5034484863281, 655.1719360351562, 231.3329315185547, 527.0253295898438, 48.50917434692383, 114.02529907226562, 1322.6943359375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000514.npy"}
{"epoch": 0.7547723935389133, "step": 515, "batch_size": 64, "mean": 230.2169647216797, "std": 397.4438781738281, "min": -1288.0675048828125, "p10": -195.77205963134764, "median": 192.73709869384766, "p90": 677.5316101074222, "max": 1257.123779296875, "pos_frac": 0.796875, "sample": [-18.621654510498047, 9.121200561523438, 503.4163513183594, 845.667236328125, 443.08624267578125, -1288.0675048828125, 410.390625, 468.3472900390625, 154.12921142578125, -467.161865234375, -192.5437774658203, 30.385513305664062, 5.949371337890625, 182.1937255859375, 280.1207580566406, 371.7796325683594, 135.4795379638672, 575.6905517578125, -262.71221923828125, 154.63055419921875, 396.3111877441406, -73.29270935058594, 108.37281799316406, 386.81829833984375, 190.17210388183594, 118.16966247558594, 552.0699462890625, -121.38619995117188, 439.241943359375, 46.65925598144531, 717.6345825195312, 31.231103897094727, -197.15560913085938, 565.443603515625, 306.17901611328125, 357.21990966796875, 537.05908203125, 13.287834167480469, 1181.446533203125, -263.7441711425781, 160.53114318847656, 886.4617309570312, -352.66168212890625, -36.120933532714844, -313.9727783203125, 208.83145141601562, 109.8021469116211, 752.4713745117188, 38.51494216918945, 110.4664306640625, 67.32295989990234, 1000.934814453125, 213.68704223632812, 151.736328125, 370.9076232910156, 583.9580078125, 465.6986389160156, -134.63143920898438, 341.2955322265625, 490.1021728515625, 245.74984741210938, 195.30209350585938, 1257.123779296875, 287.35516357421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000515.npy"}
{"epoch": 0.7562408223201175, "step": 516, "batch_size": 64, "mean": 344.97198486328125, "std": 408.1154479980469, "min": -562.3087768554688, "p10": -185.68157806396482, "median": 352.1564025878906, "p90": 844.4562744140626, "max": 1223.916748046875, "pos_frac": 0.8125, "sample": [185.54840087890625, 177.63796997070312, 678.9412841796875, -147.8072052001953, 557.8052368164062, 788.5685424804688, 365.1051940917969, -129.0337371826172, 267.9686279296875, 539.0552368164062, 1042.342041015625, 455.69403076171875, 542.7706909179688, 450.70416259765625, 291.78076171875, -88.96353149414062, 565.3742065429688, 180.17555236816406, 109.0775146484375, 1223.916748046875, -562.3087768554688, 88.4820556640625, -201.9134521484375, -435.74066162109375, 936.9510498046875, 535.764404296875, -420.9012756347656, 706.4078979492188, 1148.421142578125, 183.6971893310547, 341.4535827636719, 850.171142578125, 1170.720947265625, 354.48187255859375, 585.6393432617188, 549.1075439453125, 324.48388671875, 25.243953704833984, 411.58349609375, 262.92919921875, 90.93312072753906, -23.164033889770508, 766.949951171875, 399.8910217285156, 8.128910064697266, 1004.2669677734375, 188.74562072753906, 232.6866455078125, 39.635520935058594, 287.8831787109375, 435.1620788574219, 521.0667724609375, 90.57904052734375, 540.2131958007812, 687.7269897460938, -211.6214599609375, -290.74615478515625, -515.7652587890625, 590.988037109375, 583.0623779296875, 621.9725341796875, 831.12158203125, 349.8309326171875, -62.6766242980957], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000516.npy"}
{"epoch": 0.7577092511013216, "step": 517, "batch_size": 64, "mean": 381.9071960449219, "std": 507.1583557128906, "min": -640.1034545898438, "p10": -184.09300537109374, "median": 339.37525939941406, "p90": 905.8159362792971, "max": 2277.46435546875, "pos_frac": 0.796875, "sample": [478.4679870605469, 708.9967651367188, 166.66632080078125, 358.473876953125, 845.5076293945312, 231.48861694335938, -192.9006805419922, 771.3163452148438, 493.4520263671875, -269.2467956542969, 571.3341064453125, 1811.409912109375, 1154.525634765625, 443.49249267578125, 466.12158203125, -117.23455047607422, 257.79498291015625, 566.4500732421875, 161.5989227294922, 693.6763916015625, 1260.30322265625, 672.154052734375, 18.49546241760254, 161.85659790039062, -640.1034545898438, -327.9786071777344, 119.7280502319336, 405.4275817871094, -300.8122253417969, 2277.46435546875, 327.31982421875, 21.749404907226562, -134.7816925048828, 931.662353515625, 88.45184326171875, 510.3746337890625, -504.59130859375, 495.9144592285156, -73.50875854492188, 369.9071350097656, 759.148193359375, -163.54176330566406, 293.24273681640625, 331.0160827636719, 611.06494140625, 557.0826416015625, 674.5906982421875, 345.7962341308594, 67.10547637939453, 1262.6563720703125, 132.2656707763672, 1037.945068359375, 332.95428466796875, 830.3543701171875, 86.35205078125, 750.3995361328125, -297.085205078125, 247.82388305664062, 204.84132385253906, 357.715576171875, -155.580322265625, 672.8539428710938, 254.32083129882812, -31.686500549316406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000517.npy"}
{"epoch": 0.7591776798825257, "step": 518, "batch_size": 64, "mean": 327.5380859375, "std": 413.98773193359375, "min": -777.48583984375, "p10": -209.8187774658203, "median": 288.01515197753906, "p90": 863.6869689941407, "max": 1211.0286865234375, "pos_frac": 0.796875, "sample": [36.425106048583984, 1040.41455078125, 517.087158203125, 430.99383544921875, 168.76060485839844, 268.170654296875, 525.0265502929688, 567.4520263671875, 61.79945755004883, 311.2722473144531, 700.2438354492188, 696.2457885742188, 132.716796875, -187.3309783935547, 487.5221862792969, 167.7098846435547, 417.0005187988281, 159.1748046875, 25.980316162109375, 542.9773559570312, 1042.719970703125, -17.497718811035156, -231.05386352539062, 157.32931518554688, -19.689395904541016, 303.7445373535156, 695.84814453125, 101.778564453125, 659.187744140625, 788.4927368164062, 863.0354614257812, 103.85813903808594, 1211.0286865234375, 1112.9097900390625, 818.5694580078125, 688.685546875, 104.25939178466797, -110.57781982421875, 472.4136657714844, 633.0388793945312, 270.76239013671875, -70.27166748046875, 512.7903442382812, -249.78921508789062, 488.40496826171875, -336.62420654296875, -380.9490966796875, -777.48583984375, 360.9156188964844, 507.4437561035156, 30.796798706054688, 157.98179626464844, -219.45640563964844, 168.29995727539062, -377.05963134765625, 1173.516845703125, 863.9661865234375, 279.8289489746094, -71.75936889648438, 884.4838256835938, 294.5666809082031, 143.08058166503906, 579.806396484375, 281.463623046875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000518.npy"}
{"epoch": 0.7606461086637298, "step": 519, "batch_size": 64, "mean": 232.760986328125, "std": 456.33721923828125, "min": -940.3775634765625, "p10": -327.738247680664, "median": 230.16307067871094, "p90": 820.8544128417974, "max": 1293.2979736328125, "pos_frac": 0.734375, "sample": [2.437225341796875, 554.1441040039062, 302.8778381347656, -464.3996887207031, 1063.905517578125, -281.0782470703125, 78.28630065917969, 503.47271728515625, 1070.2845458984375, 535.4060668945312, 164.7821502685547, 575.9898681640625, -199.92852783203125, 943.3655395507812, 354.8627624511719, -940.3775634765625, 1270.065673828125, 614.343505859375, 66.04967498779297, -38.91535949707031, 293.5921630859375, -68.02955627441406, 49.669227600097656, 351.35272216796875, 393.95318603515625, 170.11416625976562, 266.6146545410156, 584.1300048828125, 473.7383728027344, 29.75856590270996, -33.2882080078125, 599.2571411132812, -162.59274291992188, -141.9573974609375, 353.75299072265625, 61.321449279785156, 569.263916015625, -92.6209487915039, -347.23077392578125, -24.551040649414062, 355.45880126953125, -341.2551574707031, 954.3609008789062, 461.06927490234375, -889.7001953125, 636.4727172851562, 1293.2979736328125, 24.97248077392578, 3.183349609375, 21.745529174804688, 166.1177978515625, 873.3297119140625, -337.3368835449219, 654.9561157226562, -305.3414306640625, 195.0577392578125, 698.4120483398438, 19.40957260131836, 370.7250061035156, 449.6605529785156, 317.54132080078125, 11.771778106689453, 265.2684020996094, -504.29608154296875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000519.npy"}
{"epoch": 0.762114537444934, "step": 520, "batch_size": 64, "mean": 295.96063232421875, "std": 456.0131530761719, "min": -837.7501220703125, "p10": -181.75666656494138, "median": 216.55307006835938, "p90": 914.0435668945316, "max": 1907.954833984375, "pos_frac": 0.71875, "sample": [-125.6692886352539, -150.1036376953125, 1096.09228515625, 134.13023376464844, 475.6786804199219, 135.05221557617188, 363.4451599121094, -68.05426025390625, -172.15338134765625, 166.1832275390625, -10.704208374023438, -56.685333251953125, 660.8865966796875, -7.4132843017578125, -70.8031997680664, 647.5028076171875, 1029.79248046875, 1907.954833984375, -258.3777770996094, 337.19549560546875, 1039.164794921875, 1265.62841796875, -371.3615417480469, 210.047607421875, -161.26828002929688, 762.965576171875, 125.85797882080078, 583.3372192382812, 330.0543212890625, 300.1771240234375, 278.832275390625, 271.0226135253906, 1175.473876953125, 148.73841857910156, 55.791534423828125, 410.4318542480469, 527.3109130859375, 330.8387145996094, 780.9523315429688, -837.7501220703125, 541.1815795898438, 172.65878295898438, 200.40187072753906, 354.82427978515625, -80.67681884765625, 795.6336059570312, 93.34496307373047, 814.0763549804688, 464.0999450683594, 64.84552001953125, 204.3854217529297, -221.61395263671875, -217.37350463867188, 290.85400390625, 287.2893981933594, 657.5660400390625, -132.3836669921875, 956.8866577148438, -185.8723602294922, 415.0641174316406, 223.05853271484375, -216.4823455810547, 43.205780029296875, 156.3106689453125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000520.npy"}
{"epoch": 0.7635829662261381, "step": 521, "batch_size": 64, "mean": 325.7584228515625, "std": 439.5833435058594, "min": -633.2882080078125, "p10": -110.3952033996582, "median": 276.8942565917969, "p90": 937.0911621093752, "max": 1549.609130859375, "pos_frac": 0.75, "sample": [482.5303955078125, 787.8856201171875, 274.5133361816406, 27.23590850830078, -47.741859436035156, -633.2882080078125, -162.34921264648438, -19.92901611328125, -109.84371948242188, 463.00115966796875, 32.984283447265625, -243.39801025390625, 36.1211051940918, 423.9288024902344, 452.3287048339844, 1168.162353515625, 122.99908447265625, 37.87045669555664, -378.33935546875, 884.448974609375, 287.93377685546875, -17.133895874023438, -150.7932586669922, 165.6193084716797, 472.37579345703125, 583.7503662109375, 612.4150390625, 160.56460571289062, 41.840492248535156, 8.4256591796875, 353.6400146484375, 574.728515625, -40.90711212158203, 1054.3800048828125, 241.1396484375, 598.2138671875, 595.5126953125, 182.10934448242188, 673.087890625, -110.63155364990234, 279.2751770019531, 362.74261474609375, 284.162841796875, 17.400562286376953, 225.99200439453125, 983.1553344726562, -103.29148864746094, 983.3495483398438, -0.70037841796875, -613.3588256835938, 588.158447265625, 419.79278564453125, 726.3944702148438, -15.762283325195312, 640.6441650390625, 959.652099609375, 164.2270965576172, 1368.5609130859375, -26.61206817626953, 1549.609130859375, 485.0249938964844, 78.07389831542969, 850.8405151367188, 755.813232421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000521.npy"}
{"epoch": 0.7650513950073421, "step": 522, "batch_size": 64, "mean": 303.8099670410156, "std": 479.4903869628906, "min": -892.7406616210938, "p10": -191.26564636230466, "median": 229.30557250976562, "p90": 1060.8266357421876, "max": 1297.0328369140625, "pos_frac": 0.765625, "sample": [278.3634033203125, 734.1441040039062, 898.478271484375, -40.410179138183594, 1161.35498046875, -6.350284576416016, 1297.0328369140625, 208.32778930664062, 102.62895202636719, -27.695240020751953, -180.39028930664062, 285.8960266113281, 739.815673828125, 42.73185729980469, 491.9301452636719, -39.629364013671875, 265.0268859863281, -81.54341125488281, 478.34930419921875, 232.43878173828125, 639.206787109375, 595.9171142578125, 139.06756591796875, 389.0093994140625, -623.6409912109375, 1117.891357421875, 280.9940185546875, 175.86422729492188, 581.5372924804688, 600.7179565429688, 173.56524658203125, 277.06549072265625, 103.98939514160156, 19.69903564453125, 321.72454833984375, -892.7406616210938, 539.6500854492188, 143.87301635742188, 226.17236328125, -325.85284423828125, -195.926513671875, 478.35546875, 1065.4296875, 111.00293731689453, -471.45751953125, 503.259765625, 753.35791015625, 371.95904541015625, -758.2786865234375, 1159.2003173828125, 1050.086181640625, -56.128265380859375, 1219.7864990234375, 94.77753448486328, 145.13296508789062, -425.1685791015625, 160.02078247070312, -146.68185424804688, 622.6138305664062, 203.71121215820312, 919.7490844726562, 82.99603271484375, 1148.2447509765625, 83.58280944824219], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000522.npy"}
{"epoch": 0.7665198237885462, "step": 523, "batch_size": 64, "mean": 333.8455810546875, "std": 538.9910278320312, "min": -972.1533203125, "p10": -214.65851440429682, "median": 329.85765075683594, "p90": 991.3598449707032, "max": 1663.6053466796875, "pos_frac": 0.71875, "sample": [-168.92820739746094, 1112.447509765625, 81.18281555175781, 237.3101043701172, 401.4573974609375, -234.25721740722656, 524.9866943359375, 519.44140625, 859.9017333984375, 370.31878662109375, -738.61767578125, 652.4929809570312, -4.119474411010742, -42.366241455078125, -427.0875549316406, 240.67466735839844, 215.53564453125, 833.20849609375, -474.5751647949219, 233.85910034179688, 500.0233154296875, 967.16015625, -100.0516357421875, 27.31901741027832, 446.301025390625, 116.18464660644531, 309.216796875, 70.92404174804688, -71.0000991821289, 1487.6368408203125, 554.9462280273438, 123.45683288574219, -94.62136840820312, -972.1533203125, 1001.7311401367188, 834.006591796875, -477.6163024902344, -166.5609893798828, 502.8328857421875, -105.26017761230469, 536.745849609375, -81.07577514648438, 1663.6053466796875, 390.5576477050781, 739.0320434570312, 81.19729614257812, 350.4985046386719, 365.3585205078125, 1546.2171630859375, 518.2238159179688, 664.04833984375, 109.87234497070312, 578.7059936523438, 46.38072967529297, 856.1971435546875, 1194.9993896484375, 791.7725830078125, -85.03990173339844, 587.61181640625, 1394.756103515625, 577.2330322265625, -634.428955078125, 66.76903533935547, -40.462947845458984], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000523.npy"}
{"epoch": 0.7679882525697503, "step": 524, "batch_size": 64, "mean": 322.07330322265625, "std": 504.83343505859375, "min": -1499.254638671875, "p10": -108.21096954345703, "median": 382.3222351074219, "p90": 856.3505676269531, "max": 1593.6534423828125, "pos_frac": 0.734375, "sample": [44.20622253417969, 854.4347534179688, 447.96484375, 499.1077880859375, 415.5269470214844, 209.3330078125, 381.29669189453125, 57.88562774658203, 541.2359008789062, 373.7004699707031, 430.31622314453125, 86.29708862304688, 519.906494140625, 493.54400634765625, 642.8944091796875, 365.1132507324219, -12.400497436523438, 806.6346435546875, 222.12123107910156, 606.9646606445312, -1499.254638671875, 532.38037109375, 353.99188232421875, 1593.6534423828125, -3.0572509765625, -27.23101806640625, 885.6055908203125, 370.06512451171875, 475.952392578125, -56.602542877197266, -39.08487319946289, 388.9168701171875, 428.9973449707031, -219.64683532714844, 800.5782470703125, 277.2502136230469, 1146.3427734375, 1420.320556640625, 642.12158203125, 304.08123779296875, -75.20723724365234, -939.431396484375, 10.690277099609375, -106.39904022216797, 745.3246459960938, -88.8270263671875, 441.46173095703125, 6.954719543457031, 420.9000244140625, 466.4860534667969, 1021.3237915039062, 1126.4342041015625, 615.9979858398438, -935.95361328125, -108.98751068115234, -65.26316833496094, 857.171630859375, 611.1856689453125, -122.44454193115234, 336.3631286621094, 537.2296142578125, 383.3477783203125, -220.45599365234375, -66.67352294921875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000524.npy"}
{"epoch": 0.7694566813509545, "step": 525, "batch_size": 64, "mean": 442.56463623046875, "std": 546.018798828125, "min": -669.1058959960938, "p10": -95.60855712890624, "median": 312.81739807128906, "p90": 1291.0304565429688, "max": 1786.6893310546875, "pos_frac": 0.8125, "sample": [360.4716796875, 1144.6630859375, 400.25714111328125, 131.96542358398438, 111.70098114013672, 467.1722717285156, 1786.6893310546875, 762.2671508789062, 378.5796203613281, -197.10940551757812, 1300.5458984375, -434.10638427734375, 942.7766723632812, 209.0412139892578, 243.11370849609375, -230.91819763183594, 136.11436462402344, 291.8479919433594, 1027.0367431640625, 710.7967529296875, 635.9075317382812, 134.89956665039062, -67.5084228515625, 108.18936157226562, -125.38653564453125, 365.8241882324219, 584.3423461914062, 546.0562744140625, 328.9822082519531, 1069.49755859375, 1268.8277587890625, 59.118125915527344, 1645.8267822265625, -4.786476135253906, 1672.0208740234375, 309.47430419921875, 119.63756561279297, 190.36013793945312, 399.7523193359375, -97.02731323242188, 1572.29248046875, -669.1058959960938, -92.15435791015625, -8.205408096313477, 266.2840881347656, 187.67733764648438, 553.305419921875, 125.08749389648438, 588.75390625, 169.00888061523438, 794.3111572265625, 1471.164794921875, -92.29812622070312, 67.26934814453125, 729.0057983398438, 290.7051696777344, 316.1604919433594, 67.2306137084961, -467.23907470703125, 908.6729736328125, 200.88156127929688, 1443.50732421875, 339.4336242675781, 875.4739990234375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000525.npy"}
{"epoch": 0.7709251101321586, "step": 526, "batch_size": 64, "mean": 417.6783142089844, "std": 547.37646484375, "min": -909.8963623046875, "p10": -211.46868591308592, "median": 402.7691650390625, "p90": 1151.8164916992187, "max": 1482.2901611328125, "pos_frac": 0.734375, "sample": [1153.59814453125, 859.0126342773438, -4.314189910888672, 281.8726806640625, 301.586669921875, 309.7204895019531, -205.67945861816406, -4.128911972045898, 366.34503173828125, -782.0656127929688, 530.4838256835938, 791.5631713867188, 742.5400390625, 247.90896606445312, 1262.031005859375, 986.1053466796875, -174.08926391601562, 729.1727294921875, 281.2440185546875, 523.7201538085938, 348.9690246582031, 603.300537109375, 157.50929260253906, 1066.408447265625, 541.3048095703125, 552.6181030273438, 1482.2901611328125, 59.31654357910156, 358.78302001953125, -220.4915008544922, -213.9497833251953, 581.0371704101562, -439.8002624511719, 1312.8319091796875, -181.66909790039062, 326.9031677246094, 1147.6593017578125, 896.8413696289062, 263.6819152832031, -909.8963623046875, 894.7400512695312, 589.3477783203125, 582.077880859375, 1179.743408203125, -424.48480224609375, 696.7183837890625, 721.8865966796875, 1057.96240234375, -56.491661071777344, 510.173095703125, -55.546844482421875, 256.382568359375, 225.87008666992188, 1122.087890625, 1001.2186889648438, -82.08279418945312, 1253.732177734375, 661.9064331054688, 439.19329833984375, -3.5357513427734375, -123.70934295654297, 26.56479835510254, -842.416748046875, 1169.8011474609375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000526.npy"}
{"epoch": 0.7723935389133627, "step": 527, "batch_size": 64, "mean": 280.2981262207031, "std": 421.3102111816406, "min": -631.3324584960938, "p10": -212.27175903320312, "median": 244.13501739501953, "p90": 781.9664184570313, "max": 1281.5142822265625, "pos_frac": 0.765625, "sample": [-560.8283081054688, 568.20166015625, 394.4524230957031, 306.6773986816406, 686.4937133789062, 147.51272583007812, 1281.5142822265625, -110.32438659667969, -140.31240844726562, 757.41455078125, -169.31394958496094, -217.83248901367188, 620.5794067382812, 281.8846740722656, -631.3324584960938, -549.4652099609375, 548.0053100585938, -131.51153564453125, 562.4030151367188, 368.71978759765625, -379.1185302734375, 932.385009765625, 78.5438232421875, 127.607666015625, 135.81468200683594, 741.4664306640625, -39.19935607910156, 108.89295196533203, 792.4886474609375, 673.6372680664062, 574.1181640625, 1032.3406982421875, 387.5283203125, 643.8104248046875, 1064.7259521484375, 1105.224853515625, 797.258056640625, -37.104408264160156, 86.80058288574219, 244.19601440429688, 630.2418823242188, 198.54730224609375, 264.90289306640625, 262.708984375, -130.04425048828125, 83.82034301757812, 241.43109130859375, 311.3698425292969, 147.17575073242188, -549.5028076171875, 244.0740203857422, 156.87600708007812, 533.5940551757812, 224.49285888671875, 75.38638305664062, 75.87565612792969, 686.4814453125, 633.8906860351562, 516.9375610351562, 454.12384033203125, 223.04278564453125, -199.29672241210938, 56.19847869873047, -287.60467529296875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000527.npy"}
{"epoch": 0.7738619676945668, "step": 528, "batch_size": 64, "mean": 343.55035400390625, "std": 568.9129638671875, "min": -609.8961181640625, "p10": -412.3058532714843, "median": 285.40643310546875, "p90": 1165.3241577148438, "max": 1835.983642578125, "pos_frac": 0.734375, "sample": [-94.4715347290039, 1171.905517578125, 327.5580139160156, 367.1804504394531, 229.82537841796875, -451.64776611328125, 31.86962127685547, 688.1776123046875, 612.94580078125, 1224.0025634765625, 433.1397705078125, -103.39714050292969, 354.2137451171875, -516.790771484375, 154.47206115722656, 463.42718505859375, 304.2728271484375, -122.47494506835938, 1149.9676513671875, 1835.983642578125, -55.808815002441406, 980.079833984375, 423.37017822265625, 675.2156372070312, 579.1375122070312, 1059.279052734375, -606.844482421875, 1375.4664306640625, -291.0608215332031, 1036.754150390625, -591.4457397460938, 266.5400390625, 329.01116943359375, 213.53970336914062, 332.212890625, 189.51014709472656, 469.6415710449219, 101.62745666503906, 411.0622863769531, 32.048744201660156, -211.52513122558594, 881.6383666992188, -163.09164428710938, -607.998291015625, 1303.5513916015625, 230.42770385742188, 563.021484375, 1651.752197265625, 120.55635070800781, 37.1387939453125, -531.307861328125, -609.8961181640625, 715.94677734375, 1194.7108154296875, 76.412353515625, 802.0210571289062, 669.15625, 177.1162109375, -320.508056640625, -14.313949584960938, 104.98709106445312, -130.54273986816406, 956.3002319335938, 102.17432403564453], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000528.npy"}
{"epoch": 0.775330396475771, "step": 529, "batch_size": 64, "mean": 237.2031707763672, "std": 555.4955444335938, "min": -1535.998291015625, "p10": -450.58307189941405, "median": 251.63294982910156, "p90": 859.9626647949221, "max": 1543.0748291015625, "pos_frac": 0.6875, "sample": [160.25624084472656, -161.78550720214844, 1543.0748291015625, 57.76824951171875, 134.76559448242188, 593.8571166992188, 405.2892150878906, 640.0126342773438, 733.9376220703125, 1369.979736328125, -354.4507751464844, 460.024658203125, 111.08699798583984, -334.75897216796875, 550.3714599609375, -124.38697052001953, 135.66949462890625, 119.4292221069336, 704.2481689453125, -4.602027893066406, -363.5023498535156, 271.3995361328125, 1092.02294921875, 785.4868774414062, 929.5613403320312, -656.7942504882812, -144.34393310546875, 377.3191223144531, -534.9369506835938, 684.36865234375, 231.86636352539062, 504.4248352050781, -93.6218032836914, 367.86151123046875, -23.300140380859375, 16.595014572143555, 708.343505859375, 85.04246520996094, 609.7404174804688, 712.7599487304688, -1535.998291015625, 172.21624755859375, 1001.0972900390625, -491.05908203125, 557.0113525390625, -990.0421752929688, 111.9205551147461, 566.4421997070312, 886.4926147460938, 145.95938110351562, -743.1019287109375, 463.37451171875, -60.41718292236328, 454.2807312011719, -440.55084228515625, 432.3797302246094, 393.8945617675781, 798.0594482421875, 606.26904296875, 1100.421630859375, -454.8825988769531, -112.99174499511719, -275.0048522949219, 295.1518249511719], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000529.npy"}
{"epoch": 0.7767988252569751, "step": 530, "batch_size": 64, "mean": 375.37451171875, "std": 677.3529052734375, "min": -2258.488037109375, "p10": -218.92864685058584, "median": 251.03358459472656, "p90": 1410.1610473632813, "max": 2161.0166015625, "pos_frac": 0.703125, "sample": [352.1918029785156, -493.8162536621094, -18.33727264404297, 144.3031005859375, 303.8215026855469, -122.36788177490234, 1641.63037109375, 614.8389282226562, 278.2685241699219, -517.887451171875, 153.76136779785156, -102.31243133544922, -1.56256103515625, -86.84646606445312, 680.6830444335938, 769.98974609375, -8.310590744018555, -52.031005859375, -342.172119140625, 1367.5877685546875, 1746.4600830078125, 449.6697082519531, 940.9849243164062, 739.5203247070312, -67.14979553222656, 739.6721801757812, -85.048095703125, 153.16513061523438, -260.31182861328125, 680.8944091796875, -8.684713363647461, 1591.7015380859375, 1094.2410888671875, 469.7176513671875, 385.7304382324219, 58.66840744018555, 770.17626953125, 200.79232788085938, -262.53875732421875, -421.2342529296875, 120.37986755371094, 1641.4544677734375, 597.1531372070312, 2161.0166015625, 27.261720657348633, 451.8940124511719, 390.8119812011719, 223.79864501953125, 1428.40673828125, 193.79147338867188, 307.597900390625, 18.27484703063965, 850.0203247070312, -2258.488037109375, 1267.6600341796875, -2.0886077880859375, -35.233978271484375, 1482.9986572265625, 496.0099182128906, 331.4107666015625, 82.62139892578125, 146.02227783203125, 500.70343017578125, 122.6318130493164], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000530.npy"}
{"epoch": 0.7782672540381792, "step": 531, "batch_size": 64, "mean": 317.74737548828125, "std": 534.1055908203125, "min": -1275.8375244140625, "p10": -223.07264862060543, "median": 245.57820892333984, "p90": 990.5006896972658, "max": 1742.7745361328125, "pos_frac": 0.765625, "sample": [626.38623046875, 138.55531311035156, 642.5718994140625, 1010.367431640625, 643.5736083984375, -409.57012939453125, 307.298828125, 220.66873168945312, 1634.9217529296875, -54.955108642578125, -235.7285919189453, 403.0937805175781, 5.827215194702148, 286.7165832519531, 281.9593505859375, 206.02992248535156, 419.99444580078125, 200.55848693847656, -1275.8375244140625, 339.2032165527344, 944.1449584960938, 633.1430053710938, 271.13031005859375, 82.40922546386719, 854.012451171875, 172.9027099609375, 468.85479736328125, 246.7498321533203, 377.70611572265625, -47.323936462402344, 1235.6510009765625, -283.58807373046875, 499.52923583984375, 567.0213623046875, 98.68006896972656, 213.95620727539062, -32.87230682373047, -391.1666564941406, -16.22595977783203, 182.9085693359375, 1742.7745361328125, 408.969970703125, 125.87264251708984, -87.81742858886719, 1682.8570556640625, -503.3713073730469, 452.3327331542969, 1330.6348876953125, 225.096435546875, 849.7340087890625, 1078.2381591796875, 57.310630798339844, 573.6278076171875, -173.75680541992188, 244.40658569335938, 129.5313262939453, 471.8284912109375, 137.54615783691406, 387.4543762207031, -22.465866088867188, 220.064697265625, 548.0187377929688, -818.77294921875, -193.5421142578125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000531.npy"}
{"epoch": 0.7797356828193832, "step": 532, "batch_size": 64, "mean": 366.9369812011719, "std": 631.0108642578125, "min": -1315.1934814453125, "p10": -348.7727935791015, "median": 313.83653259277344, "p90": 1224.7818969726566, "max": 1889.0498046875, "pos_frac": 0.734375, "sample": [1135.8917236328125, 442.150146484375, 1426.3975830078125, -224.67703247070312, 1262.877685546875, 236.43582153320312, 992.9547119140625, 765.3232421875, -1315.1934814453125, 327.7322692871094, -332.3255920410156, 1346.12451171875, 523.7166137695312, 746.04541015625, -291.5677490234375, 675.2567138671875, 565.7606201171875, -951.8858642578125, 326.5694580078125, 301.1036071777344, 391.6336975097656, 1675.660888671875, 1889.0498046875, 875.775634765625, 608.900634765625, 544.605712890625, -47.91694641113281, 1048.03662109375, 1285.5408935546875, 210.3038330078125, -116.86579132080078, 496.3428039550781, -159.23379516601562, 1020.64306640625, 891.1954956054688, 225.28451538085938, 244.36038208007812, 887.200927734375, -38.446449279785156, 240.21511840820312, 65.99364471435547, 759.2933349609375, 173.93389892578125, 588.8994140625, 218.68312072753906, -104.40250396728516, 68.93128204345703, 722.1388549804688, 177.80831909179688, -5.110603332519531, -501.7691345214844, -697.8388671875, 328.6507568359375, 1588.2152099609375, 265.70794677734375, -27.247314453125, 282.7864685058594, -901.9503173828125, 152.201904296875, 238.96507263183594, -355.82159423828125, 579.0665283203125, -594.0106811523438, 329.8650817871094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000532.npy"}
{"epoch": 0.7812041116005873, "step": 533, "batch_size": 64, "mean": 263.63397216796875, "std": 538.387939453125, "min": -1379.244140625, "p10": -293.3149230957031, "median": 220.86317443847656, "p90": 976.2145507812504, "max": 1987.7462158203125, "pos_frac": 0.6875, "sample": [411.4432067871094, -389.8458251953125, 831.559814453125, -405.3852233886719, 358.89508056640625, 198.9034423828125, 1167.4742431640625, 536.532470703125, 406.05657958984375, 311.51861572265625, 616.1092529296875, 627.7156372070312, 595.2947387695312, 683.6849975585938, -198.2579345703125, 447.5426330566406, 1017.1564331054688, 204.16867065429688, 242.98562622070312, -806.80078125, 1247.6318359375, 274.8184814453125, 56.3740234375, 584.404052734375, 58.53834915161133, 523.6254272460938, -26.49259376525879, 200.19261169433594, 190.7645263671875, 1987.7462158203125, 340.01348876953125, 367.1408386230469, -73.75604248046875, 27.509315490722656, -17.30466079711914, -1379.244140625, -204.15707397460938, 344.0850524902344, -238.6246337890625, 504.9573669433594, 161.8115692138672, 181.98402404785156, 1311.124755859375, 1015.2064208984375, -160.29220581054688, -32.50294494628906, -313.8746643066406, -4.4744110107421875, 292.7856140136719, 231.34896850585938, 343.6742858886719, -55.648345947265625, 885.2335205078125, 63.55480194091797, -306.46136474609375, -262.639892578125, -784.9254760742188, 158.05133056640625, 210.37738037109375, 1178.8863525390625, -165.56964111328125, -67.60690307617188, 779.5127563476562, 588.0425415039062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000533.npy"}
{"epoch": 0.7826725403817915, "step": 534, "batch_size": 64, "mean": 335.187744140625, "std": 510.19854736328125, "min": -959.309326171875, "p10": -243.91093749999993, "median": 270.86802673339844, "p90": 923.0258361816407, "max": 1765.6533203125, "pos_frac": 0.78125, "sample": [446.485595703125, 911.4863891601562, -41.439483642578125, 1765.6533203125, 39.18715286254883, 677.7537231445312, -512.3236083984375, 321.8661804199219, 262.40924072265625, -425.7444763183594, 69.16380310058594, 1107.67333984375, -148.84664916992188, 160.92739868164062, 995.5513305664062, 878.651123046875, 675.5416870117188, -269.9452209472656, 331.6646728515625, 837.1292724609375, 263.462646484375, -183.16427612304688, 407.2181701660156, 451.485595703125, 295.31683349609375, 382.01898193359375, 981.9781494140625, -142.11099243164062, 381.1129150390625, 190.56625366210938, 157.30545043945312, 832.852294921875, 246.95602416992188, 143.51864624023438, 648.3958129882812, -509.0599060058594, -88.44033813476562, 1575.6290283203125, 770.753662109375, 927.9713134765625, 158.892578125, 194.73143005371094, -92.58026123046875, 148.0958251953125, 756.3607177734375, 176.98399353027344, 176.1321563720703, 850.780029296875, 1453.1285400390625, 837.8085327148438, 697.9791870117188, 329.6821594238281, 519.332763671875, 49.28081130981445, 379.3629455566406, 189.5516357421875, 315.7257995605469, 278.2734069824219, 215.47439575195312, 31.366579055786133, -143.34756469726562, -959.309326171875, -333.92047119140625, -594.3820190429688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000534.npy"}
{"epoch": 0.7841409691629956, "step": 535, "batch_size": 64, "mean": 389.89483642578125, "std": 518.260009765625, "min": -969.57763671875, "p10": -206.7402252197265, "median": 386.5746765136719, "p90": 1016.0551757812501, "max": 2187.069580078125, "pos_frac": 0.828125, "sample": [-275.13751220703125, 2187.069580078125, -246.63458251953125, 475.99560546875, -232.0889434814453, 254.64642333984375, 68.25238037109375, 804.2550048828125, 112.93359375, 492.87445068359375, 256.8456726074219, 58.566017150878906, 168.49659729003906, 977.1663818359375, 414.6755065917969, 425.2430419921875, 49.94232940673828, 1217.532958984375, 265.8456115722656, 350.7254638671875, 926.6976318359375, 543.1366577148438, 566.1329345703125, 459.981689453125, 5.0716705322265625, 1405.8011474609375, 1032.7218017578125, 548.479248046875, -20.683561325073242, 317.3487854003906, 1439.6258544921875, 683.3688354492188, 954.974853515625, -437.6688537597656, -317.5873718261719, -969.57763671875, 478.2534484863281, 657.2616577148438, 176.10244750976562, 400.36895751953125, 395.8546142578125, 1082.3111572265625, 758.7052001953125, 204.32296752929688, 576.8555908203125, 48.25202941894531, 40.65349578857422, 776.041259765625, -147.5932159423828, -473.24609375, 606.907958984375, 80.98662567138672, 489.4791259765625, 85.6786880493164, -134.5150604248047, -121.72705078125, 814.0880126953125, 159.49868774414062, 1322.7216796875, 228.88302612304688, 30.68548011779785, 515.04248046875, 377.29473876953125, 559.0718994140625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000535.npy"}
{"epoch": 0.7856093979441997, "step": 536, "batch_size": 64, "mean": 468.435791015625, "std": 648.1734008789062, "min": -1122.5345458984375, "p10": -43.66725425720213, "median": 282.41058349609375, "p90": 1104.2096801757814, "max": 3790.262451171875, "pos_frac": 0.859375, "sample": [125.57344818115234, 726.020263671875, 310.65643310546875, 894.3118286132812, 1271.173828125, 30.664260864257812, 193.70172119140625, 134.41937255859375, 74.04238891601562, -30.389938354492188, 849.31591796875, 309.6325988769531, 1001.1663208007812, 828.337646484375, -188.11146545410156, 502.59039306640625, 1114.41748046875, 109.95787048339844, 761.1987915039062, 790.0988159179688, 2006.083740234375, 588.2212524414062, -160.35781860351562, 788.2542114257812, -3.6479873657226562, 1080.3914794921875, 25.741512298583984, 649.1841430664062, 684.3368530273438, 5.98486328125, 3790.262451171875, 786.7515258789062, 296.44097900390625, 162.01925659179688, 552.9191284179688, 626.2457275390625, 80.47587585449219, 1128.5653076171875, -261.1238098144531, 80.11447143554688, 151.79637145996094, -1122.5345458984375, 1404.797607421875, 900.862060546875, 713.3182983398438, -117.45896911621094, 81.39773559570312, 234.3082275390625, 268.38018798828125, 137.5684814453125, 192.95245361328125, 41.50780487060547, 207.55435180664062, 246.12521362304688, 242.93994140625, 697.1947631835938, 91.45802307128906, -49.3575325012207, -298.63824462890625, 56.804290771484375, 839.6319580078125, 674.6156616210938, 482.036376953125, 1186.989501953125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000536.npy"}
{"epoch": 0.7870778267254038, "step": 537, "batch_size": 64, "mean": 325.02301025390625, "std": 518.0490112304688, "min": -1067.0263671875, "p10": -296.09064178466787, "median": 280.0540008544922, "p90": 1000.3256652832031, "max": 1339.070068359375, "pos_frac": 0.796875, "sample": [-1067.0263671875, -659.858154296875, 22.36786651611328, 167.53939819335938, -342.82244873046875, 295.67205810546875, 34.17524719238281, 899.3556518554688, 7.649658203125, -46.940250396728516, 491.4712829589844, 6.8471832275390625, 477.9379577636719, 284.201416015625, -333.5882568359375, 836.0865478515625, 1010.762451171875, -10.093294143676758, 206.93035888671875, -61.41716766357422, 56.758392333984375, 842.842529296875, 56.195716857910156, -751.1824340820312, 828.5260009765625, -208.59620666503906, 1233.841796875, 978.883544921875, 339.7836608886719, 601.3049926757812, 1050.5018310546875, 351.76617431640625, 995.3248901367188, 291.10443115234375, 802.992919921875, 28.59680938720703, 225.45982360839844, 594.7553100585938, 284.17877197265625, -422.512939453125, 223.29891967773438, 454.3254089355469, -92.81419372558594, -870.9443359375, 86.11405944824219, 1001.3829345703125, -14.678592681884766, 226.0804901123047, 275.9292297363281, 749.3034057617188, 161.19354248046875, 1339.070068359375, 98.38426971435547, 1188.071533203125, 521.5948486328125, 109.8014144897461, 1110.6417236328125, 113.84678649902344, 997.8587036132812, 358.320068359375, 481.68756103515625, 650.09521484375, 246.099853515625, 987.0313720703125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000537.npy"}
{"epoch": 0.788546255506608, "step": 538, "batch_size": 64, "mean": 361.36297607421875, "std": 472.5987854003906, "min": -481.44873046875, "p10": -211.68746643066407, "median": 358.3570098876953, "p90": 980.2711486816407, "max": 1865.8330078125, "pos_frac": 0.765625, "sample": [101.90510559082031, 424.61236572265625, 612.3515014648438, 1126.19873046875, 470.58050537109375, -141.3927001953125, 939.0165405273438, 363.0816955566406, 589.5082397460938, 245.63455200195312, -60.16686248779297, 1398.9100341796875, 579.6298828125, 394.602783203125, 160.45982360839844, 991.4168701171875, -481.44873046875, 12.722343444824219, 338.35333251953125, 1081.1546630859375, 284.89410400390625, 498.4232482910156, 59.340484619140625, 771.4906005859375, 603.244140625, 97.36763000488281, -19.7655029296875, 1348.89697265625, -29.054580688476562, -212.24087524414062, -448.59539794921875, 802.693359375, 179.3022918701172, -123.97392272949219, 1865.8330078125, 612.605712890625, 905.8041381835938, 366.39862060546875, 353.63232421875, 654.5535888671875, 160.15521240234375, -94.56940460205078, 13.677059173583984, 612.8516845703125, 307.6655578613281, 532.703125, -437.4280090332031, 506.1153869628906, -273.0277099609375, 109.76458740234375, 546.90673828125, 388.53167724609375, 367.1396484375, 567.6358032226562, -250.49864196777344, 11.225532531738281, 175.8035888671875, 62.3940544128418, 1083.67529296875, 954.2644653320312, -218.01173400878906, -210.39617919921875, 498.7588195800781, -6.086130142211914], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000538.npy"}
{"epoch": 0.7900146842878121, "step": 539, "batch_size": 64, "mean": 332.738525390625, "std": 424.83343505859375, "min": -644.1783447265625, "p10": -128.75851135253905, "median": 284.3922119140625, "p90": 855.7549804687502, "max": 1513.31005859375, "pos_frac": 0.75, "sample": [257.5256042480469, 346.97174072265625, 56.449851989746094, -28.105012893676758, 407.9498596191406, 425.12457275390625, -215.6840057373047, 84.71125030517578, -45.0771484375, 1151.57373046875, -246.01617431640625, -24.592355728149414, 1278.539306640625, 134.1747589111328, 873.1166381835938, 662.6395263671875, 711.8444213867188, 684.6185913085938, -43.84796142578125, 786.89404296875, 1513.31005859375, 150.14483642578125, 268.805419921875, 706.879638671875, 136.94862365722656, 174.64743041992188, 84.10057067871094, 444.41436767578125, -10.105634689331055, 554.1572875976562, -365.6909484863281, 514.07421875, -644.1783447265625, 501.50885009765625, 564.9766845703125, 51.5550537109375, 38.678810119628906, 495.5928955078125, 193.52566528320312, -149.41220092773438, -62.31053161621094, 1151.205810546875, -102.83883666992188, 765.5379638671875, -70.38328552246094, -139.866943359375, 945.2578735351562, -174.27159118652344, 176.8389892578125, -86.39534759521484, 581.3919067382812, 162.61459350585938, 673.7393188476562, 369.93829345703125, 374.2306213378906, 299.97900390625, 815.2444458007812, 488.222412109375, 436.6873779296875, 468.150146484375, 11.726362228393555, 1161.2181396484375, 553.7496337890625, 12.854658126831055], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000539.npy"}
{"epoch": 0.7914831130690162, "step": 540, "batch_size": 64, "mean": 446.42547607421875, "std": 598.4830322265625, "min": -780.874755859375, "p10": -247.72854919433584, "median": 329.4580383300781, "p90": 1325.0549316406252, "max": 2299.962646484375, "pos_frac": 0.78125, "sample": [456.4935607910156, 576.41455078125, 724.842041015625, -338.52325439453125, 1907.565185546875, -21.97540283203125, 254.80862426757812, 62.89898681640625, -780.874755859375, -408.7664794921875, -39.177284240722656, 514.3499755859375, -88.33753967285156, 316.6763916015625, 2299.962646484375, -100.72217559814453, 249.30560302734375, 85.44650268554688, 629.4243774414062, 700.6464233398438, 16.665618896484375, 199.91619873046875, 391.7298583984375, 700.8243408203125, 735.307373046875, 1179.8863525390625, 518.1487426757812, 1243.037353515625, 1229.1531982421875, 408.9078674316406, 611.3099365234375, 322.58306884765625, 752.9607543945312, 111.6100845336914, 244.24432373046875, 487.6253967285156, -144.68142700195312, -16.538406372070312, 1512.5810546875, 258.0823669433594, 975.4261474609375, 1654.162353515625, 1450.135498046875, 1360.205322265625, -458.58154296875, 712.6165771484375, 232.84890747070312, 336.3330078125, 581.9819946289062, 320.1849365234375, -291.8916015625, 261.2987976074219, 645.0590209960938, 659.2230834960938, 154.97698974609375, 266.26068115234375, 693.329345703125, 172.35418701171875, 1440.040771484375, -47.83759307861328, 8.050270080566406, 643.1536865234375, -562.5927124023438, -399.3215637207031], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000540.npy"}
{"epoch": 0.7929515418502202, "step": 541, "batch_size": 64, "mean": 241.50057983398438, "std": 552.6201782226562, "min": -1153.8902587890625, "p10": -381.5276824951172, "median": 187.84393310546875, "p90": 888.6762023925785, "max": 1884.3790283203125, "pos_frac": 0.734375, "sample": [-743.1160888671875, 189.73709106445312, 523.1689453125, 75.69985961914062, 185.95077514648438, 50.528533935546875, 287.2768249511719, 97.81282043457031, 258.31494140625, -95.742431640625, -69.23129272460938, 1233.79150390625, -842.641357421875, -13.387008666992188, 929.7489013671875, 46.840126037597656, 246.50067138671875, 1194.992431640625, 99.93160247802734, 704.1658935546875, 1446.07080078125, 792.8399047851562, 5.813934326171875, -557.2308349609375, 218.38385009765625, -251.2327880859375, 425.3626708984375, 278.73236083984375, -388.42822265625, 671.0228271484375, 673.7366943359375, 60.75768280029297, 755.175537109375, 172.4301300048828, 150.02304077148438, 143.76934814453125, 165.4744873046875, 162.2649383544922, 558.4725952148438, 464.345458984375, -84.31278991699219, 158.8194122314453, 699.9658813476562, 976.0570678710938, 623.4449462890625, 335.6673583984375, -1153.8902587890625, -489.5322265625, 459.0592346191406, 573.68310546875, -178.35928344726562, 405.76202392578125, 296.5830383300781, 1884.3790283203125, 1157.919189453125, -282.44647216796875, -1046.072509765625, 576.65380859375, -365.4264221191406, -231.02920532226562, -186.903564453125, 176.062744140625, 573.0548095703125, 268.7713623046875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000541.npy"}
{"epoch": 0.7944199706314243, "step": 542, "batch_size": 64, "mean": 452.58001708984375, "std": 600.2510986328125, "min": -1150.585205078125, "p10": -225.9923385620117, "median": 404.1480712890625, "p90": 1228.6484008789066, "max": 2052.203125, "pos_frac": 0.796875, "sample": [1060.735595703125, 405.93743896484375, 242.29238891601562, 254.46334838867188, 180.68389892578125, 983.5241088867188, -283.77294921875, 2052.203125, 1277.4793701171875, 397.34527587890625, 57.33738708496094, 461.9608154296875, 778.3794555664062, -120.14498901367188, -41.132041931152344, 448.06689453125, 1887.85107421875, 425.2998046875, -44.86821746826172, 1814.024658203125, 289.8777770996094, 101.65425109863281, 454.06756591796875, 425.16375732421875, -276.9466857910156, 256.87030029296875, -77.02191162109375, -334.97601318359375, -235.39735412597656, 983.6629638671875, -204.04730224609375, 867.3062744140625, 786.2178344726562, 198.87686157226562, 1400.997314453125, -78.62576293945312, 637.6832275390625, 521.0947265625, -525.4462280273438, 606.8031005859375, 356.734130859375, 367.3120422363281, 719.261474609375, 114.2948226928711, 1438.805419921875, 276.8571472167969, 249.35153198242188, -1150.585205078125, 711.4628295898438, 184.3788299560547, 703.2271118164062, 324.88323974609375, 111.07072448730469, 402.35870361328125, 1663.6109619140625, 1020.4099731445312, 451.5917663574219, -769.7946166992188, 1114.70947265625, 196.66909790039062, 479.0786437988281, 540.036376953125, 596.9224853515625, 826.9931640625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000542.npy"}
{"epoch": 0.7958883994126285, "step": 543, "batch_size": 64, "mean": 315.81573486328125, "std": 735.4705810546875, "min": -904.5606079101562, "p10": -363.4366485595703, "median": 153.93374633789062, "p90": 1371.0229980468762, "max": 3236.733642578125, "pos_frac": 0.671875, "sample": [22.53339385986328, -99.2015380859375, 780.268798828125, -6.40362548828125, 16.036306381225586, 2225.445068359375, 5.397449493408203, -11.216751098632812, -904.5606079101562, 724.69482421875, -97.53482818603516, -775.5281982421875, 22.734981536865234, 150.03759765625, 2147.14208984375, 73.06129455566406, 37.95481872558594, 3236.733642578125, -469.5908508300781, 20.06610870361328, 554.025390625, -503.83880615234375, -314.63427734375, 194.8134002685547, -373.5774230957031, -83.1504898071289, -104.65167236328125, 556.7947387695312, -460.74139404296875, 186.3320770263672, 560.5938720703125, 225.19627380371094, -339.77484130859375, 1532.44384765625, -234.12277221679688, -166.2504425048828, 1883.543701171875, 635.1673583984375, 1059.684326171875, 65.61327362060547, 1504.453857421875, 181.8366241455078, -268.4444580078125, 292.34832763671875, 789.7324829101562, 39.93321990966797, 306.016845703125, 1713.5806884765625, -398.95635986328125, 598.0630493164062, 236.33663940429688, -145.77496337890625, 260.3956298828125, -320.8132019042969, 450.154296875, 588.9053955078125, 315.1219177246094, -222.20803833007812, 45.38383483886719, 512.139892578125, 157.82989501953125, 459.5502014160156, 636.0177001953125, 509.06634521484375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000543.npy"}
{"epoch": 0.7973568281938326, "step": 544, "batch_size": 64, "mean": 366.1580810546875, "std": 560.6038818359375, "min": -1112.775634765625, "p10": -142.72235565185545, "median": 328.06927490234375, "p90": 1038.5672424316408, "max": 1803.045654296875, "pos_frac": 0.75, "sample": [1053.0684814453125, 589.9265747070312, 339.81756591796875, 441.9128112792969, 1568.0120849609375, -941.993408203125, 381.9427490234375, -53.01476287841797, 818.2535400390625, 35.1932373046875, 1162.04345703125, -10.04692268371582, 811.16796875, 271.0845947265625, -961.819091796875, 347.7793884277344, 570.6243286132812, 190.8175048828125, -30.54681396484375, 431.2671813964844, 316.32098388671875, 488.1279296875, -321.8880310058594, 239.50439453125, 76.04537963867188, 171.65858459472656, 735.9102172851562, 87.71368408203125, -330.55853271484375, -131.94326782226562, 1411.472900390625, 208.27484130859375, -120.69420623779297, 587.4830322265625, -1112.775634765625, 1004.7310180664062, 508.5648193359375, 487.1190185546875, 752.5218505859375, -113.29350280761719, 220.20538330078125, -7.618045806884766, 268.84857177734375, 1590.802001953125, 757.3350219726562, -24.930278778076172, 188.3605194091797, 1803.045654296875, 833.351806640625, 560.6992797851562, 12.535762786865234, 157.6368408203125, 654.5435791015625, 1360.466552734375, 660.7158203125, 294.0576477050781, 384.6091003417969, -147.3419647216797, 449.5644226074219, 166.23826599121094, -98.99070739746094, 786.0423583984375, -244.60198974609375, 848.7545776367188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000544.npy"}
{"epoch": 0.7988252569750367, "step": 545, "batch_size": 64, "mean": 430.1234130859375, "std": 551.2764282226562, "min": -666.0094604492188, "p10": -98.25227508544921, "median": 302.2013702392578, "p90": 1294.9048828125, "max": 1699.100830078125, "pos_frac": 0.84375, "sample": [-92.48281860351562, 17.941558837890625, -595.05908203125, 64.72579956054688, 258.05322265625, 801.9946899414062, 223.25946044921875, 64.39442443847656, 520.9644165039062, -441.9981384277344, 416.6711730957031, 731.8072509765625, 17.84246063232422, 629.072021484375, 1071.23388671875, -65.80419158935547, 1507.857177734375, 1300.1751708984375, 1256.27587890625, 304.2188415527344, 805.0845336914062, 56.71206283569336, 318.5091552734375, 217.69459533691406, 5.563865661621094, 302.6723327636719, 36.8477783203125, 431.8389892578125, 845.5133056640625, -102.6198959350586, 1410.3502197265625, 508.59405517578125, 135.0962677001953, -100.72489929199219, 1699.100830078125, 1024.8067626953125, 99.70326232910156, 1517.320556640625, -617.8831787109375, 438.69366455078125, -22.184532165527344, 460.4187316894531, 97.14642333984375, 578.159423828125, 382.0135192871094, 666.7241821289062, 810.6142578125, 1282.6075439453125, 276.65240478515625, 301.73040771484375, 197.18336486816406, 82.45492553710938, 254.63449096679688, 427.0566101074219, 113.03616333007812, 1588.4788818359375, 1236.54736328125, -195.10067749023438, 280.3293762207031, 181.16278076171875, 577.6488037109375, -666.0094604492188, 193.5543670654297, 1399.0224609375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000545.npy"}
{"epoch": 0.8002936857562408, "step": 546, "batch_size": 64, "mean": 287.52728271484375, "std": 456.5784912109375, "min": -1355.14990234375, "p10": -225.55201568603516, "median": 267.08656311035156, "p90": 869.6376220703125, "max": 1126.4686279296875, "pos_frac": 0.734375, "sample": [-262.1741638183594, 97.06543731689453, 431.2147216796875, 46.63916015625, 257.08343505859375, 533.9874877929688, 243.48745727539062, 503.7695007324219, 180.56756591796875, 224.8585205078125, -186.40621948242188, -225.77767944335938, -225.0254669189453, -192.19830322265625, 690.69970703125, 96.55546569824219, -9.456687927246094, 433.8073425292969, 980.5045166015625, 262.70654296875, -64.50133514404297, 772.479248046875, -307.6745910644531, 1082.6781005859375, 490.47589111328125, 858.8828125, 794.2706909179688, 470.91082763671875, 874.246826171875, 301.7309265136719, 1126.4686279296875, -322.6671142578125, -60.55602264404297, 426.2890319824219, 431.14007568359375, -667.7283325195312, 622.1510009765625, 372.73162841796875, 703.8060913085938, 876.9832153320312, -1355.14990234375, 706.93603515625, 242.4715576171875, 281.0301513671875, 461.5753479003906, 787.2860717773438, 62.478851318359375, 729.74853515625, -4.482887268066406, 165.81732177734375, 142.50982666015625, 184.5618133544922, 745.4286499023438, 271.4665832519531, -197.9096221923828, 595.9470825195312, 121.40420532226562, 1019.9641723632812, -170.45553588867188, 79.8075180053711, -94.51405334472656, 383.0857238769531, -360.07696533203125, 938.7907104492188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000546.npy"}
{"epoch": 0.801762114537445, "step": 547, "batch_size": 64, "mean": 367.3738098144531, "std": 527.8897705078125, "min": -1171.3935546875, "p10": -131.31367950439449, "median": 312.4628143310547, "p90": 1005.3617675781252, "max": 2039.1790771484375, "pos_frac": 0.734375, "sample": [457.3988037109375, 164.96014404296875, -1171.3935546875, 948.0947265625, -42.94953536987305, 28.095367431640625, 276.05072021484375, 694.6470947265625, 121.67352294921875, 526.9437866210938, 192.4029541015625, 1576.99462890625, 33.196292877197266, 1651.97998046875, -34.90330505371094, -4.453742980957031, 597.848388671875, 541.4684448242188, -150.0176239013672, 910.7559814453125, -38.97826385498047, 82.02838897705078, 485.1691589355469, 217.51657104492188, 596.3118896484375, -216.02764892578125, 571.12841796875, 148.2733612060547, 216.22206115722656, 297.9615478515625, 1029.90478515625, -67.57627868652344, 493.3488464355469, 671.931640625, 555.9927368164062, 374.68292236328125, -59.2501106262207, 892.8392944335938, 326.9640808105469, -60.81072998046875, 652.3963623046875, 2039.1790771484375, 575.6812133789062, -591.5387573242188, -186.00482177734375, -46.194618225097656, -315.21392822265625, 1050.35693359375, 115.90708923339844, 434.4330139160156, 1338.6826171875, 120.48826599121094, 422.5871276855469, 424.992919921875, 215.19020080566406, 343.74774169921875, 773.7994995117188, -87.671142578125, 401.126220703125, -72.342041015625, 697.2571411132812, 1288.306396484375, 259.97564697265625, -179.64447021484375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000547.npy"}
{"epoch": 0.8032305433186491, "step": 548, "batch_size": 64, "mean": 350.507080078125, "std": 573.5116577148438, "min": -888.0849609375, "p10": -201.64198150634766, "median": 291.3077392578125, "p90": 824.2004699707032, "max": 2667.826171875, "pos_frac": 0.765625, "sample": [0.8612823486328125, 832.5164794921875, 285.10150146484375, 324.6195983886719, -106.06519317626953, 366.44659423828125, -888.0849609375, 1322.12451171875, 401.8711853027344, 296.3782653808594, 410.56195068359375, 398.48663330078125, 1097.3271484375, 464.54534912109375, 625.4375610351562, -248.174560546875, -107.15376281738281, 333.2574157714844, 747.19189453125, 197.98828125, 804.7964477539062, 313.4999694824219, 206.60592651367188, -344.05877685546875, 503.7054748535156, 32.255210876464844, 201.38868713378906, 279.023681640625, 293.1405944824219, 201.2228546142578, -200.87307739257812, 2667.826171875, 847.9823608398438, 165.03387451171875, 146.78863525390625, -398.2815246582031, -73.54501342773438, 638.9299926757812, -278.4206848144531, 2325.239013671875, 30.386226654052734, -90.34408569335938, -201.9715118408203, 545.9331665039062, -38.242523193359375, 110.32809448242188, 395.10601806640625, 137.306396484375, 429.2923278808594, -179.23348999023438, 465.1558837890625, 416.2127990722656, -5.26255989074707, 1904.775146484375, 461.37847900390625, 591.82080078125, 105.17366027832031, -255.49649047851562, 264.55487060546875, 289.4748840332031, 56.293418884277344, 589.43310546875, 704.3148803710938, 618.566650390625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000548.npy"}
{"epoch": 0.8046989720998532, "step": 549, "batch_size": 64, "mean": 265.2655944824219, "std": 516.0823364257812, "min": -844.2067260742188, "p10": -193.52015075683593, "median": 220.2188491821289, "p90": 882.9943908691407, "max": 1691.084716796875, "pos_frac": 0.671875, "sample": [246.7464599609375, 868.5772705078125, 60.89369201660156, -113.6034927368164, -32.05690002441406, 35.513790130615234, -168.3349151611328, 455.2890625, 642.6524658203125, 822.7063598632812, -149.08253479003906, 1306.42626953125, 55.858558654785156, 157.67648315429688, -79.23513793945312, -315.6304931640625, 889.1731567382812, -386.1908874511719, -172.85391235351562, -638.7640380859375, 593.78857421875, 958.3206176757812, -194.5294189453125, 288.05401611328125, 331.51605224609375, 283.8380126953125, 357.7916259765625, -113.10564422607422, 686.9481811523438, 842.5715942382812, -65.65899658203125, 235.09364318847656, 260.26251220703125, -102.21614837646484, 349.4547119140625, -844.2067260742188, 241.51632690429688, 706.7059326171875, 58.043907165527344, -43.85233688354492, -20.327484130859375, 13.612098693847656, 287.7994384765625, 718.8515014648438, 1318.09130859375, -13.364578247070312, -117.57359313964844, 247.1885528564453, -316.64453125, 1666.3719482421875, 1274.733154296875, -767.752685546875, 13.193037033081055, 648.4840087890625, 404.68212890625, -191.16519165039062, 205.34405517578125, 194.16787719726562, 641.2959594726562, 1691.084716796875, 353.54901123046875, 314.25347900390625, 93.59498596191406, 1.4301376342773438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000549.npy"}
{"epoch": 0.8061674008810573, "step": 550, "batch_size": 64, "mean": 320.65582275390625, "std": 498.62042236328125, "min": -899.0997924804688, "p10": -300.3851196289061, "median": 271.7438201904297, "p90": 905.6694519042971, "max": 2023.131591796875, "pos_frac": 0.8125, "sample": [253.78146362304688, 256.2832946777344, 1184.3309326171875, 394.1970520019531, 226.51284790039062, 757.2382202148438, 1040.9144287109375, 126.22615051269531, -101.77928161621094, 273.908203125, 541.3106689453125, 390.61553955078125, 514.3134765625, 202.61709594726562, 335.48175048828125, 1669.4857177734375, 64.25560760498047, -349.0144348144531, 83.93307495117188, 105.23138427734375, -899.0997924804688, -4.6167144775390625, 476.2667236328125, 826.359375, -104.39152526855469, 817.0419311523438, 939.6594848632812, 120.61148834228516, 250.39755249023438, 100.25840759277344, 38.57450866699219, -653.5416870117188, 304.6998291015625, -132.82571411132812, 510.0007019042969, 748.2894897460938, 65.74256896972656, 446.0194091796875, 2023.131591796875, 440.1961975097656, 305.6610107421875, 665.4126586914062, 174.98988342285156, 60.210540771484375, 1176.4422607421875, 430.52996826171875, 1028.2247314453125, 130.70469665527344, 150.88229370117188, 580.0997924804688, 454.2459411621094, 522.1663818359375, -575.5625610351562, 616.986328125, -186.91671752929688, 269.5794372558594, 250.85585021972656, 449.9467468261719, 289.9214782714844, -380.9530029296875, 510.0987548828125, -384.34466552734375, -436.4940185546875, 136.66673278808594], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000550.npy"}
{"epoch": 0.8076358296622613, "step": 551, "batch_size": 64, "mean": 522.2348022460938, "std": 507.4519348144531, "min": -433.3917541503906, "p10": -73.2684585571289, "median": 458.0789794921875, "p90": 1316.72802734375, "max": 1841.66748046875, "pos_frac": 0.828125, "sample": [-48.80099105834961, -26.489532470703125, 474.5989074707031, 790.947021484375, 642.041259765625, 558.1441040039062, -433.3917541503906, -11.983470916748047, 1277.7105712890625, -190.41783142089844, 51.87126922607422, 353.7008056640625, -66.66108703613281, 532.000244140625, 559.7490844726562, 1323.7340087890625, 485.8515319824219, 420.448486328125, 589.8489379882812, 1636.157958984375, 900.72021484375, 453.1002197265625, 561.8475952148438, 619.9061279296875, 290.08001708984375, 263.44970703125, 925.4595336914062, 360.34063720703125, 543.1459350585938, 325.7657775878906, 413.61761474609375, -76.10018920898438, 1415.6787109375, 162.94049072265625, 533.365234375, 614.7969970703125, 689.5352172851562, 642.953125, 463.0577392578125, 84.82334899902344, -106.09335327148438, 1154.8587646484375, 1300.3807373046875, 418.0503845214844, 49.72956466674805, 1531.2086181640625, -302.5429382324219, -255.06607055664062, 1492.6893310546875, 766.5637817382812, -267.50347900390625, 134.16342163085938, 1335.25048828125, 1023.5130615234375, 401.3725891113281, 401.617919921875, 783.550048828125, 337.36114501953125, 403.62823486328125, 1012.3794555664062, 177.68421936035156, 1841.66748046875, 359.306884765625, 321.71484375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000551.npy"}
{"epoch": 0.8091042584434655, "step": 552, "batch_size": 64, "mean": 581.2489013671875, "std": 688.01318359375, "min": -675.3629150390625, "p10": -54.21402282714843, "median": 512.7667999267578, "p90": 1426.2276855468751, "max": 2835.53076171875, "pos_frac": 0.84375, "sample": [768.1875610351562, 76.87458038330078, 836.5492553710938, 2835.53076171875, 710.1624755859375, 1128.457763671875, 600.3887329101562, 832.6854858398438, -55.2454833984375, 557.8722534179688, 709.6212768554688, 511.8852233886719, 128.3805694580078, 536.664306640625, 267.329345703125, 770.848876953125, 164.32296752929688, 672.3461303710938, 1355.477294921875, -333.8790588378906, 83.64584350585938, 623.3193359375, 513.6483764648438, 290.2945556640625, -675.3629150390625, 23.07408905029297, 37.970062255859375, 1456.21923828125, 167.8646697998047, -275.4674377441406, 1371.1807861328125, -524.9302368164062, 705.7028198242188, -9.067390441894531, 564.0756225585938, 105.26054382324219, 436.57208251953125, 2379.460693359375, 1245.6260986328125, 1859.8304443359375, 508.44671630859375, -31.880882263183594, 488.92279052734375, 963.8785400390625, 621.7286376953125, 2545.760498046875, 260.3717041015625, 457.73797607421875, 1449.8192138671875, 2061.98486328125, 546.950927734375, 139.66390991210938, 459.54974365234375, -377.7286071777344, -51.807281494140625, 445.4016418457031, -106.66960144042969, 135.48426818847656, 460.40728759765625, 552.957275390625, 602.0778198242188, 931.075439453125, 649.6090087890625, 32.81005096435547], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000552.npy"}
{"epoch": 0.8105726872246696, "step": 553, "batch_size": 64, "mean": 305.9494934082031, "std": 532.3888549804688, "min": -579.3229370117188, "p10": -221.588264465332, "median": 202.56253051757812, "p90": 943.5046447753906, "max": 2463.70068359375, "pos_frac": 0.71875, "sample": [2463.70068359375, 946.3470458984375, -204.03768920898438, 535.454345703125, 447.968505859375, -156.12454223632812, 52.835113525390625, -34.90155792236328, 162.41883850097656, 318.84869384765625, 193.01333618164062, 225.47300720214844, 136.0037078857422, 98.90121459960938, -27.53974151611328, 248.96531677246094, -514.2633056640625, 13.339385986328125, 317.745849609375, -229.1099395751953, 154.59255981445312, 178.4026641845703, 341.1590270996094, -499.56707763671875, 483.80322265625, -0.7948150634765625, 1139.7578125, 100.11065673828125, 0.29999542236328125, 309.31988525390625, -163.6172637939453, 478.3589172363281, -73.88623046875, -146.4967498779297, 644.94091796875, 302.326416015625, 413.364013671875, 214.0921630859375, 912.45556640625, -121.7883529663086, 356.8206481933594, -579.3229370117188, 684.94287109375, 612.1596069335938, 171.87197875976562, 200.5352783203125, -494.360107421875, 102.80926513671875, 543.0458984375, -51.95972442626953, 1189.344482421875, -5.116889953613281, 1037.1082763671875, 611.28662109375, 910.346923828125, 1071.490234375, 204.58978271484375, 910.0236206054688, -525.5393676757812, 1666.883544921875, -242.782470703125, 936.8723754882812, 27.94892120361328, 579.8976440429688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000553.npy"}
{"epoch": 0.8120411160058737, "step": 554, "batch_size": 64, "mean": 402.7427978515625, "std": 682.7725830078125, "min": -1026.617431640625, "p10": -499.47865905761716, "median": 344.0158996582031, "p90": 1327.5079589843751, "max": 2064.333740234375, "pos_frac": 0.78125, "sample": [167.3981475830078, -772.17041015625, 720.5431518554688, 309.4380798339844, 141.91409301757812, 917.7276611328125, 433.2604064941406, 449.5535583496094, -515.7366943359375, 325.1607666015625, 2064.333740234375, 454.78857421875, 405.9540100097656, 261.5914611816406, -224.16322326660156, 507.38446044921875, 620.6134643554688, 592.431640625, 552.4051513671875, 2.8515396118164062, -603.2723999023438, -698.704345703125, -461.5432434082031, 4.005989074707031, 1765.3428955078125, -186.8253173828125, 308.5079040527344, 478.2870788574219, 526.8203735351562, 832.78173828125, 564.1618041992188, 1350.956298828125, 1580.2901611328125, 182.76998901367188, 970.7681884765625, 310.2254943847656, 963.5780029296875, 4.838985443115234, 1724.5177001953125, -173.00091552734375, -100.51528930664062, 232.63763427734375, 1132.692626953125, 595.382080078125, 1024.8040771484375, 53.0720100402832, 261.43035888671875, 2.666900634765625, 859.1289672851562, 803.7898559570312, -946.4091186523438, 718.953125, 174.79934692382812, -673.4107666015625, 1076.0281982421875, -1026.617431640625, -342.8457336425781, 0.07459068298339844, 362.87103271484375, 1272.795166015625, 1735.9893798828125, 1707.9744873046875, 118.36969757080078, -127.9068603515625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000554.npy"}
{"epoch": 0.8135095447870778, "step": 555, "batch_size": 64, "mean": 478.2687072753906, "std": 737.2760620117188, "min": -1168.04052734375, "p10": -171.20047149658203, "median": 390.9259796142578, "p90": 1541.746325683594, "max": 2399.522216796875, "pos_frac": 0.78125, "sample": [507.0184326171875, 355.07379150390625, 1457.453125, 339.8422546386719, 571.9990844726562, 1296.3209228515625, 451.1876220703125, 171.03085327148438, -136.15264892578125, 207.73153686523438, -6.1347503662109375, 715.2356567382812, 716.8841552734375, 382.5419921875, -22.454345703125, 44.008087158203125, 210.98269653320312, -291.1914367675781, 2399.522216796875, 1038.6610107421875, 177.76129150390625, 1943.515625, 51.16071319580078, 516.096435546875, 613.0241088867188, 133.22500610351562, 819.0767822265625, 39.30549621582031, 471.3428649902344, 104.68388366699219, 625.280029296875, 1046.887939453125, 524.4219360351562, 440.6817932128906, 19.0545654296875, 188.01438903808594, 755.6064453125, -810.340576171875, 191.1352996826172, -133.79052734375, -937.241455078125, 1208.1983642578125, 1558.5062255859375, 2254.083251953125, -497.9859924316406, 539.675048828125, 1916.3687744140625, 74.18406677246094, 428.4836730957031, 94.16802978515625, 1502.639892578125, -539.2453002929688, 1560.158935546875, -1168.04052734375, -159.22601318359375, -110.29915618896484, 2213.35400390625, 436.3086853027344, -143.31692504882812, 252.73248291015625, 1140.4051513671875, 399.3099670410156, -176.33238220214844, 636.6064453125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000555.npy"}
{"epoch": 0.8149779735682819, "step": 556, "batch_size": 64, "mean": 279.4118957519531, "std": 448.75933837890625, "min": -437.69873046875, "p10": -250.51334686279296, "median": 230.82340240478516, "p90": 862.6483886718752, "max": 1516.070556640625, "pos_frac": 0.703125, "sample": [1217.25, 158.6317138671875, 106.01699829101562, -192.740234375, 355.7208557128906, 145.03359985351562, 589.48388671875, 457.7734680175781, -186.73291015625, 389.66058349609375, -18.564674377441406, 695.7693481445312, 276.23175048828125, -404.0845642089844, 210.25634765625, 440.740234375, 1293.4500732421875, 4.0540771484375, 703.9808349609375, -142.17254638671875, -437.69873046875, -69.55977630615234, -389.7850341796875, 124.19522857666016, -343.373779296875, -145.42999267578125, 230.42257690429688, -253.2525177001953, 879.4183959960938, -244.1219482421875, 310.28887939453125, 275.46820068359375, 261.2427673339844, 107.86280059814453, 813.5571899414062, -45.455894470214844, 477.1873779296875, 118.05467224121094, 187.12557983398438, -33.825416564941406, -114.546630859375, 736.4991455078125, 107.08866882324219, 190.3511962890625, -262.8413391113281, 641.2169799804688, 930.7198486328125, 450.14862060546875, 989.0154418945312, 77.3830337524414, 234.54043579101562, -167.7353515625, 823.5183715820312, 263.62579345703125, 1516.070556640625, -78.66362762451172, 780.609619140625, 1357.0196533203125, 501.5705871582031, 476.1561279296875, 382.5791015625, -394.6463928222656, 231.22422790527344, 289.3775939941406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000556.npy"}
{"epoch": 0.8164464023494861, "step": 557, "batch_size": 64, "mean": 441.0848083496094, "std": 695.3358154296875, "min": -1436.45458984375, "p10": -330.8173309326171, "median": 390.5575256347656, "p90": 1342.0593872070315, "max": 2470.3759765625, "pos_frac": 0.796875, "sample": [-1436.45458984375, 253.30235290527344, -964.013916015625, 723.8232421875, 587.155517578125, 223.14453125, 95.31367492675781, 486.1767883300781, 600.944580078125, 84.44839477539062, 1143.518798828125, 1362.35498046875, 986.3710327148438, 100.7843017578125, 142.0164337158203, 345.66436767578125, 469.994873046875, 797.6856689453125, -11.193721771240234, 96.22351837158203, 83.64015197753906, 1213.742919921875, 1578.4521484375, 579.04150390625, 5.03582763671875, 528.70263671875, 2.6900634765625, -434.57916259765625, -208.97567749023438, 946.125244140625, 211.212158203125, 546.115966796875, 1072.8336181640625, 555.0405883789062, 296.7274475097656, 219.79733276367188, 501.5364074707031, -704.4518432617188, 830.8319702148438, 500.3189697265625, 838.226318359375, -360.7735900878906, 544.7896118164062, 136.7451629638672, 157.9695587158203, -481.86474609375, 742.5364990234375, 1876.85888671875, -358.9252014160156, 2470.3759765625, 18.939037322998047, 2180.180908203125, 1533.45166015625, 192.69705200195312, 1048.4036865234375, -90.07136535644531, -181.1353759765625, 1452.9710693359375, -265.2322998046875, 1294.7030029296875, -70.1036605834961, 451.5879211425781, 435.45068359375, 250.54624938964844], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000557.npy"}
{"epoch": 0.8179148311306902, "step": 558, "batch_size": 64, "mean": 580.1013793945312, "std": 625.96435546875, "min": -1585.372314453125, "p10": -78.36856536865233, "median": 502.5927429199219, "p90": 1319.3754882812502, "max": 2033.400634765625, "pos_frac": 0.8125, "sample": [370.94976806640625, 1145.252685546875, 214.51585388183594, 1418.011962890625, 181.75115966796875, 332.1623840332031, 514.5601806640625, 64.97187805175781, 817.4369506835938, -414.8162536621094, -175.49276733398438, 1073.2464599609375, 1263.369873046875, 1005.051513671875, 391.966796875, 1134.23095703125, 490.62530517578125, 296.2793273925781, 1534.696533203125, 1832.009033203125, -325.20526123046875, 351.1439514160156, 318.6138916015625, 1115.796875, -26.61042022705078, 633.993408203125, 1333.7646484375, 413.45404052734375, 1285.80078125, 461.2850341796875, 520.0144653320312, 599.703857421875, 935.9984741210938, 1269.2147216796875, 1459.3565673828125, 865.0811157226562, -1585.372314453125, -84.29074096679688, 547.1453857421875, 241.1568145751953, 351.9696350097656, -3.175382614135742, 110.93870544433594, 400.631103515625, 473.3885498046875, -64.55015563964844, -180.0794219970703, 2033.400634765625, 1032.5074462890625, -1.5131473541259766, 755.267578125, 1138.6534423828125, 544.0660400390625, 539.841796875, 1220.611083984375, -664.31298828125, 868.1065673828125, 327.3682556152344, 475.88079833984375, -23.288665771484375, 281.2848205566406, 1445.6739501953125, 1161.1033935546875, 1081.89013671875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000558.npy"}
{"epoch": 0.8193832599118943, "step": 559, "batch_size": 64, "mean": 486.8303527832031, "std": 612.257568359375, "min": -798.2090454101562, "p10": -111.03283767700194, "median": 316.1628723144531, "p90": 1242.0261108398438, "max": 2247.13330078125, "pos_frac": 0.828125, "sample": [963.7036743164062, 1294.8905029296875, 653.2080078125, 624.9727783203125, 71.32825469970703, 49.65022277832031, 94.21461486816406, 151.7677764892578, 1243.0352783203125, 2247.13330078125, 1307.75439453125, 337.985107421875, 952.655029296875, -31.814729690551758, 194.51217651367188, 762.623779296875, 923.8955078125, 1239.67138671875, -798.2090454101562, -417.56756591796875, 1074.8624267578125, 294.34063720703125, 208.70596313476562, 426.38275146484375, 176.4057159423828, 1081.779052734375, -165.5126953125, -117.47076416015625, 1227.728271484375, 217.05889892578125, 80.39826202392578, 255.79827880859375, 114.94739532470703, 82.0436782836914, 184.7893524169922, -29.909347534179688, 14.51266860961914, -96.0110092163086, 228.36390686035156, 256.75677490234375, 1934.7265625, 521.9385375976562, 1008.465576171875, 105.1371841430664, -712.2015380859375, 690.334716796875, 501.4416198730469, 610.9191284179688, 1131.4130859375, -14.596214294433594, 528.59521484375, 250.27130126953125, 425.0672607421875, 1124.5540771484375, 1421.067138671875, 1892.80322265625, -371.0223693847656, 843.5687866210938, 975.3717041015625, 459.1519470214844, 169.63478088378906, 26.862777709960938, -437.17596435546875, 689.433837890625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000559.npy"}
{"epoch": 0.8208516886930984, "step": 560, "batch_size": 64, "mean": 344.74053955078125, "std": 571.6947021484375, "min": -1021.5691528320312, "p10": -267.54644470214845, "median": 255.78060913085938, "p90": 1168.5387329101563, "max": 2019.4261474609375, "pos_frac": 0.75, "sample": [2019.4261474609375, -373.92364501953125, -168.55075073242188, 1017.302978515625, -463.1128234863281, 615.4223022460938, 86.2340316772461, -68.20844268798828, 95.06561279296875, 302.7059020996094, 246.05865478515625, 168.56875610351562, 256.775634765625, 544.714111328125, 1493.1668701171875, -202.85394287109375, 695.358154296875, 1295.9012451171875, 389.20452880859375, 938.1870727539062, 715.996826171875, 2.21478271484375, 191.9796142578125, 1400.2669677734375, 853.2905883789062, -539.8630981445312, 18.420536041259766, 1160.711669921875, 394.3607177734375, -168.39535522460938, 889.1472778320312, -129.3137969970703, 27.403331756591797, 476.94140625, 499.581298828125, -509.99072265625, 102.63239288330078, 1395.6231689453125, 264.2137451171875, 561.2210083007812, 449.821044921875, -145.38739013671875, 254.78558349609375, 155.94430541992188, 743.3170776367188, 1171.8931884765625, -9.88286018371582, -58.28264617919922, -267.4851989746094, 1288.385498046875, 609.462646484375, 228.41604614257812, 348.4073791503906, 239.3612518310547, 322.4732666015625, 226.98452758789062, -267.57269287109375, 489.66204833984375, 739.9242553710938, 700.9815673828125, -731.2205810546875, 100.45848846435547, -1021.5691528320312, 0.63238525390625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000560.npy"}
{"epoch": 0.8223201174743024, "step": 561, "batch_size": 64, "mean": 291.81158447265625, "std": 644.7881469726562, "min": -2341.018310546875, "p10": -261.3713928222656, "median": 231.07373046875, "p90": 1069.5909973144535, "max": 1848.7647705078125, "pos_frac": 0.734375, "sample": [588.6134643554688, 451.579345703125, 154.54318237304688, 19.93709373474121, -1185.28173828125, 1238.02197265625, 801.9589233398438, 81.77024841308594, 650.5357055664062, 1107.5230712890625, 131.29547119140625, 503.6471252441406, 1848.7647705078125, -169.61488342285156, 169.51651000976562, 206.78079223632812, 72.51263427734375, -166.41549682617188, 367.87884521484375, 349.45794677734375, 1365.651611328125, 152.86843872070312, 567.0657348632812, 262.3966064453125, -229.29354858398438, 187.1836395263672, 884.84521484375, 393.04522705078125, -41.55182647705078, 1674.2230224609375, 594.4154052734375, 1237.3280029296875, 1.1827545166015625, 265.0816955566406, 864.6484375, -544.7453002929688, -347.5841369628906, 641.5003051757812, 20.971986770629883, 625.6439208984375, 163.4097137451172, -403.5386047363281, -183.84423828125, 800.2941284179688, 572.561767578125, -155.2294921875, 171.44561767578125, 734.3123168945312, 981.0828247070312, 222.39694213867188, -2341.018310546875, 1574.44873046875, 154.2229766845703, -130.44207763671875, 391.68701171875, -185.56439208984375, 467.7117919921875, -70.03147888183594, 614.8040771484375, 293.7955322265625, -523.8973388671875, 239.75051879882812, -270.6909484863281, -239.62576293945312], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000561.npy"}
{"epoch": 0.8237885462555066, "step": 562, "batch_size": 64, "mean": 403.70263671875, "std": 648.3153686523438, "min": -1229.0517578125, "p10": -249.23801879882808, "median": 404.1026306152344, "p90": 1062.4579589843752, "max": 2759.274658203125, "pos_frac": 0.75, "sample": [445.6430969238281, 2759.274658203125, 1358.7760009765625, 457.6830139160156, 56.069366455078125, 574.0160522460938, 718.2713012695312, 199.56346130371094, 1043.673828125, 651.5572509765625, 772.1556396484375, 963.048828125, 220.05801391601562, 454.3459777832031, 1088.6976318359375, -271.513427734375, 249.8466796875, 885.3651123046875, 245.88168334960938, 430.2297058105469, 414.4898681640625, 662.3297729492188, 468.20587158203125, -354.9080505371094, 25.356895446777344, 740.4740600585938, 968.1021728515625, -54.640777587890625, 1096.07080078125, 251.31939697265625, -267.25982666015625, -207.1871337890625, -93.64521789550781, 748.505615234375, -166.4639892578125, 2323.929443359375, -137.66766357421875, 471.77740478515625, 616.634521484375, 1042.899169921875, 900.2610473632812, 207.0227813720703, 595.912841796875, 281.7492980957031, -1229.0517578125, 236.90228271484375, -166.38954162597656, 870.0910034179688, 393.71539306640625, 987.4669799804688, -17.267391204833984, -116.10209655761719, -105.1112289428711, 59.2205924987793, 601.94287109375, 191.36280822753906, -904.4542846679688, 36.251609802246094, 95.21672821044922, -671.5384521484375, 89.73037719726562, -606.003662109375, 1070.50830078125, 1184.565185546875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000562.npy"}
{"epoch": 0.8252569750367107, "step": 563, "batch_size": 64, "mean": 309.92266845703125, "std": 755.4550170898438, "min": -3015.862548828125, "p10": -447.49941101074216, "median": 372.95025634765625, "p90": 1118.667077636719, "max": 1736.1949462890625, "pos_frac": 0.765625, "sample": [242.01763916015625, -3015.862548828125, -952.7977905273438, 463.9908447265625, 759.628173828125, 81.04296875, 1275.9151611328125, 23.228424072265625, 331.4665222167969, 718.305908203125, 1122.74267578125, 328.0689697265625, -661.7805786132812, 984.8927001953125, 842.7803344726562, 767.5072021484375, 107.85249328613281, 1700.168212890625, 630.6425170898438, -326.019775390625, 631.317138671875, 41.23413848876953, -166.47018432617188, 282.062255859375, 1109.1573486328125, -47.0592155456543, 719.8975219726562, 327.51593017578125, 1221.6590576171875, 866.5985107421875, 135.08673095703125, 732.3357543945312, -1600.79150390625, 514.8541259765625, -105.89497375488281, 196.42889404296875, 21.12125015258789, 694.8496704101562, 589.0645141601562, 397.8037109375, 748.1843872070312, 540.771240234375, 1218.014404296875, 220.90753173828125, 1039.2315673828125, -130.8763427734375, -1210.4278564453125, 1736.1949462890625, -548.1182250976562, 1363.5218505859375, -89.62570190429688, 703.75732421875, 71.53772735595703, -447.8714904785156, 763.8621826171875, -444.3370666503906, 479.25665283203125, -446.6312255859375, 114.57398986816406, 596.5590209960938, 348.0968017578125, 416.8614196777344, 310.7803649902344, 496.2687072753906], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000563.npy"}
{"epoch": 0.8267254038179148, "step": 564, "batch_size": 64, "mean": 393.664794921875, "std": 529.6962890625, "min": -499.0521240234375, "p10": -221.46762542724608, "median": 332.60986328125, "p90": 1062.34794921875, "max": 1964.5911865234375, "pos_frac": 0.765625, "sample": [347.71673583984375, 564.9871826171875, 780.2218017578125, 121.46568298339844, 698.0531005859375, 547.348388671875, 717.833740234375, -417.6221008300781, 314.7902526855469, -391.0338439941406, 448.3848571777344, 634.0162963867188, -326.0008544921875, 469.9826965332031, 1043.435302734375, 32.761322021484375, -216.8493194580078, 1225.38671875, 1070.453369140625, 397.59112548828125, 360.3204040527344, 268.5167236328125, 191.4287567138672, -208.0993194580078, -455.13409423828125, 349.3931884765625, -230.14712524414062, 291.0523986816406, 25.134490966796875, -499.0521240234375, 355.7785339355469, 1895.2828369140625, 627.9641723632812, -158.7752685546875, 76.7756118774414, 399.220458984375, 191.7906036376953, 60.308677673339844, 431.67041015625, -8.21084213256836, 587.9674072265625, 1964.5911865234375, -43.6737060546875, 647.052978515625, 770.3333129882812, 780.9127197265625, 1447.833251953125, 818.4705810546875, 1299.297119140625, 288.9805908203125, -67.18421173095703, 659.6985473632812, -4.201622009277344, 59.42546081542969, 105.02320861816406, 242.0592498779297, 216.49366760253906, 914.2750244140625, 317.50299072265625, -82.79798889160156, 1343.7454833984375, -223.4468994140625, 964.5714111328125, 159.4754638671875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000564.npy"}
{"epoch": 0.8281938325991189, "step": 565, "batch_size": 64, "mean": 426.8079833984375, "std": 725.1729736328125, "min": -1219.000732421875, "p10": -363.91133117675776, "median": 273.4913024902344, "p90": 1177.548913574219, "max": 2370.005615234375, "pos_frac": 0.671875, "sample": [-26.997089385986328, 1089.580078125, 1380.271484375, -1219.000732421875, 255.84490966796875, 992.4946899414062, 2035.6005859375, -58.79554748535156, -149.76832580566406, -323.51531982421875, 547.558349609375, 1157.2469482421875, -294.3831787109375, -886.6805419921875, 325.284423828125, -535.8861694335938, -381.2239074707031, -132.13067626953125, -104.07315063476562, 514.1160278320312, 1069.3931884765625, 56.750221252441406, 221.4207763671875, 1577.8992919921875, -537.8408813476562, -58.50865936279297, 270.2484130859375, 319.1078186035156, -604.31298828125, 1067.9990234375, 510.665283203125, 221.37986755371094, 654.0576171875, 902.2882690429688, 729.7068481445312, -1.9429473876953125, 890.1721801757812, 587.8991088867188, -271.7625427246094, -33.50218963623047, -72.57157897949219, 877.2882080078125, 694.1697998046875, 1875.3433837890625, 788.0993041992188, 2370.005615234375, 565.3106689453125, 2364.951416015625, 1119.0765380859375, -430.9047546386719, -43.1744384765625, 14.026046752929688, 967.8345947265625, 468.0204772949219, 881.312255859375, 239.41775512695312, 1051.580322265625, 271.53472900390625, 27.905879974365234, 68.3136978149414, -211.83953857421875, 275.4478759765625, 211.6536102294922, 1186.249755859375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000565.npy"}
{"epoch": 0.8296622613803231, "step": 566, "batch_size": 64, "mean": 463.4237060546875, "std": 734.962646484375, "min": -1316.117919921875, "p10": -359.15968933105466, "median": 524.9811706542969, "p90": 1522.3815307617188, "max": 2128.439697265625, "pos_frac": 0.6875, "sample": [-169.94696044921875, -644.8618774414062, 667.5772705078125, 1758.4024658203125, 388.8337707519531, -144.89804077148438, 648.5601806640625, 154.78103637695312, -1007.4862060546875, 515.4346313476562, 599.654296875, -271.10906982421875, -24.8702392578125, 632.1439819335938, 670.0805053710938, 436.49371337890625, 772.5516357421875, 37.6292610168457, 18.369831085205078, -1316.117919921875, 1070.2161865234375, 635.7825317382812, 534.5277099609375, -30.89239501953125, 1151.040771484375, -487.4656066894531, 592.7670288085938, 1202.9189453125, -59.16929626464844, 1053.1416015625, 926.7401733398438, 922.3897705078125, 746.8648071289062, 103.56932067871094, 985.1494140625, -318.3410339355469, -137.6035919189453, 1971.5050048828125, 846.7306518554688, -255.45399475097656, -273.52484130859375, 712.1780395507812, 486.0958251953125, 1530.4306640625, 1503.6002197265625, 1882.369140625, 2128.439697265625, 719.5211181640625, -789.6890869140625, 460.5494384765625, -672.58203125, 1582.0849609375, 747.61328125, 1591.2330322265625, -44.04912567138672, -371.336669921875, 1112.803466796875, 318.220458984375, 1011.6410522460938, -330.7467346191406, 31.29560089111328, -77.945068359375, 712.1709594726562, 513.1041259765625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000566.npy"}
{"epoch": 0.8311306901615272, "step": 567, "batch_size": 64, "mean": 342.280029296875, "std": 477.1089782714844, "min": -1224.62646484375, "p10": -93.70272369384762, "median": 332.47300720214844, "p90": 907.0271423339844, "max": 1809.183837890625, "pos_frac": 0.796875, "sample": [94.14971160888672, -11.309183120727539, 1182.146240234375, 530.981201171875, 347.675048828125, 633.21875, 130.95941162109375, -109.33817291259766, 980.4851684570312, 159.70648193359375, 812.4208374023438, 12.307075500488281, 1065.281982421875, 282.45196533203125, 176.28067016601562, 276.6922607421875, 1125.6688232421875, 616.3118286132812, 560.3799438476562, 86.25248718261719, 538.6975708007812, 147.58792114257812, -230.60707092285156, 657.4429931640625, 333.7701721191406, 788.1712036132812, 727.3848876953125, 1348.404541015625, -121.44017028808594, 649.12548828125, -25.152414321899414, 340.16448974609375, 348.79742431640625, -1224.62646484375, 341.4066162109375, -34.11360549926758, -5.021205902099609, 709.5859375, 1809.183837890625, 908.8864135742188, 128.83676147460938, -854.9640502929688, 308.4582824707031, 91.01947784423828, 231.63697814941406, 902.6888427734375, 346.2904052734375, 208.80230712890625, 343.63360595703125, 456.3231201171875, 141.10472106933594, 814.2734985351562, -20.59223175048828, 424.883544921875, -169.25453186035156, 331.17584228515625, 158.545166015625, 220.1017303466797, -57.220008850097656, 205.65826416015625, 354.89251708984375, 490.9953308105469, -510.75518798828125, 399.01617431640625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000567.npy"}
{"epoch": 0.8325991189427313, "step": 568, "batch_size": 64, "mean": 478.24029541015625, "std": 529.9559326171875, "min": -737.1445922851562, "p10": -137.83275909423816, "median": 407.13966369628906, "p90": 1051.257861328125, "max": 1853.343505859375, "pos_frac": 0.875, "sample": [721.3191528320312, -717.6000366210938, 1155.0732421875, 798.7506103515625, 522.5550537109375, 420.3046569824219, 770.5338134765625, 759.501953125, 409.5408630371094, -737.1445922851562, -357.672119140625, -192.61654663085938, 610.826904296875, 1853.343505859375, 196.16067504882812, 518.9869995117188, 224.67745971679688, 176.72137451171875, 298.80450439453125, -398.9788818359375, 634.461181640625, 395.5469055175781, 1017.7164306640625, 965.4384155273438, 398.4171447753906, 294.8580322265625, 151.17434692382812, 1771.0654296875, 271.4794006347656, 432.6782531738281, 946.005615234375, 763.2724609375, 1030.3251953125, 732.5260620117188, 1393.6646728515625, 371.7051696777344, 212.39524841308594, 553.6203002929688, 264.2630920410156, 870.7672729492188, 1029.33349609375, 28.87637710571289, 790.6437377929688, 1008.448974609375, 15.917953491210938, 269.49725341796875, 74.80215454101562, -10.003925323486328, 278.635498046875, 1060.22900390625, 98.7801742553711, 1305.4801025390625, 404.73846435546875, -553.938720703125, 235.10531616210938, -438.37823486328125, 1172.770751953125, 1019.6643676757812, 261.0015563964844, 707.3126220703125, 176.77365112304688, 797.1849365234375, 189.5277099609375, 180.5040740966797], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000568.npy"}
{"epoch": 0.8340675477239354, "step": 569, "batch_size": 64, "mean": 383.6678466796875, "std": 685.9461059570312, "min": -2004.428955078125, "p10": -279.7725723266601, "median": 328.0401153564453, "p90": 1290.5408447265625, "max": 1860.9090576171875, "pos_frac": 0.78125, "sample": [-579.9497680664062, 121.05471801757812, 1688.833984375, 184.36488342285156, 500.962158203125, -240.6219940185547, 686.9581909179688, 49.362457275390625, -630.4683837890625, 183.06556701660156, 933.42724609375, 1278.121826171875, 532.8480834960938, 515.1686401367188, 370.54754638671875, -554.0150146484375, 247.6731719970703, 865.5631713867188, -296.5513916015625, -110.62044525146484, -1101.760986328125, 1627.2420654296875, 3.2937278747558594, -235.3837890625, 13.324007034301758, 670.5264892578125, 1123.919677734375, 44.14469909667969, 46.043975830078125, -9.153446197509766, 1419.4351806640625, 819.086181640625, 669.0359497070312, 13.5782470703125, 1366.9158935546875, 442.4355773925781, 102.94879150390625, 314.7312316894531, 723.5601806640625, -2004.428955078125, 595.4638671875, 171.5930633544922, -122.38043212890625, 1096.2947998046875, 874.442138671875, 130.8618621826172, 878.7623901367188, 233.06594848632812, -403.2534484863281, 798.000732421875, -234.24415588378906, 1781.5009765625, 21.34276580810547, 573.2080688476562, 1144.4722900390625, -171.8897705078125, 125.38336181640625, 533.6669921875, 341.3489990234375, 46.87176513671875, 499.9236145019531, 688.313232421875, 1295.86328125, 1860.9090576171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000569.npy"}
{"epoch": 0.8355359765051396, "step": 570, "batch_size": 64, "mean": 363.5556945800781, "std": 580.7849731445312, "min": -927.0133056640625, "p10": -283.93259582519534, "median": 282.31427001953125, "p90": 1184.9621215820316, "max": 2452.930908203125, "pos_frac": 0.765625, "sample": [272.08001708984375, 1028.521728515625, 351.08258056640625, 1534.7625732421875, 2452.930908203125, 127.39295959472656, 764.52734375, 1261.63818359375, 1402.5172119140625, 215.35765075683594, 365.7454528808594, 902.1280517578125, 364.7898254394531, 491.0184326171875, 187.59498596191406, 376.1029052734375, 890.1917114257812, 414.6009521484375, 683.3091430664062, -196.1070556640625, 746.5300903320312, -279.8688659667969, -554.5004272460938, -722.0946655273438, 115.9502944946289, 870.6114501953125, -523.2811279296875, -184.93734741210938, 292.54852294921875, 390.59320068359375, 481.2111511230469, 307.76580810546875, -285.6741943359375, -441.03668212890625, 221.40411376953125, 244.4962921142578, 186.10317993164062, 474.90106201171875, 118.39531707763672, 1284.1973876953125, 460.7836608886719, -164.4989013671875, 234.00936889648438, 151.43328857421875, 1389.927001953125, 313.63330078125, 164.92799377441406, 780.6896362304688, 1084.18408203125, 695.70849609375, -36.63472366333008, 162.54257202148438, 600.6318359375, 391.60174560546875, -87.79352569580078, -927.0133056640625, 172.183837890625, 161.5575714111328, 191.5790557861328, -129.11618041992188, -285.9170227050781, 1228.1527099609375, 232.86605834960938, -151.37416076660156], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000570.npy"}
{"epoch": 0.8370044052863436, "step": 571, "batch_size": 64, "mean": 435.5760498046875, "std": 662.4684448242188, "min": -1222.17919921875, "p10": -399.49790649414064, "median": 416.9233093261719, "p90": 1303.3551391601566, "max": 2161.9296875, "pos_frac": 0.78125, "sample": [552.4705810546875, 888.1236572265625, 110.36116027832031, 1110.454345703125, -238.58078002929688, 494.8226013183594, 76.02578735351562, 1062.918212890625, 406.55133056640625, 1078.7601318359375, 1769.14599609375, -401.2730407714844, -263.3415832519531, 1666.2044677734375, 154.16648864746094, 597.009033203125, 819.994384765625, 818.6127319335938, 507.1629333496094, 821.8419189453125, 2161.9296875, 1717.6617431640625, 681.9692993164062, 1198.29248046875, -395.3559265136719, -47.9088134765625, 611.9637451171875, 332.77825927734375, 1367.4976806640625, 1092.210693359375, 248.4477996826172, -79.01409912109375, 455.6252746582031, 1637.69970703125, 805.8544921875, 58.21955490112305, -493.656494140625, 150.10916137695312, 1333.2137451171875, 521.759521484375, 865.7572631835938, -451.6420593261719, 1233.68505859375, 465.73577880859375, 374.8992919921875, 122.91400909423828, 543.0575561523438, 153.1671905517578, 575.1951904296875, 17.439743041992188, -533.2862548828125, -1222.17919921875, 253.3221435546875, 22.3212890625, -820.1624145507812, 349.51361083984375, 250.8382110595703, 427.2952880859375, -658.4320068359375, 101.79179382324219, -147.1181640625, 561.0719604492188, -79.76986694335938, 80.7247314453125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000571.npy"}
{"epoch": 0.8384728340675477, "step": 572, "batch_size": 64, "mean": 439.9151611328125, "std": 734.1136474609375, "min": -1301.7716064453125, "p10": -350.6680145263672, "median": 367.4749298095703, "p90": 1373.3884765625, "max": 2892.6845703125, "pos_frac": 0.71875, "sample": [-359.1648864746094, -191.74159240722656, -388.47161865234375, 554.7188720703125, 1927.9991455078125, 48.46514129638672, 106.20925903320312, 34.143226623535156, 1.7640113830566406, 155.61404418945312, -105.41580963134766, 1363.6248779296875, 825.4156494140625, 367.680908203125, 622.379638671875, 1013.9234008789062, 1047.6224365234375, 1148.534912109375, 1449.35205078125, 623.08251953125, 265.85809326171875, 755.5030517578125, -34.3353157043457, 1140.9381103515625, -911.6019287109375, 77.93143463134766, 280.2332763671875, -314.5941162109375, -291.9783020019531, 910.7142944335938, -592.3656005859375, 510.9992980957031, 471.993408203125, -64.09628295898438, 314.2417907714844, 2892.6845703125, 106.28545379638672, 421.8863220214844, -290.239990234375, -532.9539184570312, 1377.5728759765625, 35.71601486206055, -330.84197998046875, 575.3521728515625, -199.9166259765625, -435.4553527832031, 1134.583984375, 816.3139038085938, 712.7142944335938, -13.487632751464844, 870.94775390625, 278.84124755859375, 388.2872619628906, 1743.7701416015625, 240.03414916992188, 367.2689514160156, 597.1885986328125, 1970.284912109375, -2.1690673828125, -1301.7716064453125, 578.2936401367188, 1292.5888671875, 1475.7213134765625, 619.8916625976562], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000572.npy"}
{"epoch": 0.8399412628487518, "step": 573, "batch_size": 64, "mean": 382.56097412109375, "std": 555.9449462890625, "min": -598.6322021484375, "p10": -160.26342468261717, "median": 267.90770721435547, "p90": 944.9222961425781, "max": 2468.8544921875, "pos_frac": 0.78125, "sample": [666.7615356445312, -167.38034057617188, 29.01485252380371, -26.35736083984375, -11.468368530273438, 2468.8544921875, 877.7413940429688, -113.51781463623047, -7.763908386230469, 43.96318054199219, 422.5116271972656, 227.33485412597656, -583.6204223632812, 2141.67529296875, 446.2882080078125, 392.30938720703125, 1704.69970703125, -598.6322021484375, -7.197441101074219, -143.65728759765625, -444.0408935546875, 991.5247192382812, 640.884033203125, 511.1104736328125, 92.44507598876953, 293.25079345703125, 802.550048828125, 1001.6159057617188, 146.65597534179688, 1200.6025390625, -275.6739501953125, 950.8094482421875, 197.2460174560547, 502.795166015625, 711.3653564453125, -332.9452819824219, 328.42266845703125, 101.24005126953125, 732.824951171875, 26.16193389892578, 863.909912109375, 125.1526870727539, 280.34271240234375, -63.38605499267578, 362.825439453125, 48.387847900390625, 215.74127197265625, 255.4727020263672, 807.5934448242188, -196.445068359375, 344.4111328125, 203.11856079101562, 225.01495361328125, 693.9863891601562, 442.46380615234375, 127.23967742919922, 8.168312072753906, 678.57177734375, 686.259765625, 618.8739013671875, 46.634765625, 714.9166259765625, 931.1856079101562, 123.05769348144531], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000573.npy"}
{"epoch": 0.8414096916299559, "step": 574, "batch_size": 64, "mean": 414.46636962890625, "std": 579.702392578125, "min": -1008.83154296875, "p10": -283.70758361816405, "median": 342.71519470214844, "p90": 1211.1511352539067, "max": 1948.170166015625, "pos_frac": 0.796875, "sample": [292.6399230957031, 669.927734375, -167.81820678710938, -201.252685546875, 1123.19287109375, 62.334800720214844, 102.6695556640625, 1248.8475341796875, 1019.1744995117188, 900.1201782226562, 33.77513885498047, 1552.7442626953125, 1024.328125, 697.3538818359375, 192.30929565429688, 553.5706787109375, -14.816350936889648, 433.879638671875, -572.066162109375, 1948.170166015625, 533.3759765625, 681.5972290039062, 346.60894775390625, 579.6881103515625, 186.9267578125, 706.5046997070312, 325.0105285644531, -411.42706298828125, 55.38645553588867, -270.49151611328125, 168.01559448242188, -451.4580993652344, 251.71682739257812, 1547.0526123046875, -67.29877471923828, 767.5449829101562, -459.9369812011719, 832.30517578125, 236.4117889404297, 470.0611267089844, 890.9224243164062, -404.02197265625, 716.0050048828125, 743.438232421875, 660.963623046875, 272.1803283691406, 477.56072998046875, 122.65666198730469, -1008.83154296875, -289.3716125488281, 338.8214416503906, 78.53878021240234, 0.7153797149658203, 1380.904052734375, -44.772369384765625, 743.720703125, 352.4537658691406, 1332.765625, 206.1430206298828, 551.9862670898438, 128.92591857910156, 1712.1688232421875, 633.2411499023438, 2.053964614868164], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000574.npy"}
{"epoch": 0.8428781204111601, "step": 575, "batch_size": 64, "mean": 393.9288330078125, "std": 672.4812622070312, "min": -1319.5179443359375, "p10": -307.53247985839835, "median": 355.8787841796875, "p90": 1308.9828979492188, "max": 2153.208251953125, "pos_frac": 0.75, "sample": [1344.196533203125, 238.49375915527344, 906.0393676757812, 371.25103759765625, 562.7958984375, -859.5111083984375, 534.9510498046875, 556.1072998046875, 216.95486450195312, 148.38563537597656, 97.20410919189453, -1.3304214477539062, 224.70993041992188, 1000.3493041992188, 631.3920288085938, 296.3179016113281, -137.91360473632812, 726.55126953125, 312.9599304199219, 261.3895263671875, 340.50653076171875, -67.2412109375, 1315.0655517578125, 594.379638671875, 531.164306640625, 336.5789489746094, -1319.5179443359375, 263.8135070800781, -780.0403442382812, 542.54736328125, -836.8477783203125, 1294.7900390625, 896.2706298828125, 1439.4132080078125, 227.3106689453125, -180.4781951904297, -230.13836669921875, -103.18352508544922, 1319.6912841796875, 1026.177734375, -128.8214111328125, 118.24420928955078, -340.7013854980469, 964.6324462890625, 1115.420166015625, -833.3006591796875, 513.8238525390625, 617.2354125976562, 806.2467041015625, 2153.208251953125, 490.11663818359375, 909.5908203125, 459.3634338378906, -39.1480712890625, 1412.7952880859375, 164.36285400390625, 653.5018310546875, -1164.7840576171875, 1955.706298828125, 106.24971008300781, -68.87883758544922, 561.435546875, 149.8948516845703, 593.693359375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000575.npy"}
{"epoch": 0.8443465491923642, "step": 576, "batch_size": 64, "mean": 303.0343017578125, "std": 560.4271850585938, "min": -863.7189331054688, "p10": -351.181362915039, "median": 198.02108001708984, "p90": 1021.2107299804691, "max": 1591.7664794921875, "pos_frac": 0.734375, "sample": [180.11537170410156, 943.2913818359375, -150.06361389160156, 47.538089752197266, -312.42999267578125, -124.95741271972656, -93.43156433105469, 1569.9908447265625, 1076.9503173828125, 103.93134307861328, 291.9742736816406, 1537.13623046875, 70.01300048828125, 137.8018341064453, 738.9974365234375, 393.411865234375, 312.4179382324219, -793.38623046875, 277.5110778808594, 640.7920532226562, 523.37158203125, 1058.6282958984375, 712.1688232421875, 906.2587890625, -4.606952667236328, 1364.15087890625, -78.27275085449219, -376.69586181640625, -212.77554321289062, 755.169921875, -189.5904998779297, 213.82215881347656, 50.069305419921875, 37.23622131347656, 840.2078247070312, 776.2454833984375, 1591.7664794921875, 283.14422607421875, 150.9407958984375, 669.074462890625, 58.51355743408203, 890.3965454101562, 182.22000122070312, 222.44581604003906, 115.05464935302734, -832.7715454101562, -59.554481506347656, 552.9698486328125, -619.2593994140625, 697.673828125, 380.2603454589844, 1054.604736328125, 894.5653686523438, 933.7698364257812, -863.7189331054688, -367.7890930175781, 136.39129638671875, 247.95770263671875, 274.6910095214844, 167.403076171875, -458.8309326171875, 17.34708023071289, -273.19927978515625, 125.13394927978516], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000576.npy"}
{"epoch": 0.8458149779735683, "step": 577, "batch_size": 64, "mean": 396.7024841308594, "std": 604.195068359375, "min": -1233.643798828125, "p10": -79.43249740600585, "median": 329.4523162841797, "p90": 1271.3557495117188, "max": 2357.2216796875, "pos_frac": 0.828125, "sample": [-1233.643798828125, 2357.2216796875, 1230.7554931640625, 1413.5137939453125, 115.34522247314453, 29.926010131835938, 290.5289611816406, 118.26600646972656, 206.43524169921875, -61.429107666015625, 678.6346435546875, 395.4079284667969, 456.53546142578125, 425.84417724609375, 431.24609375, 676.9755249023438, 920.3428955078125, 1288.755859375, -509.48712158203125, 571.9420166015625, 157.7433624267578, 138.62844848632812, -737.0584716796875, 211.122802734375, 293.630859375, 779.1624755859375, 103.18962097167969, 261.77117919921875, 449.67803955078125, 360.33587646484375, 171.05551147460938, 431.0515441894531, 213.50772094726562, -74.82219696044922, 1809.99609375, 52.38470458984375, 318.3232421875, -575.0548095703125, 209.8834991455078, -323.8284912109375, 386.7584533691406, 174.98159790039062, 292.4984130859375, 472.35931396484375, -25.07019805908203, 507.116943359375, 36.67857360839844, 618.7959594726562, 685.7353515625, 1015.3856201171875, 39.55357360839844, 340.5813903808594, 1326.522216796875, -81.40834045410156, 1698.3939208984375, 600.5145874023438, 418.174072265625, -619.8992919921875, -24.623397827148438, 1603.987548828125, 111.13572692871094, 892.80517578125, 416.8846130371094, 447.278564453125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000577.npy"}
{"epoch": 0.8472834067547724, "step": 578, "batch_size": 64, "mean": 424.1553955078125, "std": 524.048583984375, "min": -737.1434326171875, "p10": -142.70306701660152, "median": 401.450927734375, "p90": 1082.5272460937504, "max": 1803.978515625, "pos_frac": 0.796875, "sample": [1803.978515625, 135.75390625, 723.64404296875, 115.139892578125, 522.746337890625, 374.90789794921875, 1248.1251220703125, 528.60009765625, 197.70138549804688, -159.4956512451172, 73.82301330566406, 828.0692749023438, 461.525146484375, 251.49942016601562, 96.63455200195312, -569.1472778320312, 175.34190368652344, 42.03893280029297, 1693.0386962890625, 802.3818969726562, 759.5964965820312, 440.33953857421875, 957.8065795898438, -103.52037048339844, -101.71537017822266, 817.8845825195312, 818.5450439453125, -88.32657623291016, -281.9050598144531, 109.65128326416016, -18.683364868164062, -376.9739990234375, 437.4083251953125, 411.6408386230469, 124.58790588378906, 545.9135131835938, -206.4039306640625, 247.4776611328125, -737.1434326171875, -65.76024627685547, 459.3974914550781, 653.5142211914062, 479.33319091796875, 391.2610168457031, 1284.6041259765625, 906.9175415039062, 267.14837646484375, 37.64159393310547, 664.41796875, 176.63316345214844, 812.03369140625, 718.3486328125, 78.39155578613281, 1117.9525146484375, 876.1383056640625, 1692.3304443359375, -26.117950439453125, 224.60496520996094, 1290.2432861328125, 999.8682861328125, 540.5609130859375, -189.36907958984375, 56.07850646972656, 597.2855224609375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000578.npy"}
{"epoch": 0.8487518355359766, "step": 579, "batch_size": 64, "mean": 257.15802001953125, "std": 568.225830078125, "min": -1570.766357421875, "p10": -396.1558868408202, "median": 301.6773376464844, "p90": 913.920300292969, "max": 2100.681640625, "pos_frac": 0.71875, "sample": [-49.510597229003906, -212.95408630371094, -1570.766357421875, 523.63720703125, -71.93185424804688, 125.00119018554688, 774.4871215820312, 327.0140686035156, 713.1082153320312, 328.9036865234375, -37.35871124267578, -137.6631317138672, -637.4344482421875, 329.028076171875, 230.15646362304688, 566.8123779296875, 309.33648681640625, 514.6137084960938, 1138.9398193359375, 93.63584899902344, 944.3734130859375, 92.600341796875, 358.6816101074219, 507.41680908203125, 45.89397430419922, 392.4423522949219, 384.15594482421875, 113.43909454345703, 551.2900390625, 242.5086212158203, 518.489501953125, -451.38775634765625, 505.4342041015625, 613.0040283203125, 2100.681640625, 1429.4482421875, 402.8001403808594, -521.0348510742188, -44.2790412902832, 473.9424133300781, 973.9993896484375, -267.2815246582031, 746.1163330078125, -205.80618286132812, 1238.7498779296875, -10.597673416137695, 35.11924362182617, -200.92164611816406, -541.3231201171875, 294.0181884765625, 842.863037109375, 409.5830383300781, 45.18211364746094, 614.8538208007812, 391.036865234375, -811.5826416015625, 1005.80322265625, 11.138181686401367, 247.1397247314453, -1025.6077880859375, 161.17271423339844, -45.90918731689453, 276.4844665527344, 356.9278564453125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000579.npy"}
{"epoch": 0.8502202643171806, "step": 580, "batch_size": 64, "mean": 488.9371643066406, "std": 663.317626953125, "min": -1163.26904296875, "p10": -72.87680892944334, "median": 327.0852813720703, "p90": 1216.588562011719, "max": 2879.746826171875, "pos_frac": 0.8125, "sample": [-123.29435729980469, 914.817626953125, 235.1970672607422, 526.025390625, -109.9725341796875, 194.21844482421875, 187.16395568847656, 1847.54541015625, 401.1153564453125, 258.0159606933594, 506.90692138671875, -6.1962890625, 2879.746826171875, 225.82327270507812, 67.48516845703125, 625.9908447265625, 338.70556640625, 133.0979461669922, 834.13525390625, 466.43341064453125, 262.85357666015625, 1097.7432861328125, 1113.8616943359375, 598.6656494140625, 528.53271484375, 77.07657623291016, 1051.056884765625, 1166.702880859375, 498.2254638671875, 147.7994384765625, -50.177894592285156, 153.83729553222656, 111.93428802490234, 119.60200500488281, 804.8426513671875, 1237.9681396484375, 69.83737182617188, -48.303504943847656, 901.7520751953125, 682.5345458984375, 1716.442138671875, 47.27537536621094, -1163.26904296875, 1092.5166015625, 44.23656463623047, 692.6646118164062, -310.4468078613281, -110.04296875, 541.9517211914062, 2750.44580078125, 18.085039138793945, -53.76324462890625, 1395.57666015625, 192.87435913085938, 339.965087890625, 438.7976379394531, -81.0683364868164, 331.5905456542969, -16.986730575561523, -316.9166564941406, 322.58001708984375, 930.9804077148438, 1253.049560546875, 306.1348876953125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000580.npy"}
{"epoch": 0.8516886930983847, "step": 581, "batch_size": 64, "mean": 348.05023193359375, "std": 600.3658447265625, "min": -863.3380737304688, "p10": -401.9725830078125, "median": 366.0399169921875, "p90": 984.2407653808594, "max": 2973.618408203125, "pos_frac": 0.765625, "sample": [173.00009155273438, 110.02723693847656, 163.06289672851562, 13.734756469726562, 402.2337646484375, 433.1258850097656, -455.26025390625, 273.84552001953125, 884.8435668945312, -612.1334228515625, 846.9747924804688, 819.3069458007812, 471.32611083984375, 340.38885498046875, 607.7791748046875, -334.6645812988281, 1404.8994140625, 720.5829467773438, 450.41375732421875, 1177.2562255859375, -216.46006774902344, 678.739501953125, -51.65692138671875, 502.75396728515625, 563.7017211914062, 527.1976318359375, 276.93328857421875, 146.7926025390625, 90.0999755859375, 995.49609375, -269.69384765625, 2973.618408203125, -2.4026832580566406, 1217.091064453125, 262.21099853515625, 415.715576171875, 957.9783325195312, 87.93598937988281, 45.684913635253906, 428.20391845703125, 407.49896240234375, 48.304046630859375, -4.35662841796875, -276.28228759765625, 1062.720458984375, 249.41551208496094, 653.5369262695312, 782.8892822265625, 1356.4091796875, 596.8214111328125, -602.3377685546875, -863.3380737304688, 52.560787200927734, -569.310546875, 621.6514282226562, 424.83599853515625, -364.5892333984375, -505.15582275390625, 46.27362060546875, -417.9940185546875, 297.3704833984375, 847.3320922851562, 391.69097900390625, 518.585205078125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000581.npy"}
{"epoch": 0.8531571218795888, "step": 582, "batch_size": 64, "mean": 372.801513671875, "std": 523.6493530273438, "min": -1092.8670654296875, "p10": -182.0078338623047, "median": 315.8799133300781, "p90": 1029.2774902343751, "max": 1529.4989013671875, "pos_frac": 0.78125, "sample": [-1092.8670654296875, 204.19021606445312, 1137.7506103515625, 319.8927001953125, 473.8277282714844, -174.06936645507812, 287.8996887207031, 439.5172119140625, 388.402099609375, -28.000003814697266, 368.6881103515625, 88.31548309326172, -12.164138793945312, 738.1358032226562, 672.3623657226562, 1461.2978515625, 50.44102478027344, -23.236557006835938, -266.67578125, 1061.971435546875, 446.39569091796875, 96.71437072753906, 589.4383544921875, -547.2852783203125, -185.4100341796875, 907.4332275390625, 215.57510375976562, 968.0167236328125, 1363.791015625, 1529.4989013671875, 816.0261840820312, 447.5922546386719, 1075.8284912109375, -19.167701721191406, -269.267578125, 163.8463134765625, 49.797828674316406, 102.50953674316406, 664.4588623046875, 593.1797485351562, 360.5496520996094, -112.27030944824219, 848.9072875976562, 1039.186279296875, 114.75379943847656, 966.0107421875, 49.088401794433594, 760.6123657226562, -384.541259765625, 249.8209991455078, 12.017444610595703, 859.5638427734375, 1006.156982421875, 311.86712646484375, -43.88976287841797, -970.2784423828125, 252.52340698242188, 230.76119995117188, 655.8522338867188, 639.2593383789062, 1000.2240600585938, 219.19923400878906, 579.2656860351562, 110.00455474853516], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000582.npy"}
{"epoch": 0.8546255506607929, "step": 583, "batch_size": 64, "mean": 232.5241241455078, "std": 655.5315551757812, "min": -1351.7745361328125, "p10": -612.8489440917969, "median": 207.93045043945312, "p90": 1001.9081176757817, "max": 2081.510986328125, "pos_frac": 0.71875, "sample": [68.05191040039062, 270.81939697265625, 253.77493286132812, -635.0154418945312, 1045.645751953125, 43.55873107910156, 831.2540893554688, 306.2883605957031, 2081.510986328125, 119.38213348388672, -273.46539306640625, 83.63954162597656, -978.5172119140625, 244.41392517089844, 205.54183959960938, -276.6020202636719, 898.4832153320312, 902.190673828125, 356.8945617675781, 1044.6441650390625, 470.95379638671875, 67.24605560302734, -593.6799926757812, -391.9176025390625, 0.46312904357910156, -810.8568115234375, 355.31097412109375, 219.79522705078125, -542.5469970703125, -805.5153198242188, -91.43821716308594, -141.69827270507812, 536.2015380859375, 599.0728149414062, 299.0520935058594, 1869.1824951171875, 369.6734619140625, 11.920692443847656, 878.3487548828125, 202.77700805664062, 3.0009765625, 336.0484619140625, 331.5400390625, -99.13612365722656, -1012.3181762695312, -1351.7745361328125, 1482.3427734375, 356.8577880859375, -281.9572448730469, 474.34197998046875, 883.744384765625, 1592.8502197265625, 91.88251495361328, 545.128173828125, 900.5263671875, -621.064208984375, 107.94917297363281, 497.4141845703125, -56.5596923828125, 195.52700805664062, 1238.2889404296875, -62.896636962890625, 24.65027618408203, 210.31906127929688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000583.npy"}
{"epoch": 0.856093979441997, "step": 584, "batch_size": 64, "mean": 487.7043151855469, "std": 578.5932006835938, "min": -589.1967163085938, "p10": -70.54629135131832, "median": 338.12327575683594, "p90": 1265.4558227539062, "max": 2337.3974609375, "pos_frac": 0.8125, "sample": [640.3590698242188, 232.81605529785156, -266.75634765625, 256.2706604003906, 79.1123046875, -422.48876953125, 632.3974609375, 495.41314697265625, 1068.775146484375, -4.977161407470703, 78.3387451171875, -31.95050811767578, 119.07833862304688, 195.26522827148438, 344.4802551269531, 1063.9447021484375, -87.08734130859375, 1241.689208984375, 878.5244140625, -401.9516296386719, 289.247802734375, 549.6436157226562, -19.216262817382812, 266.14013671875, 376.52313232421875, 227.06332397460938, 1249.7705078125, 187.64590454101562, 215.34396362304688, 217.65060424804688, 1270.4151611328125, 606.6630859375, 370.64605712890625, -589.1967163085938, 688.4940185546875, 305.4163818359375, 435.8984375, 447.6161193847656, 263.2466735839844, -11.913665771484375, 505.6346435546875, -288.51507568359375, 331.76629638671875, 1596.835693359375, 819.7488403320312, 331.6767578125, 1150.959228515625, 1253.884033203125, -255.65008544921875, 2337.3974609375, 1832.7762451171875, 33.18913269042969, 1509.3153076171875, 1306.69580078125, 458.03875732421875, 1560.7725830078125, 233.8777313232422, 607.9703369140625, -8.805870056152344, 328.7143859863281, 499.8785400390625, 356.31451416015625, 1153.117919921875, 99.13111877441406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000584.npy"}
{"epoch": 0.8575624082232012, "step": 585, "batch_size": 64, "mean": 430.7083435058594, "std": 705.8612670898438, "min": -1735.129150390625, "p10": -346.21917114257815, "median": 434.81971740722656, "p90": 1238.3836914062501, "max": 2622.450927734375, "pos_frac": 0.75, "sample": [178.75759887695312, -163.55564880371094, 621.0554809570312, 252.4043426513672, 438.48394775390625, 2622.450927734375, 927.1987915039062, -552.3854370117188, 372.5174865722656, 358.29644775390625, 1067.2930908203125, 43.80354309082031, 525.81103515625, 606.9444580078125, -15.557167053222656, 1472.8575439453125, 1254.698974609375, 627.2701416015625, 1448.1785888671875, 589.4763793945312, 555.9141235351562, 1200.314697265625, 326.21240234375, 431.1554870605469, 1005.3147583007812, -23.935440063476562, 744.362060546875, 1054.5491943359375, 666.4439086914062, 163.98416137695312, 37.40980529785156, 663.3859252929688, 1410.3968505859375, 90.24072265625, 1118.2723388671875, 178.99539184570312, 1037.2301025390625, -345.1689147949219, -73.68296813964844, 861.7083740234375, 395.3976745605469, -1735.129150390625, 828.0918579101562, -823.8050537109375, 948.4387817382812, 442.20159912109375, 449.1153564453125, -521.2744140625, 612.2105102539062, -244.3895263671875, -44.919525146484375, -981.3030395507812, -97.35057067871094, -346.6692810058594, 144.28346252441406, 1506.5166015625, 516.5858764648438, -590.6445922851562, 889.0349731445312, -34.9559326171875, 131.09703063964844, 139.76976013183594, 170.45655822753906, 2033.471923828125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000585.npy"}
{"epoch": 0.8590308370044053, "step": 586, "batch_size": 64, "mean": 290.20867919921875, "std": 579.80859375, "min": -1223.9683837890625, "p10": -355.0480102539062, "median": 333.33856201171875, "p90": 963.3406860351564, "max": 1552.0999755859375, "pos_frac": 0.6875, "sample": [1111.661376953125, -267.76727294921875, -261.341552734375, 589.3253173828125, 501.7196350097656, 475.6911315917969, -218.70584106445312, 269.14990234375, 320.6954040527344, 157.83352661132812, -676.0390625, 345.8655700683594, 606.0986938476562, 1485.8243408203125, -116.1296615600586, -182.9762725830078, 654.2523803710938, -763.1756591796875, 242.46237182617188, 178.1999969482422, -119.27315521240234, 37.105308532714844, 935.6759033203125, 692.2894897460938, 910.2588500976562, -98.67991638183594, 621.4205932617188, -230.202880859375, -231.57110595703125, 503.9905700683594, 688.9081420898438, -165.335205078125, 331.34564208984375, -1223.9683837890625, 795.2071533203125, 773.83203125, 1552.0999755859375, -332.61041259765625, 601.9612426757812, -121.4609603881836, 975.197021484375, -483.13372802734375, 469.2642822265625, 409.1519775390625, -916.6989135742188, -649.0885620117188, 364.1249084472656, 388.15399169921875, 112.71348571777344, 519.629638671875, -364.66412353515625, 335.33148193359375, 1017.0902099609375, 1503.703857421875, 846.4984130859375, 255.82530212402344, 378.33843994140625, 1526.094970703125, 111.95250701904297, -84.25662994384766, 191.95703125, 710.4304809570312, 429.1205749511719, 152.9810791015625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000586.npy"}
{"epoch": 0.8604992657856094, "step": 587, "batch_size": 64, "mean": 399.66754150390625, "std": 548.1289672851562, "min": -1409.2161865234375, "p10": -110.6885673522949, "median": 357.36602783203125, "p90": 1192.6612670898442, "max": 1984.3934326171875, "pos_frac": 0.765625, "sample": [537.26220703125, 1040.8800048828125, 400.391845703125, 82.8133544921875, 324.9131774902344, 589.1934204101562, 583.7156982421875, -51.98420333862305, -2.1220321655273438, 1488.4854736328125, -356.2113037109375, -85.61215209960938, 1984.3934326171875, 358.4573059082031, 223.48397827148438, 464.2221374511719, -33.53376770019531, 355.14910888671875, 373.4284973144531, 386.327392578125, -55.88417053222656, 567.5164794921875, 1352.85205078125, 19.097402572631836, 311.74951171875, 1083.94482421875, -244.2704315185547, 1011.6299438476562, 680.845703125, 122.49314880371094, -17.2481746673584, 42.04564666748047, -372.94940185546875, 102.33026123046875, 689.1275634765625, 147.05332946777344, 303.24169921875, -1409.2161865234375, 955.5336303710938, 154.1826171875, -337.68133544921875, 737.7528076171875, 164.69944763183594, 698.674072265625, -85.75407409667969, 1255.8409423828125, 675.5005493164062, -166.9995574951172, 196.0953369140625, 356.2747497558594, 415.5643310546875, 1387.9080810546875, 385.2376708984375, 596.8380126953125, 526.2149658203125, 586.6396484375, 921.1622314453125, -121.3747787475586, 261.095458984375, 9.282207489013672, 360.2473449707031, -52.35174560546875, 1460.87353515625, 1239.2540283203125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000587.npy"}
{"epoch": 0.8619676945668135, "step": 588, "batch_size": 64, "mean": 346.6921081542969, "std": 505.8620910644531, "min": -949.3806762695312, "p10": -305.88171081542964, "median": 316.44227600097656, "p90": 1063.8159301757812, "max": 1545.1651611328125, "pos_frac": 0.78125, "sample": [235.79493713378906, 567.42041015625, -346.6963806152344, 93.31353759765625, 174.78977966308594, 705.048583984375, 728.7406616210938, 19.966703414916992, 652.9100952148438, 768.46875, 82.1348876953125, -394.4302978515625, 4.4381103515625, 578.2069702148438, -275.7111511230469, -388.9774475097656, 587.8658447265625, 684.914794921875, 183.35040283203125, 276.37921142578125, 483.7606201171875, -167.49429321289062, -93.73149108886719, 1126.602294921875, 160.89100646972656, 429.51763916015625, 1300.5750732421875, -145.8260498046875, -318.81195068359375, 37.77678680419922, 340.2630920410156, 71.89262390136719, 294.2481689453125, 1068.2186279296875, 905.5488891601562, 473.3025207519531, 193.17970275878906, -186.29556274414062, 1304.9742431640625, 1053.54296875, 1176.5594482421875, 710.3062744140625, 35.466400146484375, -547.6278686523438, 591.3090209960938, 341.7835388183594, 164.7699737548828, 921.589111328125, -949.3806762695312, 342.40704345703125, -15.172893524169922, 1545.1651611328125, 1227.9033203125, 422.59259033203125, 301.15130615234375, 629.6404418945312, 698.6851806640625, 345.13848876953125, -162.5734100341797, 201.3828125, 117.94246673583984, 331.7332458496094, 938.4046630859375, -450.94390869140625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000588.npy"}
{"epoch": 0.8634361233480177, "step": 589, "batch_size": 64, "mean": 448.4921875, "std": 548.8944702148438, "min": -747.2820434570312, "p10": -207.369090270996, "median": 378.01927185058594, "p90": 1337.517028808594, "max": 1895.5445556640625, "pos_frac": 0.84375, "sample": [1440.2276611328125, 346.4601135253906, 256.751953125, 13.730571746826172, 727.4993286132812, 144.2387237548828, -104.64491271972656, 685.4693603515625, 395.40496826171875, -370.35369873046875, -59.75007629394531, 344.3561706542969, 125.06172943115234, 265.43463134765625, 552.102783203125, 1355.194580078125, 1895.5445556640625, 230.08209228515625, 540.5195922851562, 196.40292358398438, 420.17254638671875, 178.1005859375, 158.36671447753906, 1497.2763671875, 51.95672607421875, 477.21405029296875, 359.9190368652344, 682.871337890625, 1425.4522705078125, -83.2908935546875, 663.528564453125, 638.7210083007812, 885.110595703125, 546.733154296875, 1172.4720458984375, -362.6700134277344, -251.39373779296875, 100.70059967041016, 702.9638671875, 360.6335754394531, 499.7713928222656, -559.74462890625, 82.26882934570312, 741.89404296875, 564.0426025390625, 549.84033203125, 1017.66943359375, 97.94291687011719, 536.5170288085938, 1364.199462890625, 613.9005126953125, 236.17642211914062, 344.51788330078125, -747.2820434570312, 271.73162841796875, -638.1093139648438, 664.5042724609375, 1427.470947265625, 33.45086669921875, 306.3928527832031, 1296.2694091796875, 622.8782348632812, -420.8578796386719, 1193.4827880859375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000589.npy"}
{"epoch": 0.8649045521292217, "step": 590, "batch_size": 64, "mean": 429.2220153808594, "std": 554.9110717773438, "min": -528.1897583007812, "p10": -233.50315856933594, "median": 314.237548828125, "p90": 1058.6290649414063, "max": 2016.1876220703125, "pos_frac": 0.75, "sample": [563.15625, 1375.5126953125, 709.9746704101562, 227.48355102539062, 984.8846435546875, -328.5430908203125, 115.25341796875, 2016.1876220703125, 967.1895141601562, -219.90530395507812, 563.8865356445312, -239.330810546875, 326.77734375, 1348.760009765625, -201.57794189453125, -7.557445526123047, -199.17510986328125, 103.4168472290039, 707.541748046875, 230.89404296875, 133.1492919921875, -304.182373046875, -105.94796752929688, 263.34014892578125, -213.13385009765625, 210.591552734375, 720.8131103515625, -302.843017578125, 1330.69775390625, -111.34610748291016, 904.874267578125, 473.9981689453125, 28.966392517089844, 1270.0399169921875, 193.95175170898438, 410.7912902832031, 347.3590087890625, 659.6536865234375, 1042.011474609375, 541.1213989257812, 491.97430419921875, -245.11090087890625, -141.30258178710938, -528.1897583007812, 289.52850341796875, 64.21189880371094, 289.71514892578125, 877.2659301757812, 256.0509033203125, -313.40411376953125, 506.7585754394531, 1064.8355712890625, 301.69775390625, -201.578369140625, 704.136474609375, 692.4404907226562, 856.6807250976562, 1917.057861328125, 173.73550415039062, 964.21875, 1044.147216796875, 941.654541015625, 149.2779083251953, 775.6715087890625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000590.npy"}
{"epoch": 0.8663729809104258, "step": 591, "batch_size": 64, "mean": 406.7897644042969, "std": 585.5913696289062, "min": -1100.3389892578125, "p10": -357.6034423828125, "median": 390.29966735839844, "p90": 1103.7670288085938, "max": 1706.748779296875, "pos_frac": 0.828125, "sample": [10.977546691894531, 1249.5164794921875, 1270.87158203125, 629.1193237304688, 1007.7850341796875, -153.38719177246094, 51.65507507324219, -1100.3389892578125, 692.0017700195312, 1688.1943359375, 915.1590576171875, 161.54727172851562, 625.5733642578125, 395.099609375, 64.84060668945312, 1706.748779296875, 748.0211181640625, 574.6888427734375, 128.67007446289062, 249.8408203125, -61.77973937988281, 941.1995239257812, 161.77308654785156, 48.158287048339844, 613.1792602539062, 114.0154800415039, -735.3442993164062, 654.7239379882812, 570.71826171875, 383.4149169921875, -332.10565185546875, 72.7845458984375, 457.05865478515625, -642.6316528320312, 1069.820556640625, 190.0386199951172, 169.05508422851562, 251.45034790039062, 698.5786743164062, -434.8173828125, -368.53106689453125, 777.9022827148438, 109.09315490722656, 385.4997253417969, -619.30419921875, 1013.7489013671875, 888.404052734375, 43.946346282958984, 667.1871337890625, 1080.626953125, -5.600257873535156, 1113.6842041015625, 915.9450073242188, 146.25799560546875, 1590.8941650390625, 437.59228515625, 607.54541015625, 1184.716064453125, -789.6107788085938, 553.1972045898438, 155.3184051513672, 454.26318359375, 348.07586669921875, 237.81759643554688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000591.npy"}
{"epoch": 0.8678414096916299, "step": 592, "batch_size": 64, "mean": 352.15185546875, "std": 712.6041870117188, "min": -1364.759521484375, "p10": -422.026220703125, "median": 370.2566833496094, "p90": 1244.6780273437503, "max": 2030.1614990234375, "pos_frac": 0.75, "sample": [272.997802734375, 1176.3782958984375, -23.245027542114258, 393.61956787109375, -850.1976318359375, 100.93434143066406, 400.4864196777344, 979.20361328125, 1808.1171875, 1376.4842529296875, -1199.27392578125, 2030.1614990234375, 213.77493286132812, 775.6575927734375, -869.8707275390625, -443.0632629394531, -110.89055633544922, 357.2375793457031, 74.00787353515625, 1378.1568603515625, 108.62567901611328, 988.4542236328125, 9.457069396972656, 1106.5225830078125, 383.2757873535156, 129.38851928710938, 211.07064819335938, 988.7637329101562, 1537.015380859375, 531.0276489257812, 11.07088851928711, -295.033935546875, 200.36312866210938, 761.1513671875, 985.5550537109375, 278.693359375, -366.33612060546875, 348.9189453125, -372.9397888183594, 596.9500732421875, 454.32086181640625, -235.42538452148438, 793.1500854492188, 1471.305419921875, 276.843994140625, -288.8841552734375, 461.4559631347656, -1364.759521484375, 750.6800537109375, 430.9423522949219, 675.26318359375, 134.08021545410156, 752.560791015625, 397.3638610839844, -926.4569702148438, 929.3507690429688, 661.6641235351562, 145.2733154296875, 1273.9493408203125, -277.0844421386719, -1355.45166015625, 801.8412475585938, 667.4889526367188, -74.42644500732422], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000592.npy"}
{"epoch": 0.869309838472834, "step": 593, "batch_size": 64, "mean": 284.3349304199219, "std": 626.8619995117188, "min": -1015.6651611328125, "p10": -508.52869873046865, "median": 200.90247344970703, "p90": 1195.3561035156251, "max": 1980.654052734375, "pos_frac": 0.625, "sample": [-22.876785278320312, 649.179931640625, -268.735107421875, 104.24371337890625, -590.0712890625, -555.6907348632812, 1398.048095703125, -675.5804443359375, 1187.5224609375, 568.37109375, -158.5516357421875, 1399.8328857421875, 1275.9791259765625, 1032.2987060546875, -10.577705383300781, -787.5634765625, 1272.0489501953125, 1475.9678955078125, 715.4308471679688, 181.0355987548828, -398.48394775390625, 109.687255859375, 320.969482421875, 80.5669937133789, 72.06321716308594, 1980.654052734375, 428.7017822265625, 91.18421173095703, 462.67108154296875, -822.1104125976562, -2.0736865997314453, 446.3542785644531, -339.40875244140625, 220.76934814453125, 682.444580078125, 871.7598266601562, -136.08169555664062, 1198.71337890625, -226.59600830078125, 450.16552734375, -821.2310791015625, 628.142822265625, -291.0818176269531, -1015.6651611328125, -22.960721969604492, -188.4630126953125, 9.346887588500977, -25.83960723876953, 493.3962097167969, 851.206298828125, 541.4717407226562, 353.7696228027344, 824.4996337890625, 430.7080383300781, 300.76702880859375, 929.4208984375, -21.477630615234375, 51.71900177001953, -231.48236083984375, 675.8582153320312, -93.90740966796875, 520.9825439453125, -126.05003356933594, 742.0427856445312], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000593.npy"}
{"epoch": 0.8707782672540382, "step": 594, "batch_size": 64, "mean": 346.91796875, "std": 591.6943969726562, "min": -910.0576171875, "p10": -328.08949279785156, "median": 289.6381530761719, "p90": 1062.7650634765628, "max": 1721.0462646484375, "pos_frac": 0.75, "sample": [824.3631591796875, 25.129364013671875, 174.4501495361328, 22.617799758911133, -54.41033935546875, -543.6659545898438, 558.82763671875, 381.3304748535156, 118.64508819580078, 684.5842895507812, 510.1492919921875, 350.2914123535156, 1721.0462646484375, 734.415771484375, 140.4917755126953, 880.251953125, 806.8761596679688, 191.9325714111328, 987.5048217773438, -563.0233154296875, -558.3982543945312, 1402.4390869140625, -783.7799072265625, 1671.89111328125, -305.31866455078125, 527.7740478515625, -164.52761840820312, 410.7684020996094, 399.3426818847656, -199.12753295898438, 40.37733459472656, 469.68011474609375, 1334.8291015625, 547.907958984375, 258.94744873046875, 595.37548828125, 561.9275512695312, 30.645397186279297, 215.68832397460938, -744.4110107421875, -31.46432876586914, 920.0380249023438, 203.53518676757812, -279.5361633300781, 1081.7095947265625, 911.44091796875, 857.6799926757812, 563.9530029296875, -337.8484191894531, 463.02056884765625, -910.0576171875, 41.289031982421875, 209.68426513671875, -252.89697265625, 320.328857421875, -44.02705764770508, 1018.5611572265625, 889.9493408203125, 1485.1435546875, 218.6476287841797, 1353.859130859375, -287.947509765625, 87.69342041015625, 56.15575408935547], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000594.npy"}
{"epoch": 0.8722466960352423, "step": 595, "batch_size": 64, "mean": 393.2123718261719, "std": 564.040283203125, "min": -551.8638916015625, "p10": -183.07081756591793, "median": 280.3418731689453, "p90": 979.8756286621094, "max": 2794.038330078125, "pos_frac": 0.78125, "sample": [1003.6913452148438, 337.031494140625, 547.5350341796875, 60.647422790527344, -29.703529357910156, 985.3334350585938, 322.03631591796875, 931.7864379882812, -337.5068359375, 395.04229736328125, 1406.02294921875, 87.59883117675781, 967.1407470703125, 799.7708740234375, -155.9169464111328, 374.3399353027344, 291.1026306152344, 1839.38671875, 140.5664825439453, -126.40029907226562, 58.939876556396484, -269.158447265625, 666.2073364257812, 395.8619689941406, -214.53680419921875, -267.5651550292969, 198.39608764648438, 90.37152862548828, 259.37213134765625, 279.74053955078125, 145.93020629882812, 268.96636962890625, 276.85589599609375, 242.17291259765625, 894.4375, -13.77021598815918, 280.9432067871094, 491.4339294433594, 1447.77734375, 312.3970947265625, -29.725601196289062, -551.8638916015625, 422.8024597167969, -319.1599426269531, 414.1701354980469, 357.6510009765625, 239.7666473388672, -119.09031677246094, 661.5128173828125, 675.99609375, 22.42723846435547, 802.5661010742188, 139.47259521484375, -59.43916320800781, 471.97503662109375, 128.28582763671875, 857.9826049804688, 96.80876159667969, -194.70819091796875, 85.93849182128906, 1640.0511474609375, 479.14520263671875, 764.7110595703125, 2794.038330078125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000595.npy"}
{"epoch": 0.8737151248164464, "step": 596, "batch_size": 64, "mean": 283.9310607910156, "std": 601.953125, "min": -1041.293212890625, "p10": -416.37568054199215, "median": 167.72147369384766, "p90": 1176.4016845703131, "max": 2037.9498291015625, "pos_frac": 0.734375, "sample": [-369.0400390625, 20.13433837890625, 641.0311889648438, 56.40669250488281, 885.9652709960938, 589.7276000976562, 222.13204956054688, 479.37725830078125, 86.25497436523438, 708.1944580078125, 2037.9498291015625, -133.43508911132812, -527.25439453125, 438.2816162109375, 995.2498168945312, 1432.31689453125, 43.18269348144531, 135.20118713378906, 506.47747802734375, 626.36474609375, -121.24436950683594, 537.8862915039062, 14.812664031982422, 146.40614318847656, -523.5265502929688, 228.24586486816406, 473.4637756347656, 401.1651916503906, 305.204345703125, 141.16322326660156, -342.58917236328125, 116.3881607055664, -436.6623840332031, 1460.3619384765625, 101.13313293457031, 1031.037353515625, 1469.0419921875, 496.8835754394531, -193.38330078125, -39.066123962402344, 217.37994384765625, 532.0202026367188, 135.7621307373047, 313.9006652832031, -172.9019775390625, 166.88780212402344, 1238.70068359375, 169.75999450683594, -296.2065124511719, -76.28639221191406, 1710.94580078125, 68.51859283447266, -1041.293212890625, 168.55514526367188, 297.4559326171875, -753.9627075195312, 150.65419006347656, -669.6757202148438, 677.5015258789062, 100.62448120117188, 1439.015625, -308.7245788574219, 446.7005310058594, -484.98406982421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000596.npy"}
{"epoch": 0.8751835535976505, "step": 597, "batch_size": 64, "mean": 489.86785888671875, "std": 668.2301025390625, "min": -1221.9337158203125, "p10": -104.57267608642576, "median": 378.38226318359375, "p90": 1397.880895996094, "max": 2325.133544921875, "pos_frac": 0.78125, "sample": [1070.5977783203125, 535.0877075195312, 1142.72998046875, 396.9789733886719, 514.6136474609375, 1513.4573974609375, -186.84048461914062, 984.0713500976562, 1487.4898681640625, -346.47418212890625, -83.1274642944336, -1.66015625, 147.72329711914062, 63.48824691772461, 716.381103515625, 107.5402603149414, 1061.567138671875, 176.96510314941406, 341.8307189941406, 1606.3509521484375, 482.4515686035156, 460.45306396484375, -2.086944580078125, 1076.628173828125, 312.14617919921875, 306.0074157714844, 36.31037521362305, 170.7506866455078, 183.5233612060547, -863.1123657226562, -1221.9337158203125, -40.20367431640625, 1071.5960693359375, 450.12115478515625, 514.61767578125, 2261.9716796875, 1345.7706298828125, -54.88250732421875, -386.3813171386719, 83.14757537841797, 715.978271484375, 1201.3853759765625, 116.0661392211914, 247.08314514160156, 359.7855529785156, -81.7718276977539, 517.6838989257812, 895.8682250976562, 48.20195007324219, 1420.2138671875, 196.58990478515625, -320.92950439453125, -113.76348114013672, 2325.133544921875, 1824.7982177734375, 69.29203796386719, 537.4002075195312, 481.90081787109375, 693.9197998046875, 737.8074951171875, 81.37321472167969, 1295.5262451171875, -72.09810638427734, 738.4297485351562], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000597.npy"}
{"epoch": 0.8766519823788547, "step": 598, "batch_size": 64, "mean": 386.431640625, "std": 560.8739013671875, "min": -992.3546142578125, "p10": -143.65835647583003, "median": 396.28114318847656, "p90": 1107.7695190429688, "max": 2048.70703125, "pos_frac": 0.765625, "sample": [397.0833740234375, 631.6123046875, 319.4011535644531, 924.137451171875, -992.3546142578125, 450.2193908691406, 1207.784423828125, 571.415771484375, 173.6280517578125, 1196.3680419921875, -31.202239990234375, -827.0374755859375, -949.8992919921875, 219.15467834472656, 1241.1949462890625, -170.66905212402344, 719.5513305664062, 659.9340209960938, 1215.5799560546875, 1686.608642578125, 251.39520263671875, 1120.352783203125, 703.27001953125, -347.669921875, 319.56182861328125, 469.7701721191406, 586.4970703125, 377.09259033203125, -900.6793212890625, 117.27021026611328, 664.4764404296875, 465.5716247558594, 434.43743896484375, -113.39954376220703, -156.6264190673828, 496.833251953125, 466.4033508300781, 1063.0, 200.0641326904297, 284.92962646484375, 165.02157592773438, 196.25201416015625, -9.947044372558594, 481.0402526855469, -55.91392517089844, 681.0528564453125, 555.39892578125, 395.4789123535156, 41.97163391113281, 119.38578033447266, 347.8451843261719, 2048.70703125, 765.6903076171875, 695.4893798828125, 526.3829956054688, 314.4129333496094, 35.2442626953125, 1078.4085693359375, 573.880126953125, -17.573028564453125, -19.670913696289062, -77.24353790283203, 764.9981689453125, -19.747600555419922], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000598.npy"}
{"epoch": 0.8781204111600588, "step": 599, "batch_size": 64, "mean": 328.69097900390625, "std": 482.2472229003906, "min": -1817.329833984375, "p10": -99.11668701171875, "median": 323.3199768066406, "p90": 895.4695495605469, "max": 1456.500732421875, "pos_frac": 0.859375, "sample": [349.3523864746094, 1003.468505859375, 329.492919921875, 405.024658203125, 702.7737426757812, 170.81622314453125, 1456.500732421875, 245.6680908203125, 175.68975830078125, 483.7452392578125, 437.9534912109375, -654.3497314453125, 5.427770614624023, 81.454345703125, -6.358728408813477, 280.5968017578125, 872.2386474609375, 896.8621215820312, 505.8360290527344, 1164.6395263671875, 197.35699462890625, 587.7674560546875, 793.69189453125, 426.4210510253906, 285.7049560546875, 78.66121673583984, 95.93726348876953, 330.54046630859375, 362.4635009765625, -1817.329833984375, 577.8917846679688, 298.05535888671875, 468.7254638671875, 247.65769958496094, 971.6592407226562, 636.9452514648438, 562.6510009765625, -99.1173095703125, 1205.1494140625, 178.92807006835938, 373.19439697265625, 129.92880249023438, 842.1038818359375, -172.42430114746094, 145.21922302246094, 469.6125793457031, 575.273681640625, 210.31301879882812, 346.73419189453125, 136.59664916992188, 892.22021484375, 228.60157775878906, 28.26099395751953, 13.957298278808594, -477.5549011230469, 91.79864501953125, 274.6298522949219, 509.07672119140625, -371.72125244140625, 1131.095947265625, -99.115234375, -269.54327392578125, 317.14703369140625, 414.2231750488281], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000599.npy"}
{"epoch": 0.8795888399412628, "step": 600, "batch_size": 64, "mean": 305.56884765625, "std": 603.0245361328125, "min": -963.5167236328125, "p10": -480.2861175537109, "median": 219.2530517578125, "p90": 1044.9012451171877, "max": 2456.21533203125, "pos_frac": 0.75, "sample": [144.1728973388672, 674.765380859375, 451.55255126953125, -963.5167236328125, 834.730224609375, 492.98321533203125, -50.762752532958984, 515.7357788085938, -538.6146850585938, 500.1909484863281, 67.7164535522461, 101.55789184570312, 902.939697265625, -124.4364242553711, 247.16622924804688, 656.43017578125, -724.2009887695312, 1286.4979248046875, 95.45833587646484, 281.3988952636719, 772.475830078125, 863.9265747070312, -218.5265655517578, 272.4508361816406, -82.72618103027344, -350.17828369140625, 634.9163208007812, 1101.3551025390625, 1194.3858642578125, 805.2044677734375, 191.33987426757812, 23.396392822265625, 905.6529541015625, 128.40321350097656, 1061.950927734375, 574.412109375, 35.91343688964844, 251.12489318847656, -518.3023071289062, 1425.046142578125, 2456.21533203125, 256.7718505859375, -158.99119567871094, 90.40673065185547, 1304.8466796875, 165.4923095703125, 804.6112060546875, 102.36073303222656, -731.4067993164062, 510.0544128417969, 41.96589660644531, 1005.11865234375, 273.64874267578125, -52.24280548095703, 81.90965270996094, 763.1243286132812, -507.4152526855469, 112.98483276367188, 28.109365463256836, -416.98480224609375, -816.0549926757812, 405.634765625, -170.3511962890625, 82.61071014404297], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000600.npy"}
{"epoch": 0.8810572687224669, "step": 601, "batch_size": 64, "mean": 383.0289306640625, "std": 528.87548828125, "min": -1155.958984375, "p10": -202.57438659667966, "median": 341.6397705078125, "p90": 1048.0877075195312, "max": 1625.0374755859375, "pos_frac": 0.796875, "sample": [712.837646484375, 1263.8692626953125, 1625.0374755859375, 693.0880737304688, -81.23882293701172, 969.3348999023438, 347.8813171386719, 117.92662811279297, -179.20498657226562, 244.25836181640625, -212.58984375, -283.24725341796875, 1283.827880859375, 388.53143310546875, 300.2755126953125, 610.3071899414062, 306.6895751953125, 582.7131958007812, -159.91842651367188, 1257.846435546875, 831.6282348632812, 576.1323852539062, 781.5294799804688, 15.928607940673828, 525.3814086914062, 212.2100067138672, -67.234619140625, 518.6911010742188, 819.430908203125, 360.9542541503906, 335.3982238769531, 254.00108337402344, -764.0771484375, 375.49273681640625, -469.2313537597656, 608.4603881835938, 140.42681884765625, -81.49119567871094, 155.47793579101562, 265.61895751953125, 663.2274169921875, 715.685302734375, 109.6686019897461, -1155.958984375, 1056.0386962890625, -631.0536499023438, 44.275115966796875, 1244.4312744140625, 594.74560546875, -292.57379150390625, 882.1384887695312, 937.984619140625, 75.65037536621094, 285.3011779785156, 82.80659484863281, 927.3884887695312, 262.4934997558594, 1352.3424072265625, 1029.535400390625, 373.54669189453125, -99.34378814697266, 187.19386291503906, 456.3369140625, 229.03631591796875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000601.npy"}
{"epoch": 0.882525697503671, "step": 602, "batch_size": 64, "mean": 478.56640625, "std": 526.2677001953125, "min": -467.8316345214844, "p10": -80.38416290283203, "median": 392.7051239013672, "p90": 1013.7599426269531, "max": 2155.007080078125, "pos_frac": 0.84375, "sample": [399.40716552734375, 119.95104217529297, -156.29859924316406, 232.67828369140625, 40.990699768066406, 162.5235595703125, 923.7711791992188, 732.19287109375, 228.02284240722656, 232.9863739013672, 470.83349609375, 197.80323791503906, 582.3394165039062, 886.141357421875, 1210.7974853515625, 43.960960388183594, 583.4480590820312, 386.70166015625, 398.7085876464844, 1078.407958984375, -19.999298095703125, 200.98196411132812, -34.985809326171875, -467.8316345214844, 1018.0293579101562, 701.2066650390625, 232.54730224609375, 456.9419860839844, 36.89324951171875, -91.81776428222656, 101.52400970458984, 1003.7979736328125, 777.8375244140625, 174.89027404785156, 694.1370849609375, 1178.26025390625, 225.08834838867188, 530.7445068359375, -435.1067199707031, -353.83697509765625, 928.373779296875, 957.5638427734375, 657.1627197265625, 994.413818359375, 776.7959594726562, 790.2396240234375, 2155.007080078125, -73.85920715332031, 979.141845703125, -447.52313232421875, 903.005859375, 879.9723510742188, 246.97332763671875, 135.3795166015625, 1240.812255859375, -83.18057250976562, 280.61163330078125, 788.3374633789062, 133.19119262695312, 103.86007690429688, 870.4019165039062, 2140.956787109375, 382.9586486816406, 202.98240661621094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000602.npy"}
{"epoch": 0.8839941262848752, "step": 603, "batch_size": 64, "mean": 447.5286865234375, "std": 663.6201782226562, "min": -805.4019165039062, "p10": -344.97072448730466, "median": 334.7553253173828, "p90": 1269.3144897460938, "max": 3015.86083984375, "pos_frac": 0.75, "sample": [-374.6356201171875, -291.1618957519531, 335.90948486328125, 41.829315185546875, 1229.11669921875, -23.4130859375, 1103.6005859375, 311.12744140625, 753.3182373046875, 1163.0498046875, -58.77947998046875, 24.760398864746094, -521.3556518554688, 774.96435546875, 51.18337631225586, 427.8830871582031, 52.351470947265625, 1324.0335693359375, 1819.955322265625, 493.8359680175781, 1070.986572265625, 812.8308715820312, 698.8082275390625, 58.02461624145508, 382.0985107421875, 264.6047058105469, 473.90423583984375, 150.05557250976562, -71.23843383789062, 258.95867919921875, 1424.183837890625, -805.4019165039062, 1466.4736328125, -62.552642822265625, -8.44061279296875, 329.01605224609375, 1059.689453125, -303.14971923828125, 687.3032836914062, 443.2000732421875, 892.4161987304688, -383.0338134765625, -98.81986999511719, 5.136135101318359, -153.15359497070312, 595.4938354492188, 731.7777709960938, 153.8345489501953, -362.8940124511719, 377.01123046875, 453.512939453125, 1286.5421142578125, 845.9415283203125, -406.57440185546875, 172.1243896484375, -416.23822021484375, 3015.86083984375, 333.6011657714844, 1647.1983642578125, 36.9493408203125, 916.1904296875, 1058.8603515625, 945.3821411132812, 27.786380767822266], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000603.npy"}
{"epoch": 0.8854625550660793, "step": 604, "batch_size": 64, "mean": 490.1846008300781, "std": 676.8681030273438, "min": -950.667724609375, "p10": -291.1084899902344, "median": 442.0929412841797, "p90": 1516.035327148438, "max": 2398.3408203125, "pos_frac": 0.75, "sample": [-400.9635925292969, 175.39004516601562, -298.40386962890625, -140.4561004638672, 599.6773681640625, 1104.9278564453125, 989.4361572265625, 1121.5740966796875, 731.4451904296875, 2398.3408203125, -950.667724609375, 949.5057983398438, 165.81825256347656, -236.37918090820312, 913.8319702148438, 1053.82177734375, 761.6710205078125, 193.1007537841797, 374.3015441894531, 1118.4808349609375, 163.99114990234375, -532.7097778320312, 100.96568298339844, 1574.7789306640625, 528.6783447265625, -479.8798522949219, 1602.4537353515625, -210.76515197753906, 112.8310317993164, 254.0115966796875, -183.51553344726562, 395.15960693359375, 441.61553955078125, 538.5382690429688, -211.21658325195312, 1685.6927490234375, 711.51904296875, 10.106910705566406, -274.0859375, 754.658203125, 627.7387084960938, 438.39361572265625, 491.2789001464844, 274.5645751953125, 494.8149719238281, 2160.0546875, -403.3907470703125, 1640.39208984375, 469.8708190917969, 3.9445343017578125, 611.7817993164062, 1108.6976318359375, 830.4671630859375, -63.62434768676758, 1742.144775390625, 442.5703430175781, -380.19915771484375, -13.347808837890625, 360.1267395019531, 860.812744140625, -165.28961181640625, 513.1273803710938, 1378.9669189453125, 340.6377868652344], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000604.npy"}
{"epoch": 0.8869309838472834, "step": 605, "batch_size": 64, "mean": 360.89971923828125, "std": 603.2444458007812, "min": -1074.45556640625, "p10": -215.3929290771484, "median": 317.7875518798828, "p90": 1027.9372314453126, "max": 2600.91796875, "pos_frac": 0.71875, "sample": [-68.72642517089844, -341.677490234375, -25.58770751953125, 228.16323852539062, 98.7501449584961, 37.21467590332031, 801.5249633789062, 638.785888671875, 846.9390869140625, 215.09521484375, 554.6439208984375, -82.39602661132812, 71.4910888671875, -181.31011962890625, 399.6641540527344, 434.96142578125, 232.73411560058594, 851.0917358398438, 1550.1201171875, -224.4691925048828, 2600.91796875, -4.867828369140625, 626.5114135742188, -22.274808883666992, 263.62310791015625, 838.0206298828125, -162.92718505859375, -274.9391174316406, 426.65509033203125, 341.0960693359375, 18.78124237060547, -194.21498107910156, 1001.9445190429688, -574.461669921875, -164.06060791015625, 435.5944519042969, 1106.8031005859375, 91.65243530273438, 219.97039794921875, 775.7730712890625, 76.92720031738281, 2138.3447265625, -769.4005737304688, -1074.45556640625, 482.03680419921875, 445.02264404296875, -275.1156005859375, 1037.07763671875, 114.24671173095703, -11.78083610534668, 500.9663391113281, 450.834228515625, 1227.7279052734375, 341.3364562988281, 294.4790344238281, 406.73956298828125, 375.27874755859375, 1131.0233154296875, 460.352783203125, 141.7095489501953, 1006.609619140625, 394.8073425292969, -85.22698211669922, 901.43115234375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000605.npy"}
{"epoch": 0.8883994126284875, "step": 606, "batch_size": 64, "mean": 399.333740234375, "std": 562.7421264648438, "min": -573.9136962890625, "p10": -309.26658325195314, "median": 292.3020324707031, "p90": 1068.522045898438, "max": 2122.014892578125, "pos_frac": 0.78125, "sample": [-573.9136962890625, 70.88827514648438, 67.98714447021484, -368.7491760253906, 377.2002868652344, 972.5806884765625, 82.02969360351562, 200.14013671875, 1236.5504150390625, 817.6437377929688, 165.7743682861328, 2122.014892578125, 826.5824584960938, 9.263153076171875, -450.5874328613281, 288.61468505859375, 418.6143493652344, 295.9893798828125, 381.7743225097656, 1226.55224609375, 882.2015380859375, -346.27044677734375, 1499.283935546875, 19.33892059326172, 961.2943115234375, 569.7220458984375, 509.9228515625, 240.74635314941406, -97.44876098632812, -293.69342041015625, 595.7937622070312, 763.1283569335938, -315.9407958984375, 256.9923400878906, -475.3755798339844, 301.21893310546875, 206.37033081054688, 685.6535034179688, -36.832252502441406, -89.32656860351562, -533.2717895507812, -255.86297607421875, 2030.185546875, 1109.6397705078125, 256.7090759277344, 664.029296875, 101.85968780517578, 653.8143310546875, 797.0443725585938, 299.71148681640625, 35.13880920410156, 838.2285766601562, -71.54560852050781, 210.83692932128906, 884.9464721679688, 1404.7862548828125, 212.32252502441406, 706.9121704101562, -2.6101150512695312, 263.43646240234375, 599.5194091796875, 524.3527221679688, 537.6337890625, 285.8146057128906], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000606.npy"}
{"epoch": 0.8898678414096917, "step": 607, "batch_size": 64, "mean": 325.91705322265625, "std": 569.0267944335938, "min": -1228.8736572265625, "p10": -217.0216522216797, "median": 289.64088439941406, "p90": 973.8943298339844, "max": 2487.81787109375, "pos_frac": 0.796875, "sample": [514.623779296875, -165.21754455566406, 458.18231201171875, 176.23484802246094, 129.87982177734375, 169.9364776611328, 994.8074951171875, 426.7342529296875, 4.52459716796875, 1330.9451904296875, -250.0286865234375, 675.0140380859375, 384.4783630371094, -222.96420288085938, 126.75192260742188, 986.840087890625, -738.0369262695312, 201.62030029296875, 378.74200439453125, 72.18750762939453, -203.15570068359375, 28.674760818481445, 373.4416809082031, 441.48046875, -422.8505554199219, 642.97119140625, -185.50250244140625, 1359.9654541015625, 781.4036865234375, 59.2282829284668, 626.3902587890625, 118.61060333251953, 708.8552856445312, 246.52252197265625, 479.7813720703125, 188.9798583984375, 943.6875610351562, 63.89892578125, 1474.9591064453125, 283.7197570800781, -909.8363037109375, 583.8052368164062, -166.71710205078125, 430.200439453125, 581.5099487304688, 89.30259704589844, -12.790550231933594, 141.9427490234375, -1228.8736572265625, 295.56201171875, 436.2829895019531, 172.00619506835938, -182.40042114257812, 582.6795654296875, 300.1841125488281, 65.42140197753906, 340.9017333984375, 2487.81787109375, 1477.8240966796875, 702.7576904296875, 281.4444580078125, -294.85955810546875, 563.6117553710938, 454.59539794921875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000607.npy"}
{"epoch": 0.8913362701908958, "step": 608, "batch_size": 64, "mean": 384.8547668457031, "std": 482.8326416015625, "min": -489.4152526855469, "p10": -219.11492309570306, "median": 358.09979248046875, "p90": 939.6485839843751, "max": 1746.6431884765625, "pos_frac": 0.78125, "sample": [736.9580688476562, 822.3096313476562, 343.731689453125, 444.4731750488281, -381.1116027832031, 1535.776611328125, -425.7972106933594, 110.79139709472656, -82.92616271972656, 516.1546630859375, 786.1005249023438, 34.98014831542969, 19.558250427246094, 678.5198974609375, -149.12818908691406, 146.1597900390625, -489.4152526855469, 429.0279541015625, -63.493621826171875, 319.4478454589844, 493.02813720703125, 594.5769653320312, 87.72624206542969, 770.1152954101562, -42.080833435058594, 411.13092041015625, 668.919189453125, 451.2607116699219, 949.4304809570312, -88.48744201660156, 44.150814056396484, -47.23332214355469, 257.038818359375, 337.49810791015625, 580.5393676757812, 271.1103515625, 372.4678955078125, 443.89862060546875, 1138.6383056640625, -355.20928955078125, 1746.6431884765625, 242.9517822265625, 680.30029296875, 1114.357666015625, 1077.284912109375, 142.73907470703125, 58.04814147949219, 775.582275390625, 121.70589447021484, -336.2091979980469, 585.237548828125, -43.51677703857422, 537.8988037109375, 722.3630981445312, -249.10923767089844, 916.8241577148438, -360.1081237792969, 113.98703002929688, 1552.995849609375, 853.7727661132812, 716.5308227539062, 495.81060791015625, 309.9886779785156, 183.9896240234375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000608.npy"}
{"epoch": 0.8928046989720999, "step": 609, "batch_size": 64, "mean": 485.5028076171875, "std": 755.35693359375, "min": -1112.6187744140625, "p10": -228.04282836914055, "median": 383.15830993652344, "p90": 1143.1357421875002, "max": 3671.76953125, "pos_frac": 0.78125, "sample": [89.78732299804688, 141.8111572265625, 104.24588775634766, -613.2840576171875, 215.72850036621094, 329.50885009765625, 1243.9730224609375, 384.23944091796875, 528.577880859375, -367.51751708984375, 781.7697143554688, 1758.371826171875, 100.36604309082031, 429.7498779296875, 757.182373046875, -1112.6187744140625, -81.81985473632812, 561.78515625, 814.6001586914062, 328.4738464355469, 568.330810546875, -155.68310546875, 797.1817626953125, 644.2127685546875, -259.05413818359375, 575.820068359375, 1166.239990234375, 172.26107788085938, -467.9750061035156, 713.1507568359375, 1700.183349609375, 623.9449462890625, 2362.290771484375, 544.8441772460938, 2785.808349609375, 920.62548828125, -1.6824417114257812, 739.4331665039062, -127.44384002685547, 290.2101745605469, 299.1079406738281, 117.99617004394531, 781.021728515625, 3671.76953125, 664.14794921875, 77.86724853515625, 585.7060546875, -52.91774368286133, -620.6641235351562, 146.84573364257812, 264.23779296875, 269.5987854003906, 915.7091674804688, 16.990615844726562, -4.466819763183594, 868.97314453125, 382.0771789550781, 599.1124267578125, 660.6239624023438, 1089.225830078125, 200.8995819091797, 595.1607055664062, -363.5562744140625, -80.91728210449219], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000609.npy"}
{"epoch": 0.8942731277533039, "step": 610, "batch_size": 64, "mean": 318.61224365234375, "std": 540.3758544921875, "min": -1535.0325927734375, "p10": -377.4345520019531, "median": 270.8283233642578, "p90": 919.6821411132815, "max": 1767.1646728515625, "pos_frac": 0.765625, "sample": [669.0428466796875, 734.0537109375, 124.79847717285156, 234.1102752685547, -351.6858825683594, 603.6666870117188, 80.12867736816406, 529.7855224609375, 614.0208129882812, -504.3026123046875, 96.59387969970703, 440.44757080078125, -24.366783142089844, 1311.34130859375, 767.025390625, 337.9151611328125, 249.44174194335938, 191.46717834472656, -511.0946960449219, -1535.0325927734375, 412.70709228515625, 1767.1646728515625, 134.69268798828125, 1118.19921875, -243.297607421875, 435.2070617675781, 148.61846923828125, -431.84698486328125, 795.4955444335938, 751.2398681640625, 1212.3544921875, -30.909103393554688, 85.03216552734375, 627.6070556640625, 242.82354736328125, 1330.85205078125, -388.4696960449219, 207.80599975585938, 261.49383544921875, 671.6835327148438, 653.9735717773438, 435.55029296875, -70.47987365722656, 647.5536499023438, 9.652679443359375, 280.1628112792969, 952.5278930664062, 110.916748046875, -83.62637329101562, 335.03533935546875, 33.95137023925781, 643.7285766601562, 221.65757751464844, -57.87424850463867, 720.3770751953125, 166.53875732421875, 300.5528869628906, 350.97467041015625, -164.0738983154297, 305.2821044921875, -450.2304382324219, -424.8601379394531, 843.0420532226562, 1465.040771484375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000610.npy"}
{"epoch": 0.895741556534508, "step": 611, "batch_size": 64, "mean": 522.70654296875, "std": 730.4154663085938, "min": -707.8971557617188, "p10": -246.48788146972655, "median": 364.44456481933594, "p90": 1441.4205078125003, "max": 2578.935791015625, "pos_frac": 0.71875, "sample": [1293.9525146484375, 369.8101501464844, 2549.9072265625, 1574.007080078125, 1166.9693603515625, 787.8436889648438, -133.8773193359375, 1339.9915771484375, 3.0063095092773438, 113.50980377197266, 295.7098083496094, 827.6943969726562, -65.86036682128906, -3.158611297607422, -707.8971557617188, -201.83155822753906, -372.09912109375, -700.2081909179688, 269.6638488769531, 1375.3194580078125, -625.152587890625, 1065.9078369140625, 1917.5531005859375, -542.825927734375, 1383.6214599609375, 639.1834716796875, -18.130882263183594, 989.1234741210938, 598.5598754882812, 204.7584686279297, -92.06536865234375, 1120.9664306640625, 339.026123046875, -74.74720001220703, 334.1129150390625, 1089.2979736328125, 50.59336853027344, 218.26089477539062, -224.04598999023438, 337.3736267089844, 908.8681640625, 359.0789794921875, 50.23664855957031, 1807.1072998046875, 402.830810546875, 469.32525634765625, -265.1680603027344, 479.0268249511719, 1028.02001953125, -138.70339965820312, 556.3250732421875, -256.1058349609375, 248.03955078125, 436.9278564453125, 684.1678466796875, 434.8485107421875, 316.18341064453125, 1771.0098876953125, 450.6890869140625, 1466.1915283203125, 1229.63720703125, -8.411048889160156, 2578.935791015625, -49.66263198852539], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000611.npy"}
{"epoch": 0.8972099853157122, "step": 612, "batch_size": 64, "mean": 420.07952880859375, "std": 558.914794921875, "min": -798.2966918945312, "p10": -259.6625427246093, "median": 366.9322967529297, "p90": 1083.29169921875, "max": 2347.609619140625, "pos_frac": 0.796875, "sample": [975.2300415039062, 250.29893493652344, 324.4515686035156, -363.5426940917969, -60.255680084228516, -36.522010803222656, 995.7147827148438, 363.5189514160156, 476.3231201171875, 493.3647766113281, -621.82080078125, 690.1964721679688, 25.112030029296875, 1073.666015625, 878.4562377929688, 1023.0315551757812, 485.70587158203125, 181.64813232421875, 485.59454345703125, 1095.23583984375, 824.7667236328125, 114.28095245361328, -798.2966918945312, 1191.323974609375, 616.9685668945312, -170.88153076171875, 798.1968383789062, -297.7115478515625, 646.629638671875, 785.3189697265625, 492.1188659667969, -767.027099609375, 134.48756408691406, 404.1642761230469, 233.77874755859375, -343.21405029296875, 647.4296875, 370.34564208984375, 295.7857666015625, -108.754150390625, 159.7631378173828, 189.7486114501953, 110.10992431640625, 241.70175170898438, 2347.609619140625, -2.0446128845214844, 827.103271484375, 1174.44189453125, 1245.1466064453125, 190.2747802734375, 127.39973449707031, 731.9417114257812, 1542.7547607421875, 172.72003173828125, 97.71711730957031, -306.6510009765625, 552.462158203125, 1.6128997802734375, 1085.848876953125, -56.41221618652344, 1077.324951171875, 672.9129028320312, 819.5260620117188, 76.9581298828125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000612.npy"}
{"epoch": 0.8986784140969163, "step": 613, "batch_size": 64, "mean": 486.54583740234375, "std": 584.8431396484375, "min": -816.841796875, "p10": -123.48731689453122, "median": 472.0158996582031, "p90": 1241.7077270507816, "max": 2264.3046875, "pos_frac": 0.828125, "sample": [581.810791015625, 1280.9688720703125, 1150.098388671875, -201.50071716308594, 1118.54638671875, 1065.9744873046875, -92.87882995605469, 1101.80517578125, 859.3953857421875, 45.053016662597656, 1384.5684814453125, 82.03755950927734, 613.287841796875, 291.06298828125, 1551.703857421875, -362.80938720703125, -43.43452453613281, 487.6646728515625, 518.4563598632812, -136.60523986816406, 375.68389892578125, 9.945365905761719, 710.6232299804688, 606.4439697265625, 1326.7935791015625, 306.73321533203125, 1084.9794921875, 1095.4381103515625, 12.87176513671875, 541.8291625976562, 685.9632568359375, 94.25869750976562, -4.058160781860352, 456.36712646484375, 301.4493408203125, 1556.971923828125, 599.7498779296875, 715.1483764648438, 2264.3046875, 3.1002120971679688, -816.841796875, 1092.486572265625, 664.0872802734375, 832.3983154296875, 711.015869140625, -698.407470703125, 202.23243713378906, 1497.0235595703125, -8.424869537353516, 240.76593017578125, 725.3218994140625, -190.864501953125, -500.4143371582031, 216.50665283203125, 976.435791015625, 538.0179443359375, 179.178955078125, 416.9024963378906, 50.912376403808594, 702.28466796875, 2.0482025146484375, 76.63333892822266, 96.57869720458984, 93.25221252441406], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000613.npy"}
{"epoch": 0.9001468428781204, "step": 614, "batch_size": 64, "mean": 468.6316833496094, "std": 706.888427734375, "min": -1263.35693359375, "p10": -251.4535034179687, "median": 401.5408630371094, "p90": 1300.3713500976562, "max": 2686.19384765625, "pos_frac": 0.75, "sample": [194.6230926513672, 250.51715087890625, -655.191650390625, 925.707275390625, 19.120254516601562, 1278.825439453125, 127.00396728515625, 392.5945129394531, 529.48046875, -70.98025512695312, 1254.50390625, 397.07635498046875, 461.0641174316406, 1014.9617309570312, 297.70806884765625, -168.23016357421875, 1303.94677734375, 735.2593383789062, 266.22918701171875, -104.26100158691406, 1071.3580322265625, -17.733551025390625, 616.7008056640625, 828.4644165039062, -545.0330200195312, -102.50318145751953, 406.00537109375, -119.59780883789062, 513.8645629882812, -266.2972412109375, 618.1817016601562, 1708.787353515625, -216.818115234375, 689.7963256835938, -616.0126953125, -206.52362060546875, 623.8413696289062, -691.0418701171875, -60.35792541503906, 673.3573608398438, 1601.8377685546875, 281.1919250488281, 1643.04931640625, 353.7619934082031, 1151.244384765625, 1111.1046142578125, 1160.471435546875, 322.66912841796875, 504.27752685546875, 586.9727172851562, 9.778709411621094, 297.4705505371094, 1362.574951171875, 105.85293579101562, 548.3272705078125, 2686.19384765625, 2104.65625, 107.56500244140625, -690.9793701171875, -1263.35693359375, 1292.0286865234375, 128.71478271484375, 493.8171081542969, 734.8070678710938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000614.npy"}
{"epoch": 0.9016152716593245, "step": 615, "batch_size": 64, "mean": 395.7686767578125, "std": 630.4072875976562, "min": -1866.8206787109375, "p10": -252.87913208007805, "median": 273.7724304199219, "p90": 1105.443432617188, "max": 2207.440673828125, "pos_frac": 0.828125, "sample": [185.01083374023438, 696.0296630859375, 151.25192260742188, 1451.516845703125, 199.29835510253906, 1236.821533203125, 16.146133422851562, 68.23328399658203, 10.5655517578125, 99.9532470703125, 946.115966796875, 289.2188720703125, 273.1646728515625, -341.9329833984375, 770.6903076171875, -424.3374328613281, 856.8994140625, -1866.8206787109375, 229.84095764160156, 389.5068664550781, 274.38018798828125, 167.88058471679688, 1983.6202392578125, 147.55389404296875, 73.29393005371094, 918.966064453125, -184.65365600585938, 329.43084716796875, 529.4000854492188, 94.99295043945312, 403.29547119140625, 1032.0145263671875, 561.15380859375, 859.5330810546875, 1230.0057373046875, 32.398292541503906, 82.08512878417969, 102.32974243164062, -369.8629150390625, -174.550048828125, -603.66943359375, 1027.9718017578125, 1023.1321411132812, 78.51806640625, 1713.229248046875, 552.4487915039062, 628.3623657226562, 1136.9129638671875, 158.88204956054688, 497.52606201171875, 808.0151977539062, -282.1186218261719, 197.5608673095703, -32.374786376953125, 668.8750610351562, -376.9767761230469, 223.47772216796875, 364.16387939453125, 509.0592346191406, 850.1195678710938, -142.7108917236328, 2207.440673828125, 649.6160278320312, 141.29318237304688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000615.npy"}
{"epoch": 0.9030837004405287, "step": 616, "batch_size": 64, "mean": 322.25848388671875, "std": 565.6805419921875, "min": -782.8526000976562, "p10": -357.9785186767578, "median": 261.052978515625, "p90": 1181.9202392578125, "max": 1723.7796630859375, "pos_frac": 0.765625, "sample": [900.9625244140625, 103.85568237304688, 114.98877716064453, 623.4375610351562, 272.7436828613281, 35.353248596191406, 179.8770751953125, -211.12411499023438, 364.85760498046875, 1195.5078125, 1445.3865966796875, 585.614501953125, 154.41806030273438, 59.53398132324219, -700.6390380859375, 1186.7384033203125, 481.9596252441406, 263.1321716308594, 218.7777099609375, -97.87630462646484, 1414.2203369140625, 594.4677734375, 452.2135925292969, 392.4451599121094, -325.2826843261719, -701.1990966796875, 521.6856689453125, 199.49029541015625, -16.840721130371094, 1226.3416748046875, 1651.63134765625, 160.20518493652344, -386.44537353515625, 105.0970687866211, 134.66416931152344, -782.8526000976562, -190.72811889648438, 233.96774291992188, 835.2966918945312, 578.5272216796875, 486.61651611328125, 293.12152099609375, 363.59466552734375, 643.1687622070312, -755.0924682617188, 809.8057861328125, 179.40602111816406, -330.61138916015625, 147.8043212890625, 683.0784912109375, 307.844970703125, -353.4626770019531, 1723.7796630859375, 355.0836181640625, 715.7860107421875, 166.27517700195312, 181.30210876464844, 258.9737854003906, -293.65460205078125, 493.46722412109375, 958.3253173828125, -359.91387939453125, 1170.6778564453125, -495.2447204589844], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000616.npy"}
{"epoch": 0.9045521292217328, "step": 617, "batch_size": 64, "mean": 219.72235107421875, "std": 551.7030639648438, "min": -1229.612060546875, "p10": -405.0384399414062, "median": 218.85134887695312, "p90": 801.7510375976564, "max": 1923.3834228515625, "pos_frac": 0.703125, "sample": [215.61842346191406, 70.773681640625, -245.22540283203125, 341.29156494140625, 455.2723083496094, -344.55902099609375, -70.06879425048828, -570.2752075195312, 809.870849609375, 122.85585021972656, -327.3498229980469, 332.75286865234375, 331.9666442871094, 407.3722229003906, 201.69480895996094, 226.3477783203125, 365.6068115234375, 222.0842742919922, 40.70191955566406, 1159.8031005859375, 743.38330078125, 528.65283203125, 135.9573974609375, -7.023202896118164, 502.3429260253906, 342.7951965332031, 61.687400817871094, -100.06953430175781, 374.89794921875, -1229.612060546875, 68.29989624023438, 339.43902587890625, -508.12701416015625, -299.7267150878906, 680.6259155273438, 295.99896240234375, 1127.7685546875, 673.7410278320312, 597.332275390625, 451.2378845214844, 516.8604736328125, -430.95819091796875, 922.3154296875, 729.2837524414062, 1754.1326904296875, 287.90338134765625, 782.8048095703125, -548.2527465820312, 207.23583984375, 442.5848388671875, -1031.4256591796875, 135.6580047607422, 22.16967010498047, 201.7196044921875, 230.96749877929688, -167.75985717773438, 825.0087890625, -117.66949462890625, -81.82866668701172, -323.27508544921875, 1923.3834228515625, 213.84771728515625, -894.3140869140625, -64.29830932617188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000617.npy"}
{"epoch": 0.9060205580029369, "step": 618, "batch_size": 64, "mean": 515.9976806640625, "std": 811.6475830078125, "min": -1939.3087158203125, "p10": -411.60989379882807, "median": 429.20912170410156, "p90": 1279.8233886718751, "max": 2967.072021484375, "pos_frac": 0.78125, "sample": [-1939.3087158203125, 858.25439453125, 212.71051025390625, 310.2828674316406, 114.8992691040039, 233.8472137451172, 167.433349609375, 63.4667854309082, 2304.651123046875, 963.0352172851562, 310.670166015625, -900.4974975585938, 808.005126953125, 246.68836975097656, 1258.5758056640625, 352.4041442871094, -300.1787109375, 1286.157958984375, 665.0577392578125, -374.9568786621094, -209.0245361328125, -455.2672424316406, -358.38916015625, 44.60575866699219, -680.3925170898438, -706.292236328125, 209.53121948242188, 1265.042724609375, 920.2142333984375, 168.78756713867188, 579.8742065429688, 486.27325439453125, 662.35400390625, 1011.8363647460938, 945.5093383789062, -765.0225219726562, -427.3183288574219, 1229.6654052734375, 339.0185852050781, 434.2022399902344, 1087.1856689453125, 1293.87890625, 1196.6810302734375, 2364.89599609375, 947.1637573242188, -97.20099639892578, 424.21600341796875, 666.7779541015625, 412.63433837890625, 2967.072021484375, 843.425048828125, 101.20809173583984, 1561.040283203125, 1246.08642578125, 832.946533203125, 470.223388671875, -34.73381042480469, 1006.7635498046875, 2119.036376953125, -5.674957275390625, 412.38427734375, 1078.4556884765625, 513.4791870117188, 279.50018310546875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000618.npy"}
{"epoch": 0.9074889867841409, "step": 619, "batch_size": 64, "mean": 385.66802978515625, "std": 821.34228515625, "min": -1059.2305908203125, "p10": -543.6129394531249, "median": 217.5477294921875, "p90": 1365.6864257812504, "max": 3407.4150390625, "pos_frac": 0.71875, "sample": [-1059.2305908203125, -717.31103515625, 111.19384765625, 1186.7705078125, 1543.0196533203125, -162.64108276367188, 258.9618835449219, 73.2113265991211, 175.21690368652344, -597.505126953125, 616.154296875, 742.1392822265625, 399.8438720703125, 1885.815185546875, -733.6754150390625, 204.203857421875, 1197.32275390625, 390.9891357421875, 1108.40283203125, 1057.0496826171875, -824.8553466796875, -776.77783203125, 1291.1519775390625, 1231.25634765625, 10.851303100585938, -417.864501953125, -107.65047454833984, 667.8203735351562, 594.903564453125, -191.33551025390625, 82.20236206054688, 455.5564270019531, -301.27581787109375, 42.6021728515625, 1397.6297607421875, -334.9505920410156, 24.795408248901367, -389.79620361328125, 858.5962524414062, 659.6658325195312, -35.22640609741211, 1850.058837890625, 577.3078002929688, 587.9877319335938, 1028.574462890625, 76.46807861328125, -269.2212219238281, 2344.7705078125, 1629.7740478515625, 230.8916015625, 46.451637268066406, 200.37728881835938, 358.8197937011719, 3407.4150390625, -1036.8988037109375, -403.53704833984375, 575.2589111328125, -83.3021240234375, 1200.797607421875, 22.82241439819336, 277.7054443359375, 149.85223388671875, 15.876272201538086, 277.2732849121094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000619.npy"}
{"epoch": 0.908957415565345, "step": 620, "batch_size": 64, "mean": 366.874267578125, "std": 580.65869140625, "min": -1038.438232421875, "p10": -263.02004394531247, "median": 293.6409912109375, "p90": 1279.733337402344, "max": 1865.39599609375, "pos_frac": 0.765625, "sample": [423.8940734863281, -404.971923828125, 459.92620849609375, 624.1129760742188, 445.4852294921875, 716.5704345703125, 194.0428924560547, 237.3255615234375, -79.22257232666016, 1223.252685546875, -303.3880310058594, 224.72015380859375, -237.94937133789062, 1370.3785400390625, 502.00152587890625, 146.2471466064453, 804.98388671875, 410.9242248535156, -92.95663452148438, 1437.8446044921875, 1865.39599609375, 219.17587280273438, 420.89617919921875, 1094.33056640625, -457.8292236328125, 190.66787719726562, 557.331787109375, 247.68661499023438, 142.59718322753906, 79.99867248535156, 17.74671173095703, 1559.928955078125, 434.526123046875, 1745.07470703125, -68.40625, 524.3074340820312, 1.4265365600585938, 459.28045654296875, -81.015625, 288.235595703125, 449.4229736328125, 520.7264404296875, -127.07794189453125, -107.596923828125, 1303.9393310546875, 735.4378662109375, 702.499755859375, -1038.438232421875, 131.63345336914062, 238.98208618164062, -91.46405029296875, -927.522705078125, 1615.9677734375, 226.18765258789062, 409.3544616699219, 366.5352478027344, 430.9707336425781, 172.03720092773438, -508.06109619140625, 488.28729248046875, 260.73980712890625, 857.53125, -273.7646179199219, 299.04638671875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000620.npy"}
{"epoch": 0.9104258443465492, "step": 621, "batch_size": 64, "mean": 557.1375732421875, "std": 704.46533203125, "min": -1550.810302734375, "p10": -223.81751861572263, "median": 551.5833129882812, "p90": 1476.6358398437503, "max": 2850.878662109375, "pos_frac": 0.84375, "sample": [590.355224609375, -361.939208984375, -759.3603515625, -212.3167266845703, 1686.6768798828125, 54.853858947753906, 95.14282989501953, 734.3339233398438, 649.3168334960938, 815.6278076171875, 902.6176147460938, 338.632568359375, 272.7733154296875, 860.7322998046875, 326.07073974609375, -318.8311462402344, -278.5708312988281, 968.3643798828125, -1550.810302734375, 228.43951416015625, 1546.9178466796875, 203.05804443359375, 1028.753173828125, 524.5841064453125, 1226.7481689453125, 2850.878662109375, 200.56613159179688, 1185.423583984375, 912.1112670898438, -126.14279174804688, 1509.9278564453125, 657.9876708984375, 885.3399047851562, 246.34515380859375, 965.9945678710938, 1697.9599609375, 182.95950317382812, 673.4190673828125, 369.6988830566406, 695.443115234375, -73.72169494628906, 1372.1171875, -228.74642944335938, 578.58251953125, 235.65145874023438, 169.64556884765625, 1398.9544677734375, 321.6037902832031, 1140.8017578125, 1218.4786376953125, -940.9112548828125, 311.09271240234375, 99.3464584350586, 217.86553955078125, 251.81895446777344, 585.6239013671875, 797.3697509765625, 1561.42919921875, 282.5589904785156, 305.11395263671875, 1818.323486328125, 833.2750854492188, 705.5687255859375, 214.8802947998047], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000621.npy"}
{"epoch": 0.9118942731277533, "step": 622, "batch_size": 64, "mean": 386.73748779296875, "std": 626.3360595703125, "min": -1182.8712158203125, "p10": -223.50586853027335, "median": 369.9639129638672, "p90": 879.7986022949219, "max": 2742.759521484375, "pos_frac": 0.78125, "sample": [605.8335571289062, 359.1795654296875, -266.0307312011719, 456.85931396484375, 481.7236328125, 400.3735046386719, -567.5233764648438, 336.5343933105469, 2489.751953125, 545.4078369140625, 84.99536895751953, -323.20062255859375, 125.62513732910156, -308.043212890625, -62.167327880859375, 495.77301025390625, 160.30812072753906, 485.654541015625, 65.00436401367188, 210.95936584472656, 1121.515380859375, -124.28118896484375, 566.9329833984375, 2742.759521484375, 709.8966674804688, 52.955413818359375, 574.484619140625, 720.1943359375, 184.3565216064453, -1073.2130126953125, 240.82034301757812, -29.72271728515625, 857.455322265625, 199.45631408691406, 228.09600830078125, 886.9378662109375, 428.675537109375, -541.8493041992188, 1053.5740966796875, 561.771484375, 758.5826416015625, 863.1403198242188, 47.928199768066406, 676.5289306640625, 222.53164672851562, -39.53515625, 380.7482604980469, -61.595306396484375, 64.43345642089844, 323.2633361816406, 1438.35205078125, 438.89093017578125, 778.169189453125, 648.2846069335938, 544.2513427734375, 733.9862670898438, -1182.8712158203125, -16.00397491455078, 95.87117004394531, 1480.134765625, 627.122314453125, 545.3873291015625, -2.858074188232422, 248.62017822265625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000622.npy"}
{"epoch": 0.9133627019089574, "step": 623, "batch_size": 64, "mean": 298.5467834472656, "std": 607.906982421875, "min": -1107.687744140625, "p10": -256.56908874511714, "median": 249.71563720703125, "p90": 1036.2270751953126, "max": 2586.483642578125, "pos_frac": 0.703125, "sample": [199.67401123046875, 251.4342041015625, -806.6304321289062, 1418.453857421875, -169.0030517578125, -91.33052062988281, 573.4922485351562, -112.28659057617188, 965.655517578125, 10.71307373046875, 1358.3388671875, 273.96795654296875, 597.246337890625, 88.27549743652344, 693.4210815429688, -338.5619201660156, 674.618896484375, 398.82354736328125, -145.59735107421875, 55.5908203125, -187.8173065185547, 534.3941040039062, 1055.2061767578125, -218.46649169921875, -1107.687744140625, 191.00628662109375, 309.6741943359375, -776.9269409179688, 420.21112060546875, 593.3980712890625, -272.8987731933594, 390.5099792480469, -127.20650482177734, -57.69001007080078, -174.31564331054688, 107.93701171875, 991.9425048828125, 641.0045166015625, 832.0064086914062, 171.52664184570312, 96.04493713378906, -652.4388427734375, 346.8870544433594, 1413.818359375, -108.42611694335938, -53.84682083129883, 2586.483642578125, 495.02496337890625, -60.57798385620117, -454.2139587402344, 1056.027587890625, 520.7193603515625, 346.35552978515625, 299.80419921875, 247.9970703125, 1738.8663330078125, 53.80126953125, 155.2591552734375, 255.39752197265625, 60.4261589050293, 508.0130615234375, 268.1885986328125, 748.262451171875, 27.0166015625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000623.npy"}
{"epoch": 0.9148311306901615, "step": 624, "batch_size": 64, "mean": 462.65264892578125, "std": 559.2327880859375, "min": -1113.9554443359375, "p10": -201.21226806640624, "median": 498.3316345214844, "p90": 1149.6627807617188, "max": 1494.83349609375, "pos_frac": 0.796875, "sample": [695.5929565429688, -451.7669677734375, 634.4183349609375, 409.8582763671875, 507.76971435546875, 986.4241943359375, 1120.1622314453125, 1494.83349609375, 1359.79345703125, 235.411376953125, 246.98324584960938, -201.8826904296875, 106.0905532836914, 1134.74169921875, 795.682373046875, -141.8424072265625, 585.5537719726562, 779.7674560546875, 1216.97314453125, -853.9172973632812, 52.19477844238281, 225.0247039794922, 34.40320587158203, 828.1885375976562, 365.2716979980469, 224.09487915039062, 875.9956665039062, 848.0148315429688, 1144.682373046875, -171.19003295898438, 255.58070373535156, 1080.20458984375, -199.64794921875, 329.05810546875, -295.3098449707031, 894.06640625, 488.8935546875, -75.2662353515625, 986.4132080078125, 374.583251953125, -1113.9554443359375, 78.00293731689453, -145.004150390625, 635.1455688476562, 925.64404296875, 1006.9248657226562, 254.9917755126953, 697.276123046875, 897.5958251953125, 485.9259338378906, 64.45198059082031, -414.7490234375, 35.51347351074219, 1279.160888671875, 516.631103515625, 646.9597778320312, -222.36727905273438, 663.3698120117188, -117.45729064941406, 651.658935546875, 1256.0279541015625, 123.82988739013672, 1326.489990234375, 1151.7972412109375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000624.npy"}
{"epoch": 0.9162995594713657, "step": 625, "batch_size": 64, "mean": 505.76568603515625, "std": 518.9659423828125, "min": -307.3981018066406, "p10": 5.689547729492203, "median": 413.784912109375, "p90": 1190.024450683594, "max": 2217.088134765625, "pos_frac": 0.890625, "sample": [198.45944213867188, 41.860321044921875, -307.3981018066406, 883.6768188476562, 247.82171630859375, 61.00494384765625, 729.325439453125, 1091.7340087890625, 931.05322265625, 1013.1165771484375, 67.51187133789062, 796.6180419921875, 362.77838134765625, 45.60173034667969, 1543.532470703125, 401.11724853515625, 233.50384521484375, 61.94935989379883, 2217.088134765625, 561.8486328125, 1905.343505859375, 168.73841857910156, 79.5438003540039, 555.1248168945312, 142.72940063476562, 802.519287109375, 1579.013427734375, 20.670974731445312, 248.7589874267578, 347.3442077636719, 521.2411499023438, 574.27392578125, 520.6024169921875, -278.51739501953125, 505.5567626953125, -60.44036865234375, 87.15277862548828, 630.4476928710938, 274.9193420410156, 830.7291870117188, 1232.14892578125, 430.208740234375, 1330.8489990234375, 426.45257568359375, 260.99188232421875, 498.8402404785156, -63.29960632324219, 439.794677734375, 631.1715698242188, 710.4845581054688, 109.95978546142578, -0.7310638427734375, 775.5618896484375, 270.23602294921875, 220.33534240722656, 140.31051635742188, 155.5515899658203, 398.1114807128906, 1055.1019287109375, -301.3312072753906, 663.8139038085938, -100.86187744140625, 1379.37548828125, 1067.971923828125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000625.npy"}
{"epoch": 0.9177679882525698, "step": 626, "batch_size": 64, "mean": 287.99566650390625, "std": 633.5265502929688, "min": -2438.611572265625, "p10": -317.71809997558586, "median": 254.00843811035156, "p90": 950.3342041015627, "max": 2080.89794921875, "pos_frac": 0.71875, "sample": [559.73779296875, 40.63336181640625, 402.2249755859375, 393.0029296875, 62.663360595703125, 186.84495544433594, 736.7351684570312, 232.96310424804688, 897.0370483398438, -2438.611572265625, 1466.87353515625, 513.487060546875, 1569.7489013671875, 2080.89794921875, -350.18536376953125, 165.45480346679688, 328.1038818359375, -376.5821838378906, -2.40826416015625, -169.7783966064453, -71.91998291015625, 197.62010192871094, 420.2425537109375, 98.88562774658203, 14.96757698059082, -10.013504028320312, 155.93362426757812, -669.6876220703125, 448.8589172363281, 722.9441528320312, -87.27613830566406, 619.1085815429688, -502.6757507324219, -575.2128295898438, 526.4484252929688, 1136.8546142578125, 770.2426147460938, 394.1263122558594, 400.44000244140625, 684.1128540039062, 580.4889526367188, -241.96115112304688, 973.1758422851562, 493.18157958984375, 1780.757080078125, 556.5524291992188, -145.84747314453125, 53.520042419433594, 87.93865966796875, 366.62152099609375, -7.7523193359375, 872.685302734375, 313.6178283691406, 981.3304443359375, -17.754119873046875, 193.3082733154297, 587.9053955078125, 324.999755859375, -591.2296142578125, 22.460891723632812, 104.90690612792969, -88.0053939819336, 275.05377197265625, -17.073711395263672], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000626.npy"}
{"epoch": 0.9192364170337739, "step": 627, "batch_size": 64, "mean": 581.1422729492188, "std": 847.0567626953125, "min": -680.9352416992188, "p10": -143.391259765625, "median": 404.8114776611328, "p90": 1537.3797485351568, "max": 4179.59814453125, "pos_frac": 0.75, "sample": [573.4595336914062, 267.8970947265625, 1175.2091064453125, 380.2085876464844, -114.6450424194336, -101.47602844238281, 103.51944732666016, 614.70361328125, 563.1083984375, 442.6892395019531, 1299.2283935546875, -441.5105895996094, 195.22955322265625, 503.6576232910156, -106.71526336669922, 1078.0126953125, -143.5797119140625, 880.4188842773438, 947.2384033203125, 1128.5164794921875, -628.7388305664062, -115.72018432617188, 601.0294189453125, -22.183408737182617, 1735.4493408203125, 183.78366088867188, 429.41436767578125, 2520.21484375, 1019.6162109375, 437.19317626953125, 1822.38916015625, 76.14541625976562, 506.90960693359375, 1128.150146484375, 586.9326171875, 1614.358154296875, 1395.557373046875, 203.12850952148438, 538.138427734375, -198.02989196777344, 124.02062225341797, -59.28651428222656, 3358.490966796875, 1598.1607666015625, 756.3026123046875, -142.43533325195312, 4179.59814453125, -199.94677734375, 582.7701416015625, -142.9515380859375, 958.3693237304688, 318.7428283691406, 871.7504272460938, -680.9352416992188, 293.96380615234375, 272.77337646484375, 195.48512268066406, 211.33204650878906, -310.3215637207031, 1224.5189208984375, -135.75021362304688, 200.5014190673828, 293.2562255859375, 345.78851318359375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000627.npy"}
{"epoch": 0.920704845814978, "step": 628, "batch_size": 64, "mean": 423.3465576171875, "std": 642.8243408203125, "min": -840.4584350585938, "p10": -305.23894653320303, "median": 338.24951171875, "p90": 1287.4263061523438, "max": 2113.150634765625, "pos_frac": 0.703125, "sample": [280.3168029785156, -28.814449310302734, 387.779296875, 1381.7822265625, 1628.078857421875, 163.78848266601562, 444.8065490722656, 61.82598114013672, -139.25608825683594, -148.35723876953125, 117.6382064819336, 127.33667755126953, -840.4584350585938, 1248.7418212890625, 998.848876953125, 1198.85986328125, -51.26963806152344, -181.39642333984375, 1606.514892578125, 557.9497680664062, 2113.150634765625, 546.45263671875, 242.4031982421875, -228.584716796875, -119.17598724365234, -463.6885070800781, 535.4948120117188, 873.0282592773438, -435.45635986328125, -390.1790466308594, 879.3861694335938, 335.91748046875, 797.2194213867188, 1305.654296875, 1304.00537109375, 839.6670532226562, 403.5462951660156, 710.6425170898438, 116.99681854248047, 95.13681030273438, -120.5100326538086, 422.63427734375, 404.30877685546875, 107.27936553955078, 1608.9503173828125, 1234.5626220703125, 1184.2520751953125, 217.58673095703125, -338.09075927734375, -694.4050903320312, -141.90719604492188, -142.9263458251953, 1173.0452880859375, 837.9713134765625, -620.8699340820312, -21.305246353149414, 1150.3857421875, 340.58154296875, 559.9273681640625, 72.62298583984375, 1168.5989990234375, 467.2652587890625, -77.63282775878906, 25.520496368408203], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000628.npy"}
{"epoch": 0.922173274596182, "step": 629, "batch_size": 64, "mean": 409.25048828125, "std": 548.0084228515625, "min": -1301.7579345703125, "p10": -161.97100830078122, "median": 419.7250671386719, "p90": 1153.8191528320315, "max": 1917.30712890625, "pos_frac": 0.8125, "sample": [217.73062133789062, 1180.1715087890625, -372.30987548828125, -392.0388488769531, 309.3543701171875, 432.34210205078125, 348.26885986328125, 887.6754150390625, 1391.8272705078125, 518.9385375976562, 547.5524291992188, 796.8772583007812, -45.891517639160156, 426.85809326171875, 1281.6702880859375, -76.21088409423828, 199.0030517578125, 61.437503814697266, 11.034311294555664, 559.9202880859375, 28.081375122070312, 1196.3258056640625, 374.8966979980469, -139.79725646972656, 170.0703887939453, 432.1983947753906, -737.461181640625, -116.84558868408203, 1092.330322265625, 764.3926391601562, 109.17032623291016, 815.4909057617188, 1436.956298828125, 1917.30712890625, 518.9656372070312, 171.55198669433594, -1301.7579345703125, 1233.42626953125, 892.5294189453125, -171.4740447998047, -336.48089599609375, -466.67852783203125, 368.16473388671875, 416.46685791015625, 550.639892578125, 1058.7197265625, 276.486572265625, 356.8707275390625, 779.6809692382812, 9.036565780639648, 817.3951416015625, 268.7099609375, 40.42816162109375, 717.9677734375, 285.69976806640625, 550.271484375, 52.77887725830078, 626.9013061523438, 958.2305908203125, 482.68646240234375, 422.9832763671875, 494.029052734375, 540.4716796875, -49.996856689453125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000629.npy"}
{"epoch": 0.9236417033773862, "step": 630, "batch_size": 64, "mean": 223.78536987304688, "std": 673.7725830078125, "min": -1151.05224609375, "p10": -496.0831878662109, "median": 165.48873138427734, "p90": 946.6370544433595, "max": 3069.985595703125, "pos_frac": 0.6875, "sample": [1084.5877685546875, -1151.05224609375, -584.4738159179688, 432.1697082519531, -137.8941650390625, 680.1869506835938, 36.14373016357422, -367.69818115234375, 2431.6318359375, -178.36431884765625, 675.497314453125, -150.81105041503906, 298.1314392089844, 925.3757934570312, 731.6419677734375, -1091.4456787109375, 79.42623901367188, -580.4016723632812, 239.554931640625, 212.9717559814453, 456.6355895996094, -172.7776336669922, 144.77340698242188, 1365.6622314453125, -202.0137176513672, 3069.985595703125, 26.31852912902832, 955.7490234375, 119.59708404541016, -82.51118469238281, 31.47004508972168, -255.05975341796875, 280.94366455078125, -508.7901306152344, 298.2073059082031, -549.9249267578125, -108.45930480957031, 282.2212219238281, 160.9669952392578, 364.51593017578125, 156.7528076171875, -466.43365478515625, 260.3551940917969, -671.50732421875, 167.1943359375, 378.97174072265625, 197.0357666015625, 289.4296875, 719.435302734375, -60.329994201660156, 84.69599914550781, 1048.121337890625, 538.9103393554688, 163.7831268310547, -317.846923828125, 51.59834289550781, 286.47802734375, 186.32603454589844, 643.5607299804688, -397.31378173828125, 1248.269775390625, 188.71746826171875, 336.3441162109375, 27.026395797729492], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000630.npy"}
{"epoch": 0.9251101321585903, "step": 631, "batch_size": 64, "mean": 364.675537109375, "std": 641.3199462890625, "min": -1820.5240478515625, "p10": -355.4629669189453, "median": 298.1310729980469, "p90": 1082.320166015625, "max": 1875.9075927734375, "pos_frac": 0.71875, "sample": [1141.111572265625, 945.4434204101562, -494.4965515136719, 218.7983856201172, -327.48956298828125, 448.4649658203125, -261.2237243652344, -115.49491119384766, -2.7960548400878906, -291.2237548828125, 1680.303955078125, 1260.5509033203125, -429.7264099121094, 550.5361328125, -522.1795654296875, 197.3771209716797, 760.1468505859375, 748.0960693359375, 34.138160705566406, -367.4515686035156, 158.990478515625, -36.85404968261719, 266.06109619140625, -559.373291015625, 964.4935913085938, 58.55899429321289, 85.3152847290039, -24.492904663085938, 710.150146484375, 499.3634948730469, -1820.5240478515625, 482.6460876464844, 1875.9075927734375, 704.0736694335938, 377.532958984375, 721.0233764648438, 212.0299072265625, -166.40980529785156, -49.00737762451172, 1036.36279296875, 158.22854614257812, 559.535888671875, 1588.634521484375, 831.3667602539062, 587.578369140625, -937.6752319335938, 1083.5628662109375, 330.2010498046875, 1446.878662109375, 586.1486206054688, 102.68126678466797, 760.071533203125, -139.81756591796875, 457.7579650878906, 199.40216064453125, 146.5386962890625, 1069.425537109375, 652.8486328125, 1079.4205322265625, -158.91610717773438, 196.92074584960938, 993.1036376953125, 251.01010131835938, 825.5936279296875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000631.npy"}
{"epoch": 0.9265785609397944, "step": 632, "batch_size": 64, "mean": 386.1456298828125, "std": 608.5934448242188, "min": -1080.9925537109375, "p10": -371.7267791748047, "median": 383.0252380371094, "p90": 1058.4455810546874, "max": 2317.7626953125, "pos_frac": 0.75, "sample": [514.6013793945312, -632.988037109375, 284.74957275390625, 363.8407897949219, 628.9862060546875, -1080.9925537109375, 1048.1820068359375, 1067.8502197265625, 1301.527587890625, 923.0465698242188, 761.0769653320312, 617.5333251953125, 255.2095947265625, 965.2471923828125, 279.1551208496094, 209.83755493164062, 511.3761291503906, 402.2096862792969, 136.7400665283203, 109.7398452758789, 1468.9422607421875, 322.3309020996094, 684.06103515625, 441.350341796875, -477.4796142578125, 1053.399169921875, 688.2410888671875, -385.3918762207031, 419.1487731933594, -205.10174560546875, -1029.0216064453125, 437.0214538574219, 1059.6004638671875, 1025.0443115234375, 624.38916015625, 769.2415161132812, 1055.7508544921875, 798.6892700195312, 444.3702697753906, -294.1060485839844, 1173.162109375, -55.53534698486328, -253.1768798828125, -143.505859375, 598.5676879882812, 261.3829345703125, 652.6631469726562, 266.07464599609375, 1116.30908203125, 2317.7626953125, 943.249755859375, -39.251068115234375, -474.6663818359375, 29.69892120361328, 151.7368621826172, -339.841552734375, 354.5502624511719, 292.293701171875, -576.5769653320312, 0.93414306640625, 115.91535949707031, -158.0636749267578, -128.2859649658203, 1040.5150146484375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000632.npy"}
{"epoch": 0.9280469897209985, "step": 633, "batch_size": 64, "mean": 264.61846923828125, "std": 582.3457641601562, "min": -1209.3939208984375, "p10": -379.1044647216797, "median": 234.1685333251953, "p90": 1143.9239624023437, "max": 1436.552978515625, "pos_frac": 0.71875, "sample": [-369.1562805175781, 1249.447265625, 733.0098266601562, -782.1271362304688, 207.2340087890625, 308.779296875, -386.95806884765625, -80.36813354492188, 1143.158935546875, 199.80264282226562, 207.454345703125, 181.7880859375, 374.13006591796875, -1209.3939208984375, 374.09356689453125, 132.86854553222656, 190.67575073242188, 1144.9979248046875, 380.8972473144531, -290.0813903808594, -137.61187744140625, 381.9411315917969, 308.4151611328125, -40.522308349609375, 716.1900634765625, 260.8827209472656, 29.06010627746582, -81.00811004638672, -942.636474609375, 1040.4576416015625, 295.61138916015625, 200.5477294921875, 1144.2518310546875, 469.0237121582031, 679.2550659179688, -131.1946563720703, -286.6185302734375, -1040.8980712890625, -379.12646484375, 103.72284698486328, 1356.810791015625, 438.3554382324219, 326.29046630859375, -905.407470703125, 594.4461059570312, 621.837646484375, 577.308349609375, 1436.552978515625, 870.5677490234375, 1158.4039306640625, 90.82315063476562, 325.6484680175781, -379.0531311035156, 919.60986328125, 187.8999481201172, -85.07090759277344, 202.1667022705078, 491.581298828125, 1380.369140625, 90.2701416015625, 550.008544921875, -103.04049682617188, 108.37085723876953, 380.83843994140625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000633.npy"}
{"epoch": 0.9295154185022027, "step": 634, "batch_size": 64, "mean": 251.80764770507812, "std": 566.4796752929688, "min": -893.2117919921875, "p10": -407.45900878906247, "median": 166.234619140625, "p90": 1075.439685058594, "max": 2027.7052001953125, "pos_frac": 0.671875, "sample": [253.2149200439453, 892.3450317382812, -319.7565612792969, 1539.5880126953125, -649.7142944335938, 85.4404296875, 29.877641677856445, 192.59506225585938, 274.907470703125, 1289.40869140625, -3.268983840942383, 332.9659423828125, -105.45075225830078, 393.59686279296875, 599.9617309570312, 254.93881225585938, 332.96929931640625, -322.86322021484375, -523.2259521484375, 1263.7030029296875, 1023.5201416015625, 0.0980987548828125, -110.37953186035156, 105.0577392578125, -38.708404541015625, 58.2967529296875, 139.87417602539062, 561.5508422851562, 415.9833068847656, 57.2811164855957, 2027.7052001953125, 268.39520263671875, 761.6466064453125, -741.37646484375, 1097.69091796875, 1127.0338134765625, 534.8929443359375, -429.42730712890625, -356.19964599609375, 82.32504272460938, 816.3114624023438, -88.17271423339844, -58.66944122314453, -893.2117919921875, 206.89439392089844, -578.2025756835938, 7.169414520263672, -548.5986328125, 15.520393371582031, 762.57861328125, -236.7750244140625, 761.73974609375, 369.3642578125, 1260.1781005859375, 412.5731506347656, 123.42491149902344, 244.16738891601562, 710.131591796875, -91.42723846435547, 257.4996032714844, -18.504364013671875, -9.141036987304688, -256.5030517578125, 550.8494873046875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000634.npy"}
{"epoch": 0.9309838472834068, "step": 635, "batch_size": 64, "mean": 338.82867431640625, "std": 619.06884765625, "min": -1067.425537109375, "p10": -265.4029663085937, "median": 324.0335235595703, "p90": 981.9138488769535, "max": 2469.904296875, "pos_frac": 0.703125, "sample": [786.0667114257812, 30.539566040039062, 2469.904296875, -210.6590576171875, 445.0414733886719, -98.52828216552734, 89.8448486328125, -0.5522346496582031, -10.014045715332031, 506.80718994140625, 130.70458984375, 1122.1214599609375, -2.7444915771484375, 502.07464599609375, -491.69921875, 386.0788879394531, 453.8226013183594, 46.768035888671875, 887.263427734375, 26.891616821289062, 370.8773193359375, 176.88877868652344, 754.514892578125, 884.729248046875, 361.9906005859375, 93.86197662353516, 11.225418090820312, 415.60076904296875, -287.6207275390625, 1728.76025390625, -31.966135025024414, 368.67303466796875, 250.0130615234375, 1217.7008056640625, 803.8539428710938, -53.95793914794922, 94.50823211669922, -33.10292434692383, 302.5965270996094, 1015.242431640625, -130.30934143066406, 904.1471557617188, 893.6624755859375, 445.1163024902344, -779.5155029296875, -129.44955444335938, 584.487548828125, -49.692726135253906, -213.5615234375, 443.1556396484375, -452.66595458984375, -381.12725830078125, -1067.425537109375, 782.7338256835938, 1223.293212890625, 471.3497009277344, 325.7496643066406, 545.372314453125, -872.834716796875, 2157.5615234375, 138.02664184570312, 643.9815063476562, 322.3173828125, 366.53955078125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000635.npy"}
{"epoch": 0.9324522760646109, "step": 636, "batch_size": 64, "mean": 415.7841796875, "std": 641.3682861328125, "min": -1010.7550659179688, "p10": -268.7308837890625, "median": 322.6002960205078, "p90": 961.4151916503907, "max": 3480.59375, "pos_frac": 0.75, "sample": [46.775230407714844, -412.41473388671875, 92.68679809570312, -285.309326171875, 328.1139221191406, -364.6426696777344, 326.5240478515625, 425.80731201171875, 359.8688049316406, 109.4240951538086, 855.8350830078125, -367.5479736328125, 370.7919921875, 1012.5965576171875, -2.1928482055664062, 943.0071411132812, -110.760498046875, 2092.857666015625, 1142.993896484375, 502.80975341796875, 276.9775390625, 945.0881958007812, -155.55117797851562, -279.6273193359375, 690.134521484375, 318.6765441894531, 941.9708862304688, 401.18975830078125, 250.7563934326172, 3480.59375, -10.524398803710938, -147.3321075439453, 916.4328002929688, 1385.154052734375, -263.5116882324219, 968.4124755859375, 996.069580078125, 358.3956298828125, 527.9961547851562, -215.01992797851562, 804.05810546875, 178.1745147705078, 812.9508666992188, 895.6898803710938, 664.64306640625, 317.30340576171875, 794.212158203125, 90.33401489257812, 305.48095703125, -270.9676818847656, 181.88568115234375, 21.947357177734375, -1010.7550659179688, 302.4720458984375, 265.1864929199219, -224.0675811767578, 895.1693725585938, 765.5283203125, 248.90184020996094, 479.77166748046875, -4.7506866455078125, 284.6216125488281, 542.2623291015625, 816.6293334960938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000636.npy"}
{"epoch": 0.933920704845815, "step": 637, "batch_size": 64, "mean": 444.0483093261719, "std": 614.4473876953125, "min": -1050.9857177734375, "p10": -96.04797897338867, "median": 288.65606689453125, "p90": 1395.5146240234376, "max": 2178.625244140625, "pos_frac": 0.828125, "sample": [125.7018051147461, 36.741172790527344, 410.7004699707031, 85.36791229248047, 35.549625396728516, 828.810791015625, 248.03028869628906, 15.256227493286133, 626.3727416992188, 1130.3367919921875, 669.2135620117188, -1050.9857177734375, 140.13807678222656, 175.40505981445312, -324.6028747558594, 608.6926879882812, 865.9061889648438, 1494.0653076171875, 973.530029296875, 581.1483154296875, 1396.059326171875, 286.34954833984375, 140.24069213867188, 2105.000732421875, 109.63088989257812, 508.32940673828125, -296.5205993652344, -10.51187515258789, 149.42916870117188, -37.311737060546875, 469.7880554199219, -94.07089233398438, -189.465576171875, 590.0497436523438, 388.9362487792969, -100.84463500976562, 321.65203857421875, 708.077880859375, 822.9095458984375, 94.80523681640625, 556.7758178710938, 290.96258544921875, 257.5871887207031, 1096.85107421875, 883.0328369140625, 506.8922119140625, 1402.858154296875, 55.82160186767578, 1711.5950927734375, -782.6514282226562, 1118.25390625, 118.41880798339844, 323.24505615234375, 1551.368408203125, 365.8575134277344, 159.03843688964844, 116.69349670410156, 1394.24365234375, -69.55632019042969, 2178.625244140625, 14.499629974365234, 114.95185852050781, 112.7105712890625, -96.89530181884766], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000637.npy"}
{"epoch": 0.9353891336270191, "step": 638, "batch_size": 64, "mean": 357.00714111328125, "std": 683.345703125, "min": -3126.4248046875, "p10": -89.99835968017577, "median": 253.84050750732422, "p90": 1154.417370605469, "max": 1750.17724609375, "pos_frac": 0.796875, "sample": [1299.80126953125, -902.8610229492188, -3126.4248046875, 968.6329345703125, 316.7476806640625, 152.82801818847656, 59.97137451171875, 7.2100830078125, 1194.0718994140625, 363.74444580078125, 696.63916015625, 1383.6446533203125, 8.474594116210938, 746.8889770507812, 259.9776611328125, 610.9608154296875, 708.0608520507812, -485.3460388183594, 497.71221923828125, -9.190776824951172, 375.6063537597656, 46.200218200683594, -238.42608642578125, 107.5264663696289, 247.70335388183594, 955.9848022460938, 1594.041015625, 4.279399871826172, 633.582763671875, 178.03475952148438, 673.3015747070312, 240.88858032226562, 655.7650756835938, -71.5621566772461, -108.90299987792969, 694.43408203125, 1031.349365234375, -56.46957015991211, 295.8965759277344, -158.55039978027344, 31.714250564575195, 444.422607421875, 183.34286499023438, 1750.17724609375, -61.21678924560547, -18.176172256469727, 720.1818237304688, 103.29847717285156, 459.53839111328125, 981.6534423828125, -58.429962158203125, 10.831357955932617, 1014.4381103515625, 457.40625, 106.56222534179688, 101.29318237304688, 62.466957092285156, 96.56070709228516, 822.2855834960938, 1167.5172119140625, -97.89958953857422, 51.49492645263672, 1542.9146728515625, 1123.85107421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000638.npy"}
{"epoch": 0.9368575624082232, "step": 639, "batch_size": 64, "mean": 141.9115753173828, "std": 660.4837646484375, "min": -1207.6038818359375, "p10": -732.3835815429687, "median": 159.23845672607422, "p90": 931.4235656738285, "max": 1727.93359375, "pos_frac": 0.59375, "sample": [1343.859375, -974.8079833984375, -413.21649169921875, 970.9295043945312, 796.2462158203125, -370.56207275390625, 234.77481079101562, 319.95318603515625, -660.3021850585938, 709.6121215820312, 1463.429931640625, -42.840152740478516, 503.9859313964844, 539.9767456054688, -948.1013793945312, 268.5617370605469, -440.968017578125, -386.59661865234375, -26.537872314453125, 438.1958923339844, 663.3878173828125, -1207.6038818359375, 797.3207397460938, 551.5319213867188, -693.9332275390625, -1161.1700439453125, -1031.9482421875, 320.4624938964844, 117.53409576416016, 28.094680786132812, 539.746826171875, 525.4564819335938, 4.576591491699219, -613.146728515625, 1147.159912109375, 1137.230712890625, -410.20599365234375, 826.1282348632812, -352.479248046875, -49.450416564941406, 172.42965698242188, -1.032501220703125, 624.4961547851562, 385.7122802734375, 1727.93359375, -610.34619140625, -165.15701293945312, -26.987586975097656, -324.4324951171875, -302.11749267578125, 129.5265350341797, 583.2213745117188, 282.707763671875, 839.2430419921875, -748.8623046875, -305.4100646972656, -821.5596923828125, 146.04725646972656, 73.35249328613281, 720.8565063476562, 348.76544189453125, 1109.637451171875, 574.333740234375, 205.69699096679688], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000639.npy"}
{"epoch": 0.9383259911894273, "step": 640, "batch_size": 64, "mean": 545.9135131835938, "std": 725.3378295898438, "min": -912.5697631835938, "p10": -202.2556167602539, "median": 379.98231506347656, "p90": 1367.8445434570315, "max": 2828.490966796875, "pos_frac": 0.796875, "sample": [749.7704467773438, 558.8447265625, -495.64910888671875, -206.2161102294922, 106.73798370361328, 372.0340270996094, 309.1866455078125, 1380.2427978515625, 1318.62060546875, 194.93069458007812, 1393.1337890625, 1711.4764404296875, 934.9012451171875, 76.74966430664062, -261.7256164550781, 438.5193176269531, 844.353271484375, -129.7716064453125, 68.30302429199219, 1250.502685546875, -714.88818359375, 971.559814453125, 363.356201171875, 655.4181518554688, -45.03256607055664, -912.5697631835938, 66.77041625976562, 387.93060302734375, 2828.490966796875, 755.6612548828125, 587.9260864257812, 1023.4686279296875, -500.1629638671875, 614.248779296875, 2259.3125, 260.52642822265625, -193.01446533203125, 334.26922607421875, 13.54296875, 308.65966796875, -7.17082405090332, 254.01158142089844, -31.416336059570312, 1095.4371337890625, 1089.8499755859375, 185.76724243164062, 826.2510986328125, 705.3779296875, -472.23272705078125, 1248.1156005859375, 2049.14013671875, 355.8671569824219, 12.885040283203125, 1126.2027587890625, 12.223922729492188, 737.9132690429688, 389.0440673828125, 917.31103515625, 107.17170715332031, 769.8058471679688, 1338.915283203125, 2295.82861328125, -33.86705017089844, 285.6134338378906], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000640.npy"}
{"epoch": 0.9397944199706314, "step": 641, "batch_size": 64, "mean": 343.6410217285156, "std": 685.3895874023438, "min": -1198.50537109375, "p10": -530.9419464111328, "median": 327.4054260253906, "p90": 1003.9912231445313, "max": 2893.065673828125, "pos_frac": 0.78125, "sample": [205.4886474609375, 313.58978271484375, 540.2371826171875, 1008.3402709960938, -564.8958740234375, -110.57620239257812, 288.08740234375, 1337.680908203125, 497.605712890625, 169.1228790283203, -962.813720703125, 592.2972412109375, 31.285415649414062, -408.3525390625, 2893.065673828125, -666.8927612304688, -511.0711364746094, -445.988525390625, 62.45237731933594, 863.1373291015625, 58.793548583984375, 119.21134185791016, 86.593017578125, 585.1094360351562, 218.44833374023438, 471.0820617675781, -1050.0831298828125, 803.094482421875, 493.81842041015625, 117.86357116699219, 876.22216796875, 480.54510498046875, 675.0753173828125, 38.9952392578125, 45.05506134033203, 12.95211410522461, 1208.056640625, 477.68023681640625, 581.591552734375, 587.1372680664062, 189.3203125, 28.806886672973633, -41.69696807861328, -539.4580078125, 277.97125244140625, 748.7681884765625, -1198.50537109375, 993.8434448242188, 935.5580444335938, 969.0415649414062, 2100.64404296875, 787.6271362304688, -962.33544921875, 633.5177001953125, -15.431190490722656, 331.4891052246094, 625.12939453125, -208.7251434326172, 558.0436401367188, 323.3217468261719, 704.937744140625, 1156.3807373046875, 513.5878295898438, 1062.146728515625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000641.npy"}
{"epoch": 0.9412628487518355, "step": 642, "batch_size": 64, "mean": 418.36724853515625, "std": 866.006591796875, "min": -1465.794189453125, "p10": -313.23659362792966, "median": 326.18267822265625, "p90": 1175.9884521484378, "max": 4603.32177734375, "pos_frac": 0.734375, "sample": [88.16841125488281, 299.9844970703125, 680.2186279296875, 588.4754028320312, 177.40928649902344, -1465.794189453125, 837.1422729492188, 451.5261535644531, -70.17543029785156, -514.509765625, -177.8553466796875, 100.80337524414062, 1450.15673828125, -281.42529296875, 702.5763549804688, 1011.9797973632812, 426.288330078125, 833.5322265625, -877.4545288085938, 16.01424789428711, 5.879615783691406, -151.89361572265625, 456.1424560546875, 28.878278732299805, 311.104736328125, 481.76434326171875, -72.5999526977539, 204.71588134765625, 97.21414184570312, 410.9081726074219, 795.2109375, 623.9580688476562, 734.3251342773438, 2368.318359375, -287.1884460449219, 51.966583251953125, 341.2606201171875, 645.807373046875, -667.5948486328125, 385.9555358886719, 849.8955688476562, 203.74290466308594, 561.4515991210938, 1215.829345703125, -324.40008544921875, 1083.0263671875, 750.8547973632812, 1665.8341064453125, -126.44319152832031, -51.7169189453125, 4603.32177734375, 903.2254028320312, -603.6740112304688, 411.40411376953125, -191.23863220214844, -246.25927734375, 304.88330078125, 2379.296875, 97.99411010742188, -990.5845947265625, 485.68316650390625, 1631.0997314453125, 185.6981201171875, 935.3848876953125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000642.npy"}
{"epoch": 0.9427312775330396, "step": 643, "batch_size": 64, "mean": 335.9446716308594, "std": 531.7962036132812, "min": -722.9514770507812, "p10": -354.2138275146484, "median": 309.4313201904297, "p90": 977.4984008789064, "max": 1852.9453125, "pos_frac": 0.78125, "sample": [357.3330383300781, 276.5766906738281, 152.73390197753906, 791.9830322265625, 888.0668334960938, 663.73974609375, 444.1099548339844, 18.41620635986328, 508.40155029296875, 780.6109619140625, 17.397817611694336, -483.293212890625, 271.1916198730469, 339.2002258300781, -219.28602600097656, 890.3524169921875, 867.1810913085938, 1033.151611328125, 1004.951416015625, 471.86871337890625, 24.753799438476562, 40.8042106628418, -330.2365417480469, 147.38232421875, 215.80882263183594, 150.73501586914062, -722.9514770507812, 525.2132568359375, 472.10260009765625, 941.4031982421875, -364.48980712890625, 271.4820251464844, 285.984619140625, 254.7095947265625, 932.5631103515625, 1289.744873046875, 332.8780212402344, -486.269287109375, 629.0563354492188, 400.8440856933594, -467.51116943359375, 1852.1617431640625, 996.6431274414062, 341.68267822265625, 633.0476684570312, 106.52637481689453, 2.690095901489258, 392.4810791015625, 1852.9453125, 57.961177825927734, 767.33203125, 622.3848876953125, 621.3541259765625, -265.1324462890625, 992.9677734375, 41.95561981201172, -528.2684936523438, -313.1374816894531, 350.36279296875, -3.268524169921875, -227.604736328125, -134.32815551757812, -428.5085144042969, 149.5172119140625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000643.npy"}
{"epoch": 0.9441997063142438, "step": 644, "batch_size": 64, "mean": 334.51837158203125, "std": 688.5844116210938, "min": -1756.2186279296875, "p10": -475.72526855468743, "median": 193.88199615478516, "p90": 1158.7273315429688, "max": 2366.03857421875, "pos_frac": 0.75, "sample": [2366.03857421875, 26.356409072875977, 850.4503784179688, -1756.2186279296875, 445.0894775390625, 385.04705810546875, -630.6015625, 184.99122619628906, 800.91845703125, 313.3082580566406, 44.75537872314453, 48.63252258300781, 14.103096008300781, 392.41461181640625, 117.98570251464844, 553.67041015625, 202.77276611328125, -40.11370849609375, 169.33688354492188, -544.9667358398438, 159.08297729492188, 1973.2738037109375, 69.49147033691406, -640.3944702148438, 716.4071655273438, 806.1519775390625, -1.8792266845703125, 338.91156005859375, -194.40850830078125, 1488.9888916015625, -32.054161071777344, -359.5368347167969, 751.8118896484375, -65.29875183105469, 55.23051452636719, 1.8375492095947266, 1159.9822998046875, 539.31787109375, 667.7777099609375, 221.53515625, -38.125038146972656, -494.7819519042969, 525.212158203125, 72.5392074584961, 1073.113037109375, 533.1320190429688, -431.2596740722656, -286.455322265625, 147.00440979003906, 1010.328369140625, 1505.8062744140625, 824.08203125, 58.0247917175293, -903.3224487304688, 958.6747436523438, 1155.799072265625, 382.1204833984375, 1278.392822265625, 25.537322998046875, 445.3358459472656, 564.5341186523438, 1766.2183837890625, 176.71136474609375, -539.6451416015625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000644.npy"}
{"epoch": 0.9456681350954479, "step": 645, "batch_size": 64, "mean": 174.93692016601562, "std": 664.9188232421875, "min": -1189.0203857421875, "p10": -651.5273193359374, "median": 102.03114318847656, "p90": 1010.4605712890626, "max": 2200.49365234375, "pos_frac": 0.625, "sample": [364.57257080078125, 1025.089599609375, -564.12353515625, 910.1941528320312, -1189.0203857421875, -1061.4434814453125, -688.986083984375, 99.0272216796875, 436.5957946777344, -210.7734375, -65.34242248535156, -103.73162841796875, 52.703453063964844, -7.366132736206055, -94.73025512695312, -329.62896728515625, 392.9051513671875, 432.0740661621094, 874.40869140625, 1059.109375, 568.83203125, -349.95849609375, 105.03506469726562, -7.3136749267578125, 126.1390380859375, 44.02225875854492, 976.326171875, 8.362388610839844, -767.2412719726562, -1040.541748046875, 149.6415557861328, 407.644287109375, -516.0368041992188, 1267.21435546875, 791.2612915039062, -15.840423583984375, 1806.55517578125, 790.4588623046875, -1165.8067626953125, -547.2435302734375, 1329.3531494140625, 12.636688232421875, 111.72538757324219, 220.06996154785156, 448.5126953125, 492.11834716796875, 330.69732666015625, 312.3603820800781, 269.599365234375, 650.7921752929688, 17.93597412109375, -282.73516845703125, 1247.3349609375, 11.937828063964844, 490.0404052734375, 2200.49365234375, -150.64822387695312, 274.36859130859375, -857.3751831054688, -163.77981567382812, 559.7603149414062, 67.19542694091797, -54.33097457885742, -305.1434326171875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000645.npy"}
{"epoch": 0.947136563876652, "step": 646, "batch_size": 64, "mean": 296.4900207519531, "std": 567.9703979492188, "min": -930.6392211914062, "p10": -370.96967773437495, "median": 206.12411499023438, "p90": 1099.4735473632813, "max": 2053.802490234375, "pos_frac": 0.734375, "sample": [873.012939453125, 1164.201416015625, -118.57373809814453, 1426.408447265625, 550.156005859375, 183.39659118652344, 416.73919677734375, 674.4176025390625, 948.6556396484375, 24.157791137695312, 326.9737243652344, 504.1651611328125, -874.3638916015625, -37.448333740234375, 1101.2130126953125, 110.10292053222656, -446.2854919433594, 260.8465576171875, 1095.414794921875, -89.66277313232422, 420.289306640625, -128.8603515625, -218.2265625, 481.0527038574219, 40.07400131225586, 321.9450378417969, 403.85308837890625, 284.0718994140625, 3.648134231567383, 763.797607421875, 91.19427490234375, 474.5781555175781, 46.107635498046875, 187.6631622314453, -207.0300750732422, 173.68714904785156, 79.18963623046875, 1513.37060546875, -314.1204833984375, -930.6392211914062, 409.3116455078125, 562.6456298828125, 160.29151916503906, -430.14794921875, -119.54106140136719, -539.0757446289062, -39.41294860839844, 113.17125701904297, -395.3336181640625, 2053.802490234375, 1054.3555908203125, 245.37054443359375, 570.7560424804688, 62.25028991699219, 1183.562744140625, -208.3015594482422, -630.0050659179688, 55.93810272216797, 224.58506774902344, 186.55361938476562, 1109.9576416015625, 1035.34130859375, 446.69232177734375, 283.41998291015625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000646.npy"}
{"epoch": 0.9486049926578561, "step": 647, "batch_size": 64, "mean": 376.2207336425781, "std": 531.3851318359375, "min": -1082.4920654296875, "p10": -222.15016479492186, "median": 335.5552673339844, "p90": 1111.1393798828126, "max": 1858.520263671875, "pos_frac": 0.78125, "sample": [797.01806640625, 553.6935424804688, 278.0688781738281, 249.5560760498047, 292.3921813964844, 1054.2109375, 301.7947082519531, 350.2434997558594, 546.0739135742188, 493.2720947265625, 1587.3173828125, 219.23184204101562, 44.685882568359375, 1741.814697265625, 789.6732788085938, 599.9119873046875, -67.91072082519531, 1124.0301513671875, 636.8981323242188, -157.76573181152344, 220.38247680664062, 211.87652587890625, -6.43696403503418, -230.205322265625, 511.6801452636719, 325.49932861328125, -74.42527770996094, 1858.520263671875, -682.8883666992188, 774.4318237304688, -226.40167236328125, -303.263671875, 158.08743286132812, 831.544921875, 1138.66552734375, 624.0624389648438, 450.03131103515625, 345.6112060546875, -538.3338012695312, 217.91038513183594, 601.9341430664062, 102.15148162841797, 1131.7298583984375, -212.22998046875, -75.21673583984375, 375.3038330078125, 455.4980773925781, 417.9051208496094, -83.0864486694336, 1175.01171875, 166.0213623046875, 112.30840301513672, 1081.0609130859375, 232.28466796875, 22.241378784179688, 513.3733520507812, -1082.4920654296875, 728.7079467773438, -306.5954284667969, 357.554931640625, 481.3316345214844, 414.61126708984375, 262.6645202636719, 165.49252319335938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000647.npy"}
{"epoch": 0.9500734214390602, "step": 648, "batch_size": 64, "mean": 539.039794921875, "std": 727.60009765625, "min": -969.2520141601562, "p10": -171.07245330810545, "median": 476.8760986328125, "p90": 1443.6178588867194, "max": 3432.806396484375, "pos_frac": 0.765625, "sample": [566.0460205078125, 335.74066162109375, 186.47967529296875, 419.9629211425781, 257.93365478515625, 683.4212646484375, 1832.986572265625, 1577.781005859375, 568.4688720703125, 1294.83837890625, -24.143646240234375, -751.3941650390625, 16.492095947265625, 737.7606201171875, 487.30615234375, -159.93287658691406, 1242.85693359375, 544.5957641601562, 37.56498718261719, -175.8465576171875, -17.755720138549805, 481.4905090332031, -73.33917999267578, 935.4306640625, -305.2708740234375, 1507.3804931640625, 559.2330932617188, 149.24917602539062, -49.05900573730469, -17.62548065185547, 783.1175537109375, 674.4241943359375, -89.4074935913086, 3432.806396484375, 744.5298461914062, 710.5340576171875, -969.2520141601562, 466.0658874511719, 66.51556396484375, -637.733154296875, -209.09478759765625, 313.2359619140625, 1978.465576171875, 660.0855712890625, 622.91552734375, 462.76409912109375, 1254.8231201171875, -33.95361328125, 544.8204956054688, 726.9445190429688, 1103.525146484375, 375.9341125488281, 331.7322082519531, 569.967529296875, 1131.0189208984375, 79.58614349365234, 472.2616882324219, 819.86474609375, 308.3119812011719, 2058.09326171875, 1263.774658203125, 259.3426513671875, -429.4972839355469, 1803.37109375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000648.npy"}
{"epoch": 0.9515418502202643, "step": 649, "batch_size": 64, "mean": 259.4613037109375, "std": 628.78076171875, "min": -895.2249755859375, "p10": -393.88051452636716, "median": 190.2100067138672, "p90": 899.4631591796875, "max": 2993.20751953125, "pos_frac": 0.609375, "sample": [506.2728271484375, 23.39038848876953, 423.54302978515625, -103.18789672851562, 442.0480041503906, -316.3335266113281, -116.74642944335938, 126.37773895263672, 846.15771484375, 339.14874267578125, 647.8213500976562, -148.63330078125, -489.1303405761719, 70.86723327636719, -398.4268493652344, -152.48043823242188, -111.08628845214844, 430.3382873535156, -297.6980895996094, 299.3917541503906, 521.6593627929688, 154.1896209716797, -533.5263061523438, -174.083740234375, 873.69580078125, -806.2245483398438, 717.4210815429688, 886.6425170898438, 89.22169494628906, 535.5224609375, 394.82696533203125, -376.5622863769531, 20.75823974609375, -62.596519470214844, 892.4464111328125, 470.599609375, 800.26953125, 353.8970947265625, 751.3516235351562, -344.7142333984375, 1135.927734375, 345.2498474121094, -5.201236724853516, 1081.5484619140625, 781.5760498046875, -238.47842407226562, 1338.30517578125, 2993.20751953125, 372.77392578125, 413.73419189453125, -73.82682037353516, 226.2303924560547, 13.663040161132812, 890.6388549804688, -383.27239990234375, 1090.2120361328125, 1147.9700927734375, -550.3507690429688, -30.957721710205078, -57.56714630126953, -895.2249755859375, -712.5352783203125, 902.4703369140625, -366.9988098144531], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000649.npy"}
{"epoch": 0.9530102790014684, "step": 650, "batch_size": 64, "mean": 414.6180419921875, "std": 693.178955078125, "min": -1204.3514404296875, "p10": -508.6402282714844, "median": 320.5495300292969, "p90": 1263.604187011719, "max": 2408.072998046875, "pos_frac": 0.765625, "sample": [262.9559326171875, -484.03448486328125, 755.7052612304688, 820.1459350585938, 216.35806274414062, 867.720703125, 295.58868408203125, -833.4384155273438, 2210.27001953125, 581.1456298828125, 1119.3204345703125, -556.802001953125, 1288.0196533203125, 927.3838500976562, 835.4219360351562, 1206.634765625, 320.32763671875, 1615.5628662109375, 1111.9625244140625, 2408.072998046875, 767.2026977539062, 1.021860122680664, 1480.377197265625, 508.2975769042969, 265.60345458984375, 35.25812911987305, -152.32513427734375, 576.392333984375, 191.81112670898438, 424.1244812011719, 550.7769775390625, -23.817827224731445, -289.06005859375, -1004.0611572265625, -284.978515625, 312.6511535644531, -7.629360198974609, 1706.866455078125, -86.48814392089844, 84.50274658203125, 462.35345458984375, 134.37576293945312, -519.185546875, 665.49658203125, 553.4918823242188, 592.35791015625, 75.64195251464844, 917.007080078125, 146.98284912109375, 1442.7738037109375, 320.77142333984375, 246.28591918945312, 482.1634521484375, 955.700439453125, 729.0574340820312, -21.06793785095215, 346.81500244140625, -552.6463012695312, -1204.3514404296875, 247.36016845703125, 773.70849609375, 109.57074737548828, -524.794189453125, 130.83816528320312], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000650.npy"}
{"epoch": 0.9544787077826725, "step": 651, "batch_size": 64, "mean": 526.5450439453125, "std": 563.8071899414062, "min": -495.82080078125, "p10": -26.711966705322236, "median": 449.9176330566406, "p90": 1317.6204711914065, "max": 2054.37353515625, "pos_frac": 0.890625, "sample": [853.7871704101562, 547.9732055664062, 1961.435791015625, 216.13525390625, -349.206787109375, 10.25830078125, 379.1103210449219, 595.117919921875, 103.97593688964844, 156.0104522705078, 1098.4161376953125, 2054.37353515625, 263.3838806152344, -207.24468994140625, 75.345947265625, 107.0676040649414, 27.647525787353516, 429.53814697265625, -185.71395874023438, 836.3733520507812, 1567.7564697265625, 255.9849090576172, 859.941162109375, 298.2811279296875, 709.3543090820312, 55.710350036621094, -38.97430419921875, -57.823974609375, 121.71209716796875, 640.6680908203125, 174.2362518310547, 1.9001541137695312, 969.3978881835938, 318.923583984375, 1244.0745849609375, 889.4144897460938, 546.4509887695312, 42.04217529296875, 720.52197265625, 3.4803333282470703, 391.7633972167969, 1216.92724609375, -307.18072509765625, 505.353759765625, 644.2506103515625, 647.3707275390625, 58.18035888671875, 995.4894409179688, 118.27024841308594, 554.7858276367188, 211.2135467529297, 303.6048583984375, 1386.5677490234375, 1769.802490234375, 481.74395751953125, -495.82080078125, 1091.890380859375, 470.297119140625, 1631.58740234375, 485.5912170410156, 1014.3091430664062, 1349.14013671875, 727.5509033203125, 149.35806274414062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000651.npy"}
{"epoch": 0.9559471365638766, "step": 652, "batch_size": 64, "mean": 405.647216796875, "std": 657.3677368164062, "min": -1570.7515869140625, "p10": -252.38703002929685, "median": 313.20367431640625, "p90": 1221.6790893554687, "max": 2408.080078125, "pos_frac": 0.78125, "sample": [394.4510498046875, 242.53952026367188, 819.1635131835938, -830.3416137695312, 551.0350952148438, 350.3501281738281, 197.03526306152344, -414.0016174316406, 741.0879516601562, 161.474853515625, 1225.9444580078125, 1345.041015625, 1124.8492431640625, -1570.7515869140625, 71.34294891357422, 478.0271911621094, 19.736419677734375, 817.3502197265625, 1396.0289306640625, -377.9718322753906, 1211.7265625, 324.6771240234375, 88.58483123779297, -168.54083251953125, 42.450477600097656, -18.1329345703125, 806.0064697265625, -95.10032653808594, 398.9902038574219, 117.32489776611328, 162.8881072998047, 2408.080078125, -268.4347229003906, 2160.725341796875, 325.5106506347656, 1563.5750732421875, 43.73088073730469, 45.87353515625, 75.21112060546875, 31.873504638671875, 560.0875854492188, -2.2065505981445312, 128.40353393554688, -104.4040298461914, 878.1981201171875, 848.411865234375, 945.40185546875, -214.94241333007812, 538.7217407226562, 694.697998046875, 1568.0758056640625, 66.55146789550781, 818.475341796875, 219.60940551757812, -378.6828918457031, 377.2526550292969, 1138.187255859375, -351.186767578125, 405.18017578125, 962.58544921875, 233.9141387939453, 350.97808837890625, 301.730224609375, -23.0291748046875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000652.npy"}
{"epoch": 0.9574155653450808, "step": 653, "batch_size": 64, "mean": 369.4931640625, "std": 631.78564453125, "min": -1180.08056640625, "p10": -293.719091796875, "median": 240.785888671875, "p90": 1214.8829589843751, "max": 2048.3798828125, "pos_frac": 0.734375, "sample": [701.0318603515625, -0.4380645751953125, -131.43421936035156, 1758.4542236328125, 210.2566680908203, -312.1527404785156, 129.8655548095703, -222.674560546875, 83.6944580078125, 821.2294311523438, 683.368408203125, -408.8380126953125, 42.93787384033203, 1394.3956298828125, 781.2046508789062, 1143.41845703125, 844.4715576171875, 1230.06005859375, 1179.4697265625, -92.86395263671875, 175.9306640625, 312.3546447753906, 1712.7620849609375, 435.8686828613281, 439.43414306640625, 737.8994140625, 999.0784301757812, 148.8704833984375, 48.388404846191406, -428.4751892089844, 1403.549072265625, 232.5388946533203, 483.85498046875, -1180.08056640625, -250.70724487304688, -179.2935333251953, 354.30572509765625, 511.9698791503906, -549.8519287109375, 2048.3798828125, 1967.832275390625, 758.5773315429688, 31.372373580932617, -26.327857971191406, 670.8607177734375, 532.0072021484375, 19.34613037109375, -242.08335876464844, 214.27630615234375, 241.44537353515625, 362.8556823730469, 121.0792236328125, 758.6849365234375, 126.93673706054688, 520.4453125, -649.0584106445312, -348.1763000488281, 350.72119140625, 701.2837524414062, 85.65553283691406, 390.6664123535156, -240.12989807128906, 240.12640380859375, -233.06980895996094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000653.npy"}
{"epoch": 0.9588839941262849, "step": 654, "batch_size": 64, "mean": 348.8302001953125, "std": 551.5304565429688, "min": -467.76361083984375, "p10": -312.0627227783203, "median": 291.6698455810547, "p90": 1218.9852661132816, "max": 1704.64404296875, "pos_frac": 0.671875, "sample": [18.853782653808594, 387.82489013671875, -145.65716552734375, 229.655029296875, 421.5499267578125, -46.49528503417969, -321.7386169433594, 288.1494140625, -52.33396530151367, 664.6788330078125, 646.5173950195312, -298.6047058105469, 89.98479461669922, 517.5709228515625, -443.88909912109375, -224.0933837890625, -35.16234588623047, 1629.7386474609375, 948.743408203125, 577.2907104492188, 591.6498413085938, 1599.1185302734375, 433.46826171875, 355.76800537109375, 1246.72412109375, 507.619873046875, -55.66740417480469, 660.4761352539062, 81.68972778320312, -110.58111572265625, -76.64277648925781, 1241.5013427734375, 442.08673095703125, 40.023040771484375, 33.10792541503906, -47.92870330810547, 891.1784057617188, 295.1902770996094, 223.33572387695312, 307.1961975097656, 573.1534423828125, 533.5045776367188, 711.0381469726562, -170.40933227539062, -438.2279052734375, 846.2069702148438, 286.29931640625, 1629.2105712890625, 1166.44775390625, -102.01307678222656, 33.53118896484375, 1332.47021484375, 531.3611450195312, 316.4928283691406, -68.51519012451172, -348.2166442871094, 576.0304565429688, -354.8477478027344, 1704.64404296875, 943.9359130859375, 2.173065185546875, -467.76361083984375, -317.8304443359375, -105.43976593017578], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000654.npy"}
{"epoch": 0.960352422907489, "step": 655, "batch_size": 64, "mean": 380.19952392578125, "std": 598.4489135742188, "min": -655.0719604492188, "p10": -298.3244293212891, "median": 280.7362976074219, "p90": 1073.878369140625, "max": 2318.57861328125, "pos_frac": 0.71875, "sample": [-447.60009765625, -655.0719604492188, 1247.905029296875, 352.1937561035156, 22.0994930267334, 1070.6417236328125, 744.3795166015625, 846.8876953125, 330.8616638183594, -300.4316711425781, 635.9282836914062, 50.12273406982422, -88.436279296875, -541.34423828125, -35.63379669189453, 1066.742919921875, -159.17349243164062, 893.5638427734375, -193.2377166748047, 1075.2655029296875, -210.483642578125, 437.0428466796875, 926.3538818359375, -468.8738708496094, -293.40753173828125, 510.4276123046875, 266.34979248046875, 106.51785278320312, -42.86595153808594, 295.122802734375, -71.13741302490234, 1679.59716796875, 167.24147033691406, 254.09144592285156, 609.0938110351562, 788.432373046875, -636.4378662109375, 933.0523681640625, 510.51397705078125, 884.381103515625, 382.68365478515625, 106.97977447509766, 564.3126831054688, 2318.57861328125, 1594.5545654296875, -44.012359619140625, 419.3935852050781, -360.6694030761719, 1038.3541259765625, 497.32757568359375, 40.95942687988281, 47.6942138671875, 74.28179931640625, 160.40830993652344, 1163.3507080078125, 184.2041473388672, 1624.4547119140625, -146.1327667236328, -96.36408996582031, 494.95452880859375, 545.9216918945312, 251.70570373535156, 155.57261657714844, 753.5802001953125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000655.npy"}
{"epoch": 0.9618208516886931, "step": 656, "batch_size": 64, "mean": 620.030517578125, "std": 679.1792602539062, "min": -529.27392578125, "p10": -110.03804702758788, "median": 431.25865173339844, "p90": 1582.4131835937505, "max": 2685.41552734375, "pos_frac": 0.8125, "sample": [1357.833740234375, 741.625244140625, -529.27392578125, -18.687580108642578, 233.97283935546875, 369.65887451171875, 290.64520263671875, 540.97802734375, 390.1337585449219, 898.5512084960938, 871.834716796875, 663.098388671875, 301.19256591796875, 115.43130493164062, 424.3547668457031, 1432.5810546875, -157.54437255859375, 57.454315185546875, 1646.626953125, 188.6595458984375, 787.46728515625, 11.2923583984375, 1990.3343505859375, -156.90066528320312, -120.69662475585938, 216.4757080078125, 982.8348388671875, 1075.917236328125, 429.712646484375, 1066.23193359375, 457.6867980957031, 2685.41552734375, 635.2265014648438, -122.43617248535156, 2080.633056640625, 802.779296875, 432.8046569824219, 991.9412841796875, 2088.8447265625, 118.83425903320312, 2157.4912109375, -155.5527801513672, 1281.368408203125, 17.8922119140625, 1106.1732177734375, 1020.2211303710938, 358.565673828125, 2092.9150390625, 243.73495483398438, 210.28921508789062, 760.4187622070312, 174.7776336669922, -99.68973541259766, 220.82579040527344, 972.6879272460938, -21.804908752441406, -114.47303771972656, 301.25018310546875, -58.69384002685547, 788.9130249023438, 851.65380859375, -20.029897689819336, 653.480712890625, 666.0125732421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000656.npy"}
{"epoch": 0.9632892804698973, "step": 657, "batch_size": 64, "mean": 494.6385803222656, "std": 664.4813232421875, "min": -1425.0482177734375, "p10": -135.2067283630371, "median": 417.80279541015625, "p90": 1293.8541503906251, "max": 2384.08544921875, "pos_frac": 0.859375, "sample": [1177.096923828125, 30.019004821777344, 822.0592651367188, 443.0352783203125, 1374.634033203125, 1479.237060546875, 801.6234741210938, 109.04283142089844, 1145.29248046875, -49.60693359375, -311.43798828125, 411.675048828125, 1356.482177734375, -231.3782196044922, 188.3127899169922, 493.5003662109375, 63.37080383300781, 401.6113586425781, 118.28874969482422, 1317.4599609375, 2384.08544921875, 375.28619384765625, 44.75395965576172, 2356.14208984375, 1238.77392578125, 1233.8922119140625, 214.7863311767578, 64.3387451171875, 622.90869140625, 605.9666748046875, 953.1543579101562, 423.9305419921875, 34.81181335449219, -1425.0482177734375, 535.0662841796875, 512.8404541015625, 223.5348358154297, -107.09404754638672, 62.4288330078125, 304.3199157714844, 694.59033203125, -992.895263671875, -290.8744812011719, 559.730224609375, 867.8515625, 108.57762908935547, 279.3348388671875, 654.9387817382812, 462.8955078125, 35.21381378173828, 736.09912109375, 716.4083251953125, 160.86253356933594, 716.2586669921875, 443.91180419921875, 162.61795043945312, 1176.8802490234375, 2206.989990234375, 387.4620666503906, -257.1295471191406, 215.56704711914062, -147.25502014160156, 574.8919067382812, 384.74383544921875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000657.npy"}
{"epoch": 0.9647577092511013, "step": 658, "batch_size": 64, "mean": 385.14642333984375, "std": 545.0371704101562, "min": -954.8048095703125, "p10": -337.5264648437499, "median": 354.1825408935547, "p90": 1119.8713256835938, "max": 1716.13330078125, "pos_frac": 0.75, "sample": [624.4769897460938, 624.0816650390625, 626.8404541015625, -278.87078857421875, 288.5442810058594, -156.2474365234375, 1120.9573974609375, 306.2324523925781, -424.51531982421875, 412.1759338378906, 365.68292236328125, 1273.8826904296875, 42.43675231933594, 274.4600524902344, 306.8209228515625, -450.49755859375, 342.6821594238281, 313.8545837402344, 122.84476470947266, -103.95954895019531, 482.4809875488281, 181.2838134765625, 404.81988525390625, 151.5828399658203, -954.8048095703125, 954.7532348632812, -483.4631042480469, 622.107177734375, 662.9453125, -155.36209106445312, -154.31077575683594, 244.486083984375, -15.10715103149414, 431.98211669921875, 291.083251953125, 1461.8238525390625, -5.154245376586914, 1117.337158203125, 660.0631103515625, 1085.21142578125, 84.62007141113281, 325.43023681640625, 588.199951171875, 1245.5382080078125, 1074.3497314453125, 626.838623046875, -106.12881469726562, 492.8064270019531, 828.8421630859375, 866.1874389648438, -494.95904541015625, 1189.6593017578125, -362.66461181640625, 1716.13330078125, 665.2518310546875, 935.2672119140625, 491.231201171875, 1319.1761474609375, -276.2534484863281, 708.50439453125, 100.18037414550781, 431.9228820800781, 119.46929168701172, -535.871337890625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000658.npy"}
{"epoch": 0.9662261380323054, "step": 659, "batch_size": 64, "mean": 391.2337646484375, "std": 537.7700805664062, "min": -547.7483520507812, "p10": -240.65662689208983, "median": 338.67823791503906, "p90": 1033.5119506835938, "max": 2234.4267578125, "pos_frac": 0.765625, "sample": [624.028076171875, -167.52584838867188, 221.02676391601562, 881.7755126953125, -149.0478973388672, -234.84681701660156, 1091.125, -243.14654541015625, -484.9923400878906, -151.4673614501953, 977.066162109375, 309.9384765625, 436.50653076171875, -270.98248291015625, 66.16566467285156, 137.5018310546875, 561.250244140625, 337.6264343261719, 256.8995361328125, 238.03689575195312, 314.88446044921875, 613.4838256835938, -403.5147399902344, 2234.4267578125, -547.7483520507812, 1291.889892578125, 447.3047180175781, 590.2528076171875, 1150.29150390625, 298.011474609375, 357.4033508300781, 351.3886413574219, 0.14730072021484375, -5.6121978759765625, 613.284423828125, 339.73004150390625, 211.09117126464844, 33.957157135009766, 58.29347229003906, 1004.43212890625, -430.32000732421875, -8.037498474121094, 638.3729858398438, -136.2553253173828, 674.5352172851562, 79.89862060546875, -10.033073425292969, 610.6493530273438, 1599.569580078125, 1830.4132080078125, 798.9423217773438, 220.53746032714844, 692.8560791015625, 496.0187072753906, 446.9512939453125, 243.99534606933594, 868.6119995117188, 1045.9747314453125, 533.734130859375, -272.6773986816406, 807.531982421875, 115.43617248535156, 362.6363525390625, 439.2835693359375], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000659.npy"}
{"epoch": 0.9676945668135095, "step": 660, "batch_size": 64, "mean": 488.71484375, "std": 627.0621337890625, "min": -1113.11474609375, "p10": -181.92402038574218, "median": 409.97743225097656, "p90": 1236.7150268554688, "max": 2082.219482421875, "pos_frac": 0.765625, "sample": [-171.0357208251953, 267.4941101074219, 1037.2862548828125, 421.99676513671875, 390.09423828125, 497.7662353515625, 674.9555053710938, 1244.050048828125, 1049.7197265625, 828.180419921875, 327.93389892578125, -269.73193359375, 1147.799072265625, 508.2200927734375, 812.3641357421875, 1588.6705322265625, -184.54795837402344, -14.731948852539062, -148.9220428466797, -237.39279174804688, 1109.5235595703125, 1446.184326171875, 49.49842071533203, 428.8414001464844, 942.1300659179688, 1898.7757568359375, 279.57354736328125, -205.17210388183594, 646.791015625, 311.9586181640625, 193.55618286132812, 1219.5999755859375, -705.7149047851562, 2082.219482421875, 256.5259704589844, 555.3264770507812, 800.3932495117188, 54.30519104003906, 423.6457824707031, 894.9695434570312, 181.9916229248047, 163.27633666992188, 564.2135620117188, 788.9486694335938, -130.27191162109375, -30.828704833984375, 69.11756134033203, 1623.08740234375, 927.8745727539062, 1044.302734375, -175.80149841308594, -50.806190490722656, -538.5187377929688, 687.9576416015625, 1089.3272705078125, 384.87274169921875, 114.1041488647461, 335.5348815917969, 545.6428833007812, 1784.525390625, -142.092529296875, -1113.11474609375, 303.3479919433594, 397.9580993652344], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000660.npy"}
{"epoch": 0.9691629955947136, "step": 661, "batch_size": 64, "mean": 436.3536376953125, "std": 631.232421875, "min": -915.065673828125, "p10": -26.940075302124008, "median": 346.05877685546875, "p90": 1025.093658447266, "max": 3254.3125, "pos_frac": 0.859375, "sample": [238.2321014404297, -32.93946075439453, 753.5514526367188, 26.149763107299805, 347.0548400878906, 59.5024299621582, 580.0391845703125, -270.9609680175781, 627.4609375, 292.62213134765625, 1076.088623046875, 352.56097412109375, 644.8989868164062, 656.9554443359375, 3254.3125, 886.461181640625, 154.11691284179688, 58.6182861328125, -915.065673828125, 162.9010009765625, -12.941509246826172, 906.1054077148438, 137.32122802734375, 345.0627136230469, 169.55572509765625, 188.8307342529297, 322.4318542480469, 873.1968994140625, 1194.091064453125, 56.595794677734375, 254.51292419433594, 586.50146484375, 471.2364196777344, 1160.1834716796875, 468.62017822265625, 263.3931884765625, 187.566650390625, 312.4764709472656, 193.09727478027344, 591.81884765625, -445.3720397949219, 380.8453369140625, -8.895492553710938, 800.8486938476562, 743.5364990234375, -700.1925048828125, 774.11376953125, 713.5010986328125, 8.788885116577148, 593.0567016601562, 2535.33642578125, -652.5243530273438, 254.52391052246094, 48.78257751464844, 724.416748046875, 1342.2491455078125, 69.70696258544922, 590.62646484375, -223.8134765625, 1306.3409423828125, 475.62274169921875, 492.35162353515625, 480.31591796875, 0.24777984619140625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000661.npy"}
{"epoch": 0.9706314243759178, "step": 662, "batch_size": 64, "mean": 342.0297546386719, "std": 630.2185668945312, "min": -1100.287353515625, "p10": -403.1084350585937, "median": 369.1773223876953, "p90": 1014.9060119628907, "max": 2136.427001953125, "pos_frac": 0.71875, "sample": [-97.34578704833984, 840.634521484375, 1098.93896484375, -1100.287353515625, 558.3203735351562, -484.03143310546875, 457.34967041015625, 725.4229736328125, 1811.34423828125, 112.15604400634766, 814.6593017578125, -411.6260986328125, -1087.996826171875, -170.73048400878906, 1391.3251953125, 133.6368408203125, 30.116933822631836, 373.8721618652344, 2136.427001953125, 812.8588256835938, 1023.9386596679688, -225.82699584960938, 406.9509582519531, -86.28729248046875, 185.8536376953125, 868.23828125, 89.97624206542969, 993.829833984375, 280.6430358886719, 194.86590576171875, 334.86468505859375, -171.80145263671875, 455.25634765625, -239.49078369140625, 643.1383666992188, 952.4226684570312, 598.914306640625, 434.3770446777344, 650.2100219726562, 522.2666015625, 35.85694885253906, 216.5218048095703, 969.649658203125, 1704.6322021484375, 538.389892578125, 1191.71923828125, -744.9210815429688, 364.48248291015625, 918.0333251953125, 253.6276092529297, -383.23388671875, 10.44051742553711, -92.57908630371094, 452.3431091308594, -507.0545654296875, -249.83084106445312, 544.2332763671875, 537.5687255859375, 269.2895812988281, 497.9260559082031, 557.9224243164062, -723.8657836914062, -249.38397216796875, -79.24800109863281], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000662.npy"}
{"epoch": 0.9720998531571219, "step": 663, "batch_size": 64, "mean": 411.10906982421875, "std": 730.5654296875, "min": -1677.65966796875, "p10": -259.7738281249999, "median": 283.7246551513672, "p90": 1151.859814453125, "max": 3507.5009765625, "pos_frac": 0.765625, "sample": [864.9564819335938, 404.1650390625, 827.6475219726562, 14.961973190307617, 357.38037109375, 852.0410766601562, 115.39945983886719, 356.5075988769531, -56.12799072265625, 308.6804504394531, 123.02709197998047, 161.227294921875, 1590.908935546875, 830.51220703125, 65.59959411621094, 862.61376953125, 797.556884765625, 28.29901695251465, -1677.65966796875, 132.02838134765625, 258.76885986328125, 626.3792114257812, -292.90936279296875, 642.8381958007812, 634.1297607421875, 381.5780029296875, 595.8775024414062, 184.21511840820312, 414.14361572265625, 3507.5009765625, -475.0997314453125, 1138.809326171875, 1488.96337890625, -95.44639587402344, 535.0513305664062, 1894.481201171875, 1549.0665283203125, 63.249019622802734, 17.706361770629883, 1775.4173583984375, 983.6348876953125, -336.2385559082031, 148.89459228515625, 342.848388671875, -182.45758056640625, 528.0316772460938, 904.768310546875, 151.6424102783203, 132.90069580078125, 963.6461181640625, 179.0282440185547, -39.962989807128906, 1157.452880859375, -990.8975219726562, 894.095703125, -68.45332336425781, -485.0741271972656, -105.90396881103516, 131.78372192382812, 586.6914672851562, -365.33843994140625, -67.1734390258789, -102.95736694335938, 145.57113647460938], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000663.npy"}
{"epoch": 0.973568281938326, "step": 664, "batch_size": 64, "mean": 493.97576904296875, "std": 653.6135864257812, "min": -789.1929931640625, "p10": -194.19190368652343, "median": 418.52610778808594, "p90": 1240.3517578125004, "max": 2338.40380859375, "pos_frac": 0.78125, "sample": [1083.0335693359375, 82.86810302734375, 538.9525756835938, 754.6937255859375, 838.60595703125, 103.4781723022461, 343.54034423828125, 1867.3046875, -170.12820434570312, 1585.295654296875, -453.902587890625, 792.7151489257812, 430.67633056640625, 311.4200744628906, 678.5078735351562, 279.891845703125, -655.7772216796875, 439.182373046875, 23.135902404785156, 528.3444213867188, 1279.163818359375, 1145.29150390625, 28.63671875, 1149.790283203125, -195.28421020507812, 174.42428588867188, 1966.6756591796875, -231.75491333007812, 508.06195068359375, -476.4124450683594, 1125.2706298828125, 176.45730590820312, -52.89141845703125, 1028.724609375, 472.1979675292969, 406.3758850097656, -789.1929931640625, 292.6476745605469, 694.5530395507812, 1131.5098876953125, 111.27503967285156, 540.8052978515625, 913.1564331054688, 23.26708984375, 934.1173095703125, 2338.40380859375, 764.3341064453125, 257.6000671386719, 2226.990234375, -191.6431884765625, -125.17373657226562, 1288.0863037109375, 1038.177978515625, 232.4626007080078, 158.63937377929688, 529.426025390625, 265.7774963378906, 337.0064392089844, -128.4717559814453, -92.41806030273438, 448.55230712890625, -5.560394287109375, 870.90966796875, -357.3572692871094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000664.npy"}
{"epoch": 0.9750367107195301, "step": 665, "batch_size": 64, "mean": 406.1307067871094, "std": 506.5909729003906, "min": -543.9229125976562, "p10": -141.95775756835937, "median": 341.0124969482422, "p90": 1037.8103271484379, "max": 1784.196044921875, "pos_frac": 0.75, "sample": [-181.83279418945312, 32.49081039428711, 359.6478576660156, 505.7364196777344, -57.540870666503906, 494.4068298339844, 941.1513671875, -9.82183837890625, 1402.7095947265625, 461.23748779296875, 335.1290283203125, 583.801025390625, 88.06851196289062, -543.9229125976562, 569.3600463867188, -43.5404052734375, -116.33206176757812, 258.6263732910156, -33.79136657714844, 537.3556518554688, 801.9108276367188, 183.39198303222656, 426.2016906738281, 312.1537780761719, -62.47093200683594, 346.8959655761719, -537.7171020507812, 1498.8980712890625, 715.1781005859375, -34.7899055480957, 790.8041381835938, 300.362548828125, 607.6834716796875, 402.1833801269531, 1136.00244140625, 291.5810546875, 1079.235595703125, 588.99169921875, 315.4142150878906, 255.71224975585938, 290.447509765625, 1585.1990966796875, 500.9413757324219, -381.02142333984375, 427.93267822265625, 868.384765625, 282.90728759765625, -347.1229248046875, 459.05340576171875, 893.5930786132812, 141.6385498046875, 240.23326110839844, 825.501953125, 172.01986694335938, 1703.8079833984375, -220.2037811279297, 592.0581665039062, -136.80380249023438, 144.4324951171875, 770.5294189453125, 1784.196044921875, -8.335067749023438, 546.5801391601562, -144.16659545898438], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000665.npy"}
{"epoch": 0.9765051395007343, "step": 666, "batch_size": 64, "mean": 413.43023681640625, "std": 580.15673828125, "min": -988.5257568359375, "p10": -256.16258239746094, "median": 353.69898986816406, "p90": 1032.8391479492188, "max": 2513.106689453125, "pos_frac": 0.765625, "sample": [750.4779052734375, 314.557861328125, 715.3316040039062, 450.8644104003906, -363.29437255859375, 684.564697265625, 1033.15185546875, 980.4697265625, 131.07357788085938, 1013.5262451171875, -36.5450439453125, 204.66476440429688, 379.8971252441406, -342.8331298828125, -108.38080596923828, 275.91064453125, 318.08087158203125, -231.75775146484375, -107.07333374023438, -131.32763671875, 1278.342041015625, 13.699165344238281, 757.4165649414062, 416.60498046875, 2098.7333984375, -47.44776916503906, 699.967041015625, 358.9316101074219, -164.51638793945312, -621.71044921875, 358.7224426269531, 569.7376708984375, 297.76812744140625, 1121.047607421875, 890.968017578125, 382.1959228515625, 948.5494995117188, -275.8555908203125, 417.60943603515625, 568.0687255859375, 70.15887451171875, 701.1331787109375, -266.6217956542969, 1213.72705078125, 266.37896728515625, 339.85162353515625, -988.5257568359375, 1032.1094970703125, 348.675537109375, 73.03347778320312, 377.55230712890625, 569.3673095703125, 749.757080078125, -19.682064056396484, 331.33819580078125, 173.70858764648438, 213.9716796875, 342.7749938964844, 2513.106689453125, 234.1461944580078, -368.17987060546875, 897.8995971679688, 373.0098571777344, 1280.6531982421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000666.npy"}
{"epoch": 0.9779735682819384, "step": 667, "batch_size": 64, "mean": 479.6250915527344, "std": 613.2628784179688, "min": -698.2890014648438, "p10": -238.25177001953122, "median": 455.54188537597656, "p90": 1311.5103271484381, "max": 2266.107177734375, "pos_frac": 0.8125, "sample": [309.8559875488281, -524.4299926757812, 637.7468872070312, -620.7462158203125, 37.542633056640625, 1034.4827880859375, 823.11669921875, 503.5795593261719, 569.3991088867188, 1391.003662109375, 280.2720642089844, 136.48367309570312, 107.46637725830078, 471.5076904296875, 329.17724609375, -632.1094970703125, 724.4262084960938, 577.7145385742188, 1683.33935546875, 992.8666381835938, 780.0128173828125, 274.8537292480469, 1126.02587890625, 157.64382934570312, 639.5490112304688, 299.49517822265625, 1002.7432861328125, -189.73828125, 898.920654296875, 706.8018188476562, 763.9644775390625, 552.1364135742188, 917.4454956054688, 569.1126708984375, 439.5760803222656, -252.56265258789062, 306.699462890625, 426.93658447265625, 383.2055358886719, -25.52625274658203, -43.99925231933594, 1551.1856689453125, 1743.60693359375, 1.899078369140625, 711.6544799804688, 1492.2049560546875, 649.7574462890625, 161.94454956054688, 125.28082275390625, 747.7952880859375, 24.589927673339844, 701.0540161132812, -415.03656005859375, 192.18336486816406, 2266.107177734375, 1726.856689453125, 81.0436019897461, 884.2583618164062, -568.0352783203125, 361.9317932128906, -698.2890014648438, -204.85971069335938, -103.78207397460938, 696.6618041992188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000667.npy"}
{"epoch": 0.9794419970631424, "step": 668, "batch_size": 64, "mean": 330.46356201171875, "std": 711.972900390625, "min": -1223.20166015625, "p10": -496.23199462890625, "median": 269.52391052246094, "p90": 1176.52763671875, "max": 2541.0732421875, "pos_frac": 0.703125, "sample": [508.6945495605469, 278.91619873046875, 1189.093505859375, -1223.20166015625, 139.96095275878906, 703.8753051757812, 1344.878662109375, 208.5433807373047, 1021.7173461914062, 390.6048583984375, -482.328125, 506.45440673828125, 50.26500701904297, 113.11813354492188, 378.21527099609375, 867.782470703125, 208.3720245361328, 44.17585754394531, -324.2421875, -586.9483032226562, -993.3712158203125, 1096.137939453125, -358.341796875, 671.0980834960938, -136.59530639648438, 887.4994506835938, 304.7860412597656, 378.9333190917969, 39.00376892089844, 622.6560668945312, 342.2501220703125, 539.472412109375, 1147.207275390625, -524.4874267578125, 953.5728759765625, 453.32012939453125, -751.3800048828125, 1067.59130859375, 84.17133331298828, -191.42022705078125, 246.71246337890625, -141.06195068359375, 1402.20703125, -30.595876693725586, -502.1907958984375, 1071.0693359375, -23.26812744140625, -292.08160400390625, 260.1316223144531, -44.329017639160156, 324.18646240234375, 216.9451904296875, 302.0780029296875, 226.66506958007812, -121.18618774414062, 1217.79638671875, 412.3696594238281, 312.12750244140625, 2303.04736328125, -1189.44384765625, -50.76232147216797, 1679.047119140625, 59.07781982421875, 2541.0732421875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000668.npy"}
{"epoch": 0.9809104258443465, "step": 669, "batch_size": 64, "mean": 478.28851318359375, "std": 637.1280517578125, "min": -1391.9324951171875, "p10": -168.2071662902831, "median": 436.46607971191406, "p90": 1317.1952148437501, "max": 2415.11572265625, "pos_frac": 0.8125, "sample": [183.0876922607422, 51.20478820800781, -85.31084442138672, 996.6419677734375, -955.6890869140625, 325.94476318359375, 717.9298706054688, 740.20556640625, 363.4356384277344, -91.8770523071289, 2415.11572265625, -57.67190933227539, 26.639846801757812, 327.0118713378906, -283.0970153808594, 457.5926818847656, 1349.9649658203125, -92.74492645263672, 129.55401611328125, 226.2039031982422, 553.6593017578125, 778.8955078125, 1342.6065673828125, 1086.0035400390625, 204.5894012451172, 1775.63916015625, 99.19500732421875, 1323.80810546875, 968.739501953125, 101.44976806640625, 598.3657836914062, 1185.645751953125, -200.54812622070312, 525.8193359375, 1301.76513671875, -550.4249877929688, -280.6379699707031, 848.9056396484375, 864.08984375, 818.8507690429688, 1472.3826904296875, 306.1687927246094, 143.76019287109375, 295.2716979980469, -1391.9324951171875, 731.7003173828125, 905.7796630859375, 660.4144287109375, 459.660400390625, 254.37322998046875, -577.0918579101562, 999.9583129882812, 415.3394775390625, 483.7779235839844, 555.593505859375, 776.2140502929688, 47.69174575805664, 253.94454956054688, 541.78662109375, 1217.125244140625, -76.33901977539062, 1440.154052734375, 337.7060852050781, 266.466796875], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000669.npy"}
{"epoch": 0.9823788546255506, "step": 670, "batch_size": 64, "mean": 475.13616943359375, "std": 636.8126831054688, "min": -1235.189697265625, "p10": -317.0511047363281, "median": 332.7907409667969, "p90": 1324.7405029296876, "max": 1771.3065185546875, "pos_frac": 0.796875, "sample": [274.0053405761719, 488.1278076171875, 241.67498779296875, 1370.9188232421875, 432.63037109375, -189.1116943359375, -222.58773803710938, 3.9534130096435547, 1097.5810546875, 595.7627563476562, 273.87847900390625, -106.8179931640625, -1235.189697265625, 1692.535400390625, 1060.0992431640625, 1771.3065185546875, 827.4933471679688, 434.6686096191406, 1110.1153564453125, 531.4649658203125, 163.3888702392578, -532.1116333007812, 1189.4134521484375, -200.477294921875, -538.2930297851562, 332.17431640625, 1541.9625244140625, 105.353515625, 985.6559448242188, -463.8868713378906, 1346.2486572265625, 772.1445922851562, -175.6353302001953, 276.7417907714844, 239.05166625976562, 1368.9473876953125, 1115.389404296875, -318.15155029296875, 227.65036010742188, 1186.69189453125, 259.7940673828125, 1197.795166015625, 769.293701171875, -337.41729736328125, 1749.86279296875, 830.593994140625, 333.40716552734375, 288.6880798339844, -314.4833984375, -323.8253173828125, 9.765472412109375, 1274.5548095703125, 173.35247802734375, 329.44976806640625, 1261.52587890625, 172.83755493164062, 300.7174987792969, 915.6322631835938, 17.73651123046875, 892.8411254882812, 372.49908447265625, 499.57958984375, 80.2008056640625, 579.545166015625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000670.npy"}
{"epoch": 0.9838472834067548, "step": 671, "batch_size": 64, "mean": 392.805419921875, "std": 622.9658203125, "min": -1438.388671875, "p10": -167.8800094604492, "median": 399.63035583496094, "p90": 1189.6399291992188, "max": 1719.4854736328125, "pos_frac": 0.765625, "sample": [74.37138366699219, -1394.864013671875, 507.9355163574219, 128.62710571289062, -130.960205078125, 745.5385131835938, 89.54953002929688, 399.09295654296875, 582.6729125976562, 409.5718994140625, 442.72039794921875, -89.92789459228516, 121.44215393066406, 1145.2138671875, 1105.0809326171875, 83.15171813964844, 319.8321838378906, 134.093505859375, 920.23095703125, 739.6783447265625, 106.06986236572266, 1430.267333984375, -1438.388671875, 1124.0211181640625, 730.3923950195312, 32.75077819824219, 1703.7548828125, 527.9421997070312, 237.1832275390625, 476.4917297363281, -128.80026245117188, -471.0336608886719, -367.0113525390625, 469.72393798828125, 1191.4822998046875, -135.603271484375, 368.95147705078125, 406.9186706542969, 638.3448486328125, 1305.442626953125, 1404.337890625, 369.6183776855469, -174.28855895996094, 148.99362182617188, 684.4712524414062, -25.242996215820312, 314.7109069824219, 882.3748779296875, -480.4515075683594, -30.947486877441406, 589.69287109375, 436.4781188964844, -27.61295509338379, 200.20074462890625, 1185.341064453125, 692.1090087890625, -152.92672729492188, -591.1861572265625, 133.29823303222656, 1478.5078125, 404.97943115234375, 1719.4854736328125, 1035.48291015625, 400.1677551269531], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000671.npy"}
{"epoch": 0.9853157121879589, "step": 672, "batch_size": 64, "mean": 396.134033203125, "std": 695.1176147460938, "min": -1125.5809326171875, "p10": -392.5641387939453, "median": 326.74114990234375, "p90": 1306.2592651367193, "max": 2590.458740234375, "pos_frac": 0.703125, "sample": [313.68048095703125, 38.018585205078125, -287.0430908203125, 138.24276733398438, -78.55732727050781, 163.74476623535156, 822.8596801757812, 419.0103759765625, 1642.2978515625, 290.8855285644531, 48.76677703857422, -62.68824005126953, -401.4400939941406, -143.8146209716797, 563.8243408203125, 238.22280883789062, -88.34255981445312, -508.9886474609375, 456.2923583984375, 1657.499267578125, 87.40765380859375, 884.2415771484375, -36.464088439941406, -152.72201538085938, 1389.408935546875, -1125.5809326171875, 1129.242431640625, -934.096923828125, 486.31304931640625, 2590.458740234375, -122.82472229003906, 918.6817016601562, 976.4752197265625, 966.1937866210938, 1128.316162109375, 582.6273803710938, 526.1741333007812, -371.85357666015625, 278.2383117675781, 1711.58251953125, -829.6793212890625, -22.208091735839844, 402.08721923828125, 1040.441162109375, 825.4498901367188, 568.3536376953125, 885.0209350585938, 207.45654296875, 795.4273071289062, 661.3162841796875, 1356.086181640625, 23.554962158203125, -131.99832153320312, 649.404541015625, 51.38677978515625, 690.5130615234375, 1189.9964599609375, 94.15856170654297, -98.92515563964844, -723.7142333984375, 517.8427124023438, 339.80181884765625, 1418.770751953125, -692.2550048828125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000672.npy"}
{"epoch": 0.986784140969163, "step": 673, "batch_size": 64, "mean": 377.1641540527344, "std": 637.5220336914062, "min": -1326.3653564453125, "p10": -255.27341156005858, "median": 375.02545166015625, "p90": 1024.1608398437506, "max": 3373.58056640625, "pos_frac": 0.78125, "sample": [-145.8921356201172, -689.823974609375, 327.8944396972656, 882.8502197265625, 348.85009765625, 532.4140625, 826.98291015625, 398.1463928222656, 1218.81640625, 375.1814880371094, -227.9412078857422, 1471.96533203125, 466.4131774902344, 532.362548828125, 152.72752380371094, 223.457275390625, -434.540283203125, 815.5114135742188, 1485.2568359375, 350.9979248046875, -112.55876159667969, 285.5643310546875, 334.6116943359375, -91.874267578125, 89.66325378417969, -1326.3653564453125, 254.72756958007812, 619.1802368164062, -564.9892578125, 173.2548370361328, 713.1780395507812, 1379.407470703125, 187.13893127441406, -13.394855499267578, 715.5405883789062, 439.83880615234375, 388.4940185546875, -516.7538452148438, 604.9574584960938, 792.9398193359375, 713.0933837890625, 19.91385841369629, 56.827606201171875, 669.8331298828125, 397.8017578125, 81.90311431884766, 490.61456298828125, -184.23472595214844, 441.0061340332031, 3373.58056640625, 179.54208374023438, 61.9320182800293, 580.2049560546875, 596.6337890625, -595.0084228515625, 434.55584716796875, 549.4539184570312, 1105.6083984375, 188.44384765625, -266.9872131347656, -61.765342712402344, 581.7687377929688, 1084.7225341796875, 374.8694152832031], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000673.npy"}
{"epoch": 0.9882525697503671, "step": 674, "batch_size": 64, "mean": 487.11492919921875, "std": 666.9636840820312, "min": -657.0740966796875, "p10": -260.0626556396484, "median": 440.7497100830078, "p90": 1158.471557617188, "max": 3308.49755859375, "pos_frac": 0.765625, "sample": [-143.43255615234375, 33.91953659057617, 737.1055297851562, 650.2362670898438, 464.2071228027344, 777.2605590820312, 81.55699157714844, 48.45561218261719, -562.7434692382812, 1195.826904296875, 1371.2490234375, 1031.66259765625, -35.42283248901367, 456.76806640625, -207.6250457763672, 452.0506286621094, -222.82125854492188, 1978.17333984375, 587.6943359375, 3308.49755859375, 553.4080200195312, 538.69384765625, 20.755352020263672, 926.1446533203125, 429.44879150390625, -319.1925048828125, 1017.2472534179688, 457.229736328125, 706.268798828125, -170.46368408203125, 1071.30908203125, 1828.76611328125, -100.68612670898438, 262.65911865234375, 396.34027099609375, 813.514892578125, -276.02325439453125, -296.46795654296875, -314.8507995605469, 837.1629028320312, 727.5333251953125, 271.9779357910156, 352.83819580078125, 1669.0263671875, 332.3351745605469, 785.445068359375, 982.19384765625, 944.4706420898438, 596.00732421875, 214.3439178466797, -657.0740966796875, 185.6090850830078, 1703.27685546875, -308.2815856933594, -62.793907165527344, 690.822021484375, 189.4891357421875, 716.4066772460938, 344.4201965332031, 539.6480712890625, 162.73040771484375, -89.90757751464844, 186.63668823242188, 314.32037353515625], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000674.npy"}
{"epoch": 0.9897209985315712, "step": 675, "batch_size": 64, "mean": 502.36358642578125, "std": 671.8565063476562, "min": -595.5717163085938, "p10": -141.88766174316402, "median": 313.2147521972656, "p90": 1289.98291015625, "max": 2787.55712890625, "pos_frac": 0.8125, "sample": [218.9851837158203, 1287.4903564453125, 456.51922607421875, 76.32856750488281, 112.08089447021484, 132.90455627441406, 825.6228637695312, -23.994049072265625, 557.3114013671875, 311.3122253417969, 809.8807373046875, -275.18896484375, 742.0744018554688, 1050.712646484375, -162.294921875, -35.79988098144531, 103.23941040039062, 1364.073974609375, 509.8266296386719, 484.8586730957031, 651.4148559570312, 24.524620056152344, 2340.031494140625, 777.5576782226562, 23.39220428466797, 337.8219299316406, 42.813961029052734, 777.2163696289062, 585.67529296875, 309.34088134765625, 371.006103515625, 1291.0511474609375, 312.05224609375, 2010.3046875, 95.80496215820312, 1269.0029296875, 2787.55712890625, -174.7180938720703, 1106.7718505859375, 2070.696044921875, 298.4396057128906, 199.05123901367188, -19.282485961914062, 54.69232177734375, 314.37725830078125, -94.27072143554688, -595.5717163085938, 209.59921264648438, 553.4733276367188, 214.84268188476562, 112.13032531738281, 1379.1041259765625, 630.1287841796875, -548.7804565429688, -591.0774536132812, 367.55975341796875, 152.6151123046875, 672.5374145507812, -45.746612548828125, 1285.2996826171875, 282.5016174316406, 1125.985107421875, 993.873779296875, -355.4754943847656], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000675.npy"}
{"epoch": 0.9911894273127754, "step": 676, "batch_size": 64, "mean": 466.677734375, "std": 563.8510131835938, "min": -900.5250854492188, "p10": -67.67688674926757, "median": 413.1147766113281, "p90": 1285.5497558593752, "max": 2311.710205078125, "pos_frac": 0.796875, "sample": [104.61837005615234, 1599.6756591796875, 806.2943725585938, 969.3776245117188, -55.14598083496094, -59.576210021972656, 474.4132385253906, 712.400390625, 16.39284896850586, -900.5250854492188, 550.057861328125, 40.831634521484375, -54.21852111816406, 1063.8155517578125, 1416.5078125, 2311.710205078125, -12.713813781738281, -71.14860534667969, 15.387222290039062, -147.17501831054688, 103.35782623291016, 435.6292724609375, 619.0155029296875, 7.999359130859375, -135.86239624023438, 676.8514404296875, 101.050537109375, 1263.602294921875, 348.8601989746094, 1649.1409912109375, -25.494781494140625, 597.7706298828125, 287.0462646484375, 644.5150146484375, 84.18335723876953, 703.287841796875, 1549.941650390625, 430.3480224609375, 557.861572265625, 1359.183837890625, 554.09765625, -4.656211853027344, 149.1174774169922, -245.6936798095703, 1294.955810546875, 666.0703125, 63.895782470703125, 816.5469360351562, 442.6612548828125, 332.7321472167969, 927.1349487304688, 480.083251953125, 378.2980651855469, 434.2454833984375, 61.645973205566406, 270.6402587890625, 927.8851318359375, -263.611083984375, 1243.7056884765625, 127.99771118164062, 395.88153076171875, 297.8256530761719, -154.0706329345703, 630.7173461914062], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000676.npy"}
{"epoch": 0.9926578560939795, "step": 677, "batch_size": 64, "mean": 446.36956787109375, "std": 700.0260620117188, "min": -798.1535034179688, "p10": -211.0741973876953, "median": 327.8285369873047, "p90": 1140.8051513671876, "max": 3515.512939453125, "pos_frac": 0.78125, "sample": [233.6812744140625, 2854.539794921875, 504.5494689941406, 39.804527282714844, 417.281005859375, -437.37054443359375, -145.86935424804688, -192.51315307617188, 379.4130859375, -11.9647216796875, 654.5397338867188, 557.7667236328125, 62.80979919433594, 1253.09033203125, 307.15594482421875, 716.7041015625, 835.6181030273438, 55.50062561035156, 225.8667449951172, 276.08807373046875, 1152.19580078125, 246.1250762939453, 1067.717041015625, -247.8268585205078, 391.0784912109375, 1084.0772705078125, 226.13319396972656, 483.1412048339844, 358.27880859375, 1116.0518798828125, 316.52435302734375, 8.817901611328125, -188.68214416503906, 225.90170288085938, 190.78753662109375, -521.7837524414062, 888.5704345703125, 328.75201416015625, 3515.512939453125, 220.83599853515625, 222.67758178710938, 484.4605712890625, 864.8297729492188, 700.4115600585938, -67.9855728149414, 1151.4136962890625, 326.9050598144531, 216.30990600585938, 768.27490234375, 214.2643585205078, 1342.1671142578125, -41.41630935668945, 601.027587890625, 359.1768493652344, 541.59423828125, -440.7124938964844, -747.76171875, 1701.8345947265625, 648.013427734375, 621.9922485351562, 761.55029296875, -219.0289306640625, -798.1535034179688, -93.09225463867188], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000677.npy"}
{"epoch": 0.9941262848751835, "step": 678, "batch_size": 64, "mean": 523.98193359375, "std": 714.199462890625, "min": -1661.43505859375, "p10": -160.75921478271482, "median": 465.3302764892578, "p90": 1236.8371459960938, "max": 2979.2919921875, "pos_frac": 0.8125, "sample": [237.40945434570312, 843.317626953125, 535.010986328125, 603.4755859375, 515.3168334960938, -170.48324584960938, 613.3247680664062, -1661.43505859375, 66.53289794921875, -231.05368041992188, -88.37098693847656, 1085.0526123046875, 64.15492248535156, 1216.6192626953125, -128.07040405273438, 507.11260986328125, 836.1005249023438, -300.4000244140625, 829.1128540039062, 811.6080322265625, 873.897216796875, 1350.03466796875, 16.359207153320312, 2233.061767578125, 1154.8875732421875, 2979.2919921875, 164.7893524169922, 223.6036834716797, -14.874202728271484, 1058.8040771484375, -471.1916809082031, 104.5357894897461, 159.73809814453125, 795.9534301757812, 204.91107177734375, 953.7525634765625, 140.23403930664062, 1139.4031982421875, 304.902099609375, 827.094482421875, -138.06980895996094, 753.6318969726562, 1245.501953125, 1542.8531494140625, 284.2601318359375, 315.5924987792969, 2105.655517578125, 2085.734130859375, 94.59852600097656, 451.3873291015625, 358.8302001953125, -389.6415710449219, 479.2732238769531, 56.65797424316406, 44.29595184326172, 105.77156066894531, 592.6884765625, 1030.1099853515625, 589.7548828125, 514.3319091796875, 392.6531677246094, -26.08285140991211, 918.3170166015625, -256.7820739746094], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000678.npy"}
{"epoch": 0.9955947136563876, "step": 679, "batch_size": 64, "mean": 362.1433410644531, "std": 578.2069702148438, "min": -777.6499633789062, "p10": -307.8836059570312, "median": 246.5436553955078, "p90": 1244.0425415039065, "max": 1880.03466796875, "pos_frac": 0.78125, "sample": [450.6427001953125, 21.7924861907959, 1184.903564453125, 1569.3507080078125, 427.1812438964844, 799.3965454101562, 781.0455322265625, 437.8966369628906, 1269.3878173828125, 313.2655334472656, 419.86102294921875, 472.4624938964844, 451.5157775878906, 141.0472869873047, 951.3802490234375, -388.2159729003906, -140.34414672851562, 163.21810913085938, 233.1947021484375, 947.8236083984375, 664.3541259765625, 1394.181884765625, 1.3365936279296875, 127.74429321289062, 195.2790985107422, -322.3935241699219, 170.73919677734375, 1540.9202880859375, -605.1730346679688, 545.4229736328125, 801.7811889648438, 66.660400390625, -72.09696960449219, 53.92377471923828, 828.6504516601562, 359.08685302734375, 1667.920166015625, 142.2889404296875, -139.37155151367188, 18.588478088378906, 115.50759887695312, 37.88420104980469, 527.570556640625, 259.8926086425781, -389.0367126464844, 1363.20458984375, 216.1743621826172, -274.0271301269531, -722.5654296875, -108.42977905273438, 817.316650390625, -558.7105712890625, 338.4953918457031, -165.89878845214844, 440.6809997558594, 295.73138427734375, -85.68179321289062, 193.67337036132812, 844.0525512695312, 1880.03466796875, 225.9967041015625, 644.0720825195312, -777.6499633789062, 112.23621368408203], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000679.npy"}
{"epoch": 0.9970631424375918, "step": 680, "batch_size": 64, "mean": 483.0449523925781, "std": 711.842041015625, "min": -2201.535888671875, "p10": -200.60778198242184, "median": 361.66685485839844, "p90": 1493.9922119140626, "max": 2109.31787109375, "pos_frac": 0.828125, "sample": [1131.59326171875, 364.9379577636719, 358.395751953125, 787.7052001953125, 180.84519958496094, 1031.6268310546875, 1711.150390625, 507.85784912109375, 224.57577514648438, 354.7611083984375, 497.6444091796875, 196.38377380371094, 790.6135864257812, -406.3210754394531, 1173.63623046875, 1497.24951171875, 602.73388671875, -576.4783935546875, -227.7501983642578, 798.1368408203125, 1065.357421875, 137.35821533203125, -89.94160461425781, 315.28045654296875, -208.86477661132812, 951.5865478515625, 789.8953247070312, 239.96031188964844, 624.1834716796875, 1370.8330078125, 246.5104217529297, 20.06256675720215, 191.58267211914062, 1722.906982421875, 980.489990234375, 168.4573211669922, 180.85508728027344, 508.66162109375, 1889.377197265625, -650.7796020507812, 155.81878662109375, 522.27880859375, 464.36627197265625, 603.0582885742188, 153.26458740234375, -2201.535888671875, 1653.96337890625, -161.16949462890625, -537.7661743164062, 1335.815673828125, 511.910400390625, 2109.31787109375, 19.498077392578125, 34.57432556152344, 195.24740600585938, 1486.391845703125, 732.3642578125, 1553.4273681640625, -181.34146118164062, 172.32794189453125, -131.91952514648438, 245.06289672851562, 0.047943115234375, 726.8038330078125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000680.npy"}
{"epoch": 0.9985315712187959, "step": 681, "batch_size": 64, "mean": 385.8537292480469, "std": 605.7545776367188, "min": -665.2570190429688, "p10": -364.2838134765625, "median": 364.14414978027344, "p90": 1103.2621459960938, "max": 2237.37646484375, "pos_frac": 0.6875, "sample": [700.6250610351562, -423.4295654296875, 529.5761108398438, 138.67779541015625, 844.516845703125, -665.2570190429688, 137.92166137695312, 333.7727966308594, -424.9667663574219, 326.6544189453125, -239.44366455078125, 482.2549743652344, -478.7701416015625, -348.49847412109375, 1061.71875, 1538.254638671875, 477.4807434082031, -38.50127410888672, -272.37945556640625, 1105.14501953125, -82.44336700439453, -9.290916442871094, 790.88525390625, 421.7236022949219, 1565.724853515625, 594.1597290039062, 248.73501586914062, 1025.3133544921875, 849.575927734375, 636.6077880859375, 109.8266372680664, 446.20233154296875, 212.4404296875, -110.50204467773438, -169.39019775390625, 273.91192626953125, 161.19427490234375, -581.5899658203125, 464.8574523925781, -160.1449432373047, -357.3948974609375, 2237.37646484375, -622.4024658203125, 770.4360961914062, -76.74080657958984, 394.5155029296875, -367.2362060546875, 211.8491668701172, 93.68141174316406, -268.727294921875, 855.3018188476562, -163.61294555664062, 799.6243286132812, 517.75390625, 1246.3233642578125, 514.2962036132812, 1098.8687744140625, 977.90185546875, 1062.1043701171875, 640.0664672851562, 162.0082244873047, 1256.998291015625, 1386.9805908203125, 851.5177001953125], "npy": "/root/dynamic-dpo-v4/outputs/llama-3-8b-base-new-dpo-hh-helpful-s_star0.4-4xh200-batch-64-20260421-214335-rerun/margin_logs/step_0000681.npy"}