Files
qwen3-8b-base-new-dpo-hh-ha…/margin_logs/margins.jsonl
ModelHub XC 252813cb7c 初始化项目,由ModelHub XC社区提供模型
Model: jackf857/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4
Source: Original Platform
2026-05-12 02:17:33 +08:00

662 lines
1.0 MiB

{"epoch": 0.0, "step": 1, "batch_size": 64, "mean": -0.0029816031455993652, "std": 0.38981664180755615, "min": -0.7835464477539062, "p10": -0.5016929626464843, "median": 0.02667522430419922, "p90": 0.4355194091796875, "max": 1.2425384521484375, "pos_frac": 0.53125, "sample": [-0.2990684509277344, 0.05040740966796875, 0.4813804626464844, -0.7835464477539062, 0.16756057739257812, -0.21320724487304688, 0.066741943359375, 0.169891357421875, -0.06363677978515625, -0.33983612060546875, 0.20204925537109375, -0.003765106201171875, -0.7424850463867188, -0.039760589599609375, 0.008941650390625, 0.2320232391357422, 0.3860015869140625, 0.11869239807128906, -0.36592864990234375, -0.047290802001953125, -0.28316688537597656, 0.0283660888671875, -0.351715087890625, 0.11574554443359375, 0.86297607421875, -0.7426376342773438, 0.1338043212890625, -0.21837997436523438, 0.426910400390625, -0.12430953979492188, 0.2183837890625, -0.4932708740234375, 0.13604736328125, 0.1666259765625, 0.024984359741210938, -0.42929840087890625, -0.6993560791015625, -0.413604736328125, 0.22283935546875, -0.0557861328125, 1.2425384521484375, -0.2928791046142578, -0.14715576171875, 0.3737640380859375, -0.14208221435546875, 0.19033432006835938, 0.3464927673339844, 0.20479965209960938, 0.04190826416015625, -0.00957489013671875, -0.5053024291992188, 0.4848480224609375, 0.2988262176513672, 0.045352935791015625, 0.427978515625, -0.5745201110839844, 0.5770988464355469, 0.1401214599609375, -0.027454376220703125, -0.6424560546875, -0.2728919982910156, -0.428192138671875, 0.5285491943359375, 0.438751220703125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000001.npy"}
{"epoch": 0.0015117157974300832, "step": 2, "batch_size": 64, "mean": 0.029325813055038452, "std": 0.47058698534965515, "min": -1.2616119384765625, "p10": -0.39437751770019525, "median": -0.11953926086425781, "p90": 0.6386299133300786, "max": 1.48486328125, "pos_frac": 0.4375, "sample": [-0.43146514892578125, 0.07180404663085938, -0.20481109619140625, -0.00714111328125, 0.5232467651367188, 0.06253433227539062, -0.07450485229492188, -0.35506439208984375, -0.14567184448242188, -0.2234630584716797, -0.31732177734375, 1.456878662109375, 0.14324188232421875, -0.41083526611328125, -0.4837646484375, -0.12252044677734375, -0.1322479248046875, 0.45180511474609375, -0.6440353393554688, -1.2616119384765625, 0.7379837036132812, 0.0069866180419921875, 0.14553451538085938, 0.2057647705078125, -0.11970138549804688, 0.1814441680908203, -0.2711448669433594, -0.22872161865234375, 0.23077011108398438, 0.2108001708984375, 0.348419189453125, -0.10046005249023438, 0.4903106689453125, -0.209228515625, 0.3726234436035156, -0.2670707702636719, 0.056774139404296875, 0.1702728271484375, -0.3437042236328125, -0.5232925415039062, 0.1266021728515625, -0.31758880615234375, -0.4544639587402344, -0.13794708251953125, 0.5147171020507812, 0.03656768798828125, 1.48486328125, -0.2191619873046875, -0.22581100463867188, -0.11937713623046875, -0.1849536895751953, 0.9678802490234375, 0.3454742431640625, -0.16698455810546875, -0.2411823272705078, -0.1938018798828125, 0.999603271484375, -0.17424774169921875, 0.908782958984375, -0.3559761047363281, -0.17584609985351562, 0.688079833984375, 0.04034423828125, -0.2581329345703125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000002.npy"}
{"epoch": 0.0030234315948601664, "step": 3, "batch_size": 64, "mean": -0.026119887828826904, "std": 0.4387124478816986, "min": -1.4166717529296875, "p10": -0.4598175048828125, "median": -0.006603240966796875, "p90": 0.4366516113281251, "max": 1.359405517578125, "pos_frac": 0.484375, "sample": [-0.37566375732421875, 0.23485183715820312, -0.17536163330078125, 0.5390396118164062, -1.4166717529296875, -0.01930999755859375, -0.011260986328125, -0.28614044189453125, -0.104217529296875, 0.2634124755859375, -0.1735076904296875, 0.44366455078125, 0.07244873046875, -0.107452392578125, -0.5330810546875, -0.9506912231445312, -0.200103759765625, 0.4202880859375, 0.05163764953613281, 0.414642333984375, -0.3824920654296875, -0.10361099243164062, 0.6924972534179688, 0.48990631103515625, -0.3680419921875, 0.25499725341796875, 0.2889251708984375, 0.00958251953125, 0.14304542541503906, 0.2736968994140625, -0.5632228851318359, 0.12537384033203125, 0.12093353271484375, -0.40206146240234375, -0.038074493408203125, -0.2542743682861328, -1.0967941284179688, -0.7733154296875, -0.19484329223632812, -0.36724853515625, -0.02677154541015625, 0.10428237915039062, 0.16277313232421875, 0.561798095703125, 0.18677520751953125, -0.1341705322265625, 0.2722587585449219, 0.013427734375, -0.43447113037109375, -0.06104278564453125, 0.1268157958984375, -0.47068023681640625, 0.2641410827636719, -0.0221099853515625, -0.3976593017578125, 0.22234344482421875, -0.22405242919921875, 0.8594970703125, 0.012676239013671875, 1.359405517578125, -0.00194549560546875, 0.064300537109375, -0.353240966796875, 0.3024749755859375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000003.npy"}
{"epoch": 0.0045351473922902496, "step": 4, "batch_size": 64, "mean": -0.09968316555023193, "std": 0.43002378940582275, "min": -1.46099853515625, "p10": -0.6136440277099608, "median": -0.06336784362792969, "p90": 0.36873741149902345, "max": 0.9605560302734375, "pos_frac": 0.375, "sample": [-0.0510711669921875, -0.8007431030273438, -0.380157470703125, -0.07479095458984375, 0.36246490478515625, -0.18966293334960938, -0.04729270935058594, -0.7150650024414062, -0.23600006103515625, 0.1410369873046875, -0.2966804504394531, 0.1404876708984375, 0.356781005859375, 0.5804443359375, -0.4389190673828125, -1.10406494140625, 0.647613525390625, 0.5118484497070312, -0.1147613525390625, 0.1829376220703125, -0.82025146484375, -0.2997169494628906, 0.0713043212890625, 0.0289154052734375, -0.09989166259765625, 0.20183563232421875, -0.31121063232421875, 0.3714256286621094, -0.1462860107421875, -0.2486114501953125, 0.252960205078125, -0.002971649169921875, -0.3486175537109375, -0.08733367919921875, -0.388671875, 0.9605560302734375, -0.1121978759765625, -0.2801933288574219, -1.46099853515625, -0.01836395263671875, 0.057708740234375, -0.16152381896972656, -0.06646347045898438, 0.10154342651367188, 0.7017974853515625, -0.027984619140625, 0.18838882446289062, 0.024105072021484375, -0.3725738525390625, 0.19111251831054688, -0.3352546691894531, -0.4026947021484375, -0.08211135864257812, 0.03760528564453125, -1.1468963623046875, -0.060272216796875, -0.46466064453125, 0.4182929992675781, 0.2608528137207031, -0.5442085266113281, 0.24173736572265625, -0.643402099609375, -0.006256103515625, -0.02465057373046875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000004.npy"}
{"epoch": 0.006046863189720333, "step": 5, "batch_size": 64, "mean": -0.057308852672576904, "std": 0.5109280347824097, "min": -1.3619537353515625, "p10": -0.7629455566406249, "median": 0.0095062255859375, "p90": 0.5408508300781251, "max": 0.7597427368164062, "pos_frac": 0.5, "sample": [0.06749725341796875, 0.25983428955078125, 0.12149238586425781, 0.090850830078125, 0.2657012939453125, -0.2509918212890625, -0.043361663818359375, -0.6129608154296875, -0.28943634033203125, 0.6941986083984375, 0.533203125, -1.301361083984375, -0.303985595703125, 0.5130805969238281, 0.73870849609375, 0.54412841796875, -0.0986480712890625, -0.00293731689453125, 0.33172607421875, -0.2979278564453125, -0.016155242919921875, 0.16066551208496094, -0.08575057983398438, 0.02194976806640625, 0.282073974609375, -1.209625244140625, -0.7993698120117188, -0.13091278076171875, -1.3619537353515625, 0.10773086547851562, 0.02216339111328125, -0.05814361572265625, -0.4203033447265625, -0.025182723999023438, -0.6695785522460938, -0.37221527099609375, 0.3271198272705078, -0.059051513671875, 0.3648338317871094, -0.08028411865234375, 0.030231475830078125, -1.10540771484375, 0.7597427368164062, 0.5650100708007812, -0.25485992431640625, 0.1627941131591797, 0.39836883544921875, -0.6779556274414062, 0.51605224609375, -0.05416107177734375, 0.24566650390625, 0.733978271484375, -0.40288543701171875, 0.6006622314453125, -1.329315185546875, 0.08945465087890625, -0.1424407958984375, 0.19567489624023438, 0.055797576904296875, 0.26140594482421875, 0.2826976776123047, -0.8268203735351562, -0.1523590087890625, -0.5759201049804688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000005.npy"}
{"epoch": 0.007558578987150416, "step": 6, "batch_size": 64, "mean": 0.06428647041320801, "std": 0.4112168550491333, "min": -0.839019775390625, "p10": -0.42820663452148433, "median": 0.06373786926269531, "p90": 0.49914512634277347, "max": 1.405792236328125, "pos_frac": 0.5625, "sample": [-0.12801742553710938, 0.2694587707519531, 0.4464836120605469, -0.5591049194335938, -0.17350387573242188, -0.1508636474609375, 0.00051116943359375, 0.2749786376953125, 0.05301666259765625, 0.13184356689453125, -0.279632568359375, -0.4649658203125, 0.2159442901611328, -0.141937255859375, 0.1030120849609375, -0.21630859375, 0.0655517578125, -0.7683830261230469, 0.3217048645019531, -0.0474700927734375, 0.490020751953125, -0.823455810546875, 0.24263763427734375, -0.09461784362792969, 0.27106475830078125, 0.42159271240234375, -0.767059326171875, -0.839019775390625, 0.061923980712890625, 0.16101455688476562, -0.13071441650390625, 0.38567352294921875, 0.535675048828125, 0.12804031372070312, 0.10810089111328125, -0.22629547119140625, 0.23101043701171875, 0.6341629028320312, 0.01082611083984375, 0.4130859375, 0.3071136474609375, 0.17597198486328125, 0.09113311767578125, -0.07659912109375, 0.14630126953125, -0.18292999267578125, 1.1349945068359375, 0.741363525390625, 0.19751358032226562, -0.187652587890625, 0.199676513671875, -0.3753509521484375, -0.03498268127441406, -0.08752250671386719, -0.45085906982421875, -0.10222625732421875, -0.10262680053710938, -0.18599700927734375, 1.405792236328125, -0.0735321044921875, 0.7454986572265625, 0.5030555725097656, 0.3468666076660156, -0.18665313720703125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000006.npy"}
{"epoch": 0.009070294784580499, "step": 7, "batch_size": 64, "mean": -0.08962073922157288, "std": 0.47593727707862854, "min": -1.33172607421875, "p10": -0.7346477508544922, "median": -0.07745170593261719, "p90": 0.4260574340820313, "max": 1.46685791015625, "pos_frac": 0.46875, "sample": [0.04972076416015625, -0.1839752197265625, 0.40950775146484375, -0.09022140502929688, -0.23196792602539062, 0.4295158386230469, 0.2560005187988281, -0.6590995788574219, -0.205413818359375, 0.1305999755859375, 0.5134162902832031, 0.3429145812988281, 0.120147705078125, 0.0498504638671875, 0.05149078369140625, 0.4140472412109375, 0.141326904296875, -0.15079879760742188, 0.09723663330078125, -0.41762542724609375, -0.17227554321289062, 1.46685791015625, 0.5286865234375, 0.4842376708984375, -0.7805938720703125, -0.4140434265136719, -0.202301025390625, -0.714447021484375, 0.03505706787109375, 0.13144493103027344, -0.7433052062988281, 0.1246795654296875, -0.180572509765625, -0.22855377197265625, 0.056682586669921875, -0.7668075561523438, -0.09143447875976562, -0.25323486328125, 0.4179878234863281, 0.1985321044921875, 0.10950469970703125, -0.024784088134765625, 0.23018646240234375, -0.5153350830078125, 0.15003204345703125, -0.7877426147460938, -0.23253631591796875, 0.035247802734375, -0.5349235534667969, 1.159393310546875, -0.3236122131347656, 0.562774658203125, -0.0646820068359375, -0.5452308654785156, -0.15069580078125, -0.5943603515625, 0.06171417236328125, -0.45098876953125, -1.33172607421875, -0.91143798828125, -0.7794036865234375, -0.10480499267578125, 0.00572967529296875, -0.66131591796875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000007.npy"}
{"epoch": 0.010582010582010581, "step": 8, "batch_size": 64, "mean": -0.03348389267921448, "std": 0.47948718070983887, "min": -0.8033294677734375, "p10": -0.5605331420898437, "median": -0.07758140563964844, "p90": 0.4251411437988283, "max": 1.9595947265625, "pos_frac": 0.359375, "sample": [-0.1359081268310547, -0.682098388671875, 1.9595947265625, -0.19326019287109375, -0.1292572021484375, 0.19982528686523438, -0.621307373046875, -0.5634765625, 0.132720947265625, -0.8033294677734375, -0.3839111328125, -0.3758544921875, -0.025182723999023438, -0.33463287353515625, -0.498382568359375, 0.5154953002929688, 0.7166595458984375, -0.20733261108398438, -0.10620880126953125, -0.6179656982421875, 0.110321044921875, -0.694915771484375, -0.0187225341796875, 1.562347412109375, 0.0380859375, -0.2505226135253906, 0.746978759765625, -0.380889892578125, -0.08347320556640625, -0.07138824462890625, 0.3758201599121094, -0.40097808837890625, -0.34002685546875, 0.24608993530273438, -0.6447830200195312, -0.416717529296875, 0.3834419250488281, 0.012285232543945312, 0.633880615234375, 0.4430122375488281, -0.4057464599609375, -0.07168960571289062, 0.24706268310546875, -0.27368927001953125, 0.3566246032714844, -0.0631103515625, -0.3173942565917969, -0.06911468505859375, -0.12416648864746094, -0.17742919921875, 0.1804943084716797, -0.24460601806640625, -0.035144805908203125, 0.12146377563476562, 0.238677978515625, -0.0020904541015625, -0.5477561950683594, -0.5536651611328125, 0.28658294677734375, 0.17038726806640625, -0.016632080078125, 0.31899261474609375, -0.112640380859375, -0.14441299438476562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000008.npy"}
{"epoch": 0.012093726379440665, "step": 9, "batch_size": 64, "mean": 0.06733927130699158, "std": 0.44315221905708313, "min": -1.459869384765625, "p10": -0.3665611267089844, "median": 0.09519290924072266, "p90": 0.6377990722656252, "max": 0.8797111511230469, "pos_frac": 0.578125, "sample": [0.118682861328125, -1.0063095092773438, 0.5713424682617188, -0.5961761474609375, 0.092041015625, 0.8797111511230469, -0.2540550231933594, 0.608062744140625, -0.208770751953125, 0.4366607666015625, 0.29868316650390625, 0.261016845703125, 0.276763916015625, 0.10462188720703125, -1.459869384765625, -0.09932327270507812, 0.75592041015625, 0.7855377197265625, -0.1670684814453125, -0.0347442626953125, 0.5341796875, 0.08177566528320312, -0.1517486572265625, 0.15204238891601562, 0.12946701049804688, 0.4322052001953125, -0.253936767578125, 0.1970062255859375, -0.032825469970703125, -0.083221435546875, 0.35085296630859375, -0.13860321044921875, -0.0163116455078125, 0.3607940673828125, -0.8383445739746094, 0.126708984375, 0.13556289672851562, -0.36833953857421875, 0.650543212890625, -0.03031158447265625, -0.11443328857421875, -0.381622314453125, -0.3624114990234375, -0.020263671875, -0.23238372802734375, -0.22237014770507812, 0.4928474426269531, 0.07253646850585938, 0.09834480285644531, -0.1771240234375, -0.194549560546875, -0.03521728515625, 0.0743560791015625, 0.32659912109375, 0.15081787109375, 0.1268463134765625, 0.6929206848144531, 0.5869979858398438, 0.20494461059570312, 0.00717926025390625, 0.73577880859375, 0.7681732177734375, 0.10652542114257812, -0.9950027465820312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000009.npy"}
{"epoch": 0.013605442176870748, "step": 10, "batch_size": 64, "mean": 0.12282583117485046, "std": 0.44259899854660034, "min": -0.8345794677734375, "p10": -0.40459518432617186, "median": 0.07452964782714844, "p90": 0.6654479980468752, "max": 1.336669921875, "pos_frac": 0.578125, "sample": [0.549774169921875, 0.2828254699707031, -0.19845199584960938, 0.231842041015625, 0.6865005493164062, -0.7357330322265625, 0.07460403442382812, -0.7368927001953125, 0.408050537109375, -0.3916358947753906, -0.03496551513671875, -0.43780517578125, -0.39186859130859375, 0.1990203857421875, -0.8345794677734375, -0.13375091552734375, 0.21783065795898438, 0.140167236328125, -0.140716552734375, 0.5197830200195312, -0.10595321655273438, -0.19989013671875, -0.0046234130859375, 0.7026596069335938, 0.3368339538574219, 0.07445526123046875, -0.01959991455078125, 1.238983154296875, 0.1780242919921875, 0.451080322265625, 0.20704269409179688, 0.6953659057617188, 0.24458694458007812, 0.0211029052734375, -0.197906494140625, 0.6156845092773438, 0.7663803100585938, 0.19816207885742188, -0.5505523681640625, 0.1820068359375, -0.041515350341796875, -0.03076934814453125, -0.1236419677734375, 1.336669921875, 0.059703826904296875, 0.2623710632324219, 0.01139068603515625, 0.6163253784179688, 1.3350067138671875, -0.1507720947265625, 0.0971832275390625, -0.4100494384765625, 0.302764892578125, -0.027057647705078125, -0.4237098693847656, -0.02098846435546875, -0.176361083984375, 0.2856616973876953, -0.3716888427734375, 0.5207748413085938, 0.2451324462890625, 0.43552398681640625, 0.045318603515625, -0.024261474609375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000010.npy"}
{"epoch": 0.015117157974300832, "step": 11, "batch_size": 64, "mean": 0.004195868968963623, "std": 0.41499000787734985, "min": -1.064300537109375, "p10": -0.48200073242187497, "median": 0.03537178039550781, "p90": 0.4709819793701172, "max": 1.082855224609375, "pos_frac": 0.53125, "sample": [-0.19631195068359375, -0.3141365051269531, 0.810333251953125, 0.5819015502929688, -0.011346817016601562, 0.17719650268554688, -0.2957954406738281, -0.04380989074707031, 0.574462890625, 0.172576904296875, 0.09334564208984375, -0.237945556640625, -0.506591796875, 0.259246826171875, -0.402313232421875, -0.100555419921875, 0.3461151123046875, 0.0441741943359375, -0.285888671875, 0.0121917724609375, 0.3576622009277344, -0.1259613037109375, 0.33056640625, 0.21107864379882812, -0.13301849365234375, -0.380828857421875, -0.0225982666015625, -0.37958526611328125, -0.18877029418945312, -0.2585105895996094, 0.07755661010742188, 0.044002532958984375, -0.1383075714111328, 0.47467041015625, 0.36115264892578125, 0.22783660888671875, 0.9678726196289062, -0.8209381103515625, -0.35953521728515625, -0.42462158203125, 1.082855224609375, 0.3583030700683594, -0.15735244750976562, 0.2027587890625, 0.6310272216796875, -0.7218780517578125, -0.8196868896484375, 0.23228073120117188, 0.3361358642578125, -0.57562255859375, -0.5250720977783203, 0.3059844970703125, 0.02674102783203125, 0.07875823974609375, 0.20331573486328125, -1.064300537109375, -0.1515960693359375, -0.269134521484375, 0.14551544189453125, 0.05810546875, 0.4623756408691406, 0.1625213623046875, -0.4227294921875, 0.192657470703125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000011.npy"}
{"epoch": 0.016628873771730914, "step": 12, "batch_size": 64, "mean": -0.0058082640171051025, "std": 0.3620968759059906, "min": -1.0873260498046875, "p10": -0.46809883117675777, "median": 0.016571044921875, "p90": 0.31931438446044924, "max": 0.9600372314453125, "pos_frac": 0.53125, "sample": [-0.36237335205078125, 0.11598968505859375, -0.1125640869140625, 0.26313209533691406, 0.29218101501464844, 0.4814300537109375, -0.24670791625976562, 0.38495635986328125, 0.2084808349609375, 0.2548255920410156, -0.0383758544921875, -0.510650634765625, 0.19952392578125, -0.3997802734375, 0.24935531616210938, -0.4117851257324219, 0.15802001953125, 0.18680572509765625, -0.4922332763671875, 0.31729698181152344, -0.14544677734375, 0.28106689453125, -0.7490997314453125, 0.4921112060546875, 0.03717803955078125, 0.16724777221679688, 0.16692733764648438, 0.9600372314453125, 0.0174560546875, 0.2467193603515625, -0.20958328247070312, -0.305419921875, 0.2016143798828125, -1.0873260498046875, -0.2267894744873047, -0.045307159423828125, 0.2569770812988281, -0.011013031005859375, -0.0283050537109375, -0.3389129638671875, 0.3201789855957031, 0.2854766845703125, -0.0391693115234375, -0.17390060424804688, 0.10880279541015625, 0.0499725341796875, -0.2618370056152344, 0.15842819213867188, -0.01654815673828125, -0.38657379150390625, -0.7361488342285156, 0.0015163421630859375, 0.28711700439453125, 0.274566650390625, 0.89508056640625, -0.5650100708007812, 0.3357810974121094, -0.296966552734375, -0.5813331604003906, -0.04790496826171875, 0.10297393798828125, 0.01568603515625, -0.1570892333984375, -0.16248703002929688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000012.npy"}
{"epoch": 0.018140589569160998, "step": 13, "batch_size": 64, "mean": 0.09430572390556335, "std": 0.3625533878803253, "min": -0.6329345703125, "p10": -0.27704696655273436, "median": 0.00505828857421875, "p90": 0.5117507934570313, "max": 1.0798797607421875, "pos_frac": 0.5, "sample": [0.06884765625, 0.22531890869140625, -0.0824432373046875, -0.10640335083007812, 0.9639892578125, 0.21715164184570312, 0.35730743408203125, -0.16692352294921875, 1.0798797607421875, -0.1250457763671875, -0.208709716796875, -0.17989730834960938, -0.02321624755859375, 0.374725341796875, -0.16528701782226562, -0.3739509582519531, 0.4710693359375, 0.25865936279296875, 0.3904571533203125, 0.2552680969238281, -0.468994140625, -0.21486663818359375, 0.520782470703125, 1.003173828125, -0.049930572509765625, -0.48914337158203125, 0.1594696044921875, 0.021558761596679688, 0.38970947265625, -0.0081024169921875, -0.045330047607421875, -0.09066009521484375, -0.6329345703125, 0.3056221008300781, 0.7093353271484375, 0.27484893798828125, -0.056118011474609375, 0.22420120239257812, 0.41164207458496094, -0.16449737548828125, 0.2101459503173828, 0.17887115478515625, 0.427276611328125, 0.21760940551757812, -0.05600738525390625, -0.0935516357421875, 0.7441635131835938, -0.2914276123046875, -0.17006301879882812, 0.80670166015625, -0.28211212158203125, 0.0989990234375, -0.052722930908203125, 0.044414520263671875, 0.018218994140625, -0.15357208251953125, 0.4906768798828125, -0.14359664916992188, -0.488006591796875, 0.13541412353515625, -0.23496246337890625, -0.0520782470703125, -0.08415985107421875, -0.265228271484375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000013.npy"}
{"epoch": 0.019652305366591082, "step": 14, "batch_size": 64, "mean": -0.0942547619342804, "std": 0.33739882707595825, "min": -1.238006591796875, "p10": -0.4626434326171875, "median": -0.062221527099609375, "p90": 0.2578567504882813, "max": 0.6434326171875, "pos_frac": 0.421875, "sample": [-0.06197357177734375, -0.1273956298828125, -0.009246826171875, -0.26006317138671875, -0.3518257141113281, -0.19011306762695312, -0.202850341796875, 0.049468994140625, -0.2988319396972656, -0.4704322814941406, -0.17908859252929688, -1.1562423706054688, 0.055255889892578125, -0.062469482421875, 0.24007797241210938, 0.15935707092285156, 0.2654762268066406, -0.27490997314453125, -0.008544921875, -0.4444694519042969, -0.3611602783203125, 0.0593109130859375, 0.029958724975585938, -0.1511383056640625, -0.28081512451171875, -0.4982452392578125, -1.238006591796875, 0.19672393798828125, 0.03227806091308594, -0.2285308837890625, 0.02602386474609375, -0.734588623046875, -0.5203170776367188, 0.6434326171875, -0.23611068725585938, 0.48309326171875, -0.043727874755859375, 0.47174835205078125, 0.4634132385253906, -0.420166015625, -0.48675537109375, -0.1607513427734375, 0.142608642578125, 0.09259033203125, 0.04164886474609375, -0.15645980834960938, 0.15430450439453125, 0.0983123779296875, 0.21990585327148438, 0.2675189971923828, -0.12136077880859375, 0.01132965087890625, 0.236083984375, -0.354095458984375, -0.11737442016601562, 0.4917030334472656, 0.0875396728515625, -0.25243568420410156, 0.0594482421875, -0.05098152160644531, -0.34195518493652344, -0.11664581298828125, 0.1160125732421875, -0.2568511962890625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000014.npy"}
{"epoch": 0.021164021164021163, "step": 15, "batch_size": 64, "mean": -0.0212060809135437, "std": 0.4226672053337097, "min": -1.154205322265625, "p10": -0.4553474426269531, "median": -0.056804656982421875, "p90": 0.46765975952148436, "max": 1.0021438598632812, "pos_frac": 0.421875, "sample": [-0.3367156982421875, 0.3599853515625, 0.20716094970703125, 0.02660369873046875, 0.19657135009765625, -0.4593048095703125, -1.147369384765625, -0.573333740234375, -0.494903564453125, -0.2535400390625, -0.11773300170898438, -0.11510658264160156, -0.78955078125, -0.10405349731445312, -0.2726287841796875, -0.318511962890625, -0.260498046875, 0.21599578857421875, -0.130462646484375, -0.3375358581542969, 0.41727447509765625, 0.46915435791015625, 0.185760498046875, -0.44611358642578125, 0.1682720184326172, 0.803955078125, 0.15966796875, 0.46417236328125, 0.5324478149414062, -0.208343505859375, 0.10857772827148438, 0.3879070281982422, 0.2199249267578125, -0.09786224365234375, -0.0655517578125, -0.3946685791015625, -1.154205322265625, -0.4069366455078125, -0.2310333251953125, 0.20757293701171875, -0.12518310546875, -0.7127532958984375, 0.43326568603515625, -0.384429931640625, -0.04345130920410156, 0.06427001953125, -0.01798248291015625, 0.0527801513671875, 0.701385498046875, 0.9163818359375, 0.6213531494140625, -0.20292282104492188, 0.12346839904785156, -0.180633544921875, -0.04691886901855469, 0.4117279052734375, -0.0521087646484375, -0.02496337890625, -0.1959991455078125, -0.07931137084960938, 0.27999114990234375, 1.0021438598632812, -0.2508392333984375, -0.06150054931640625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000015.npy"}
{"epoch": 0.022675736961451247, "step": 16, "batch_size": 64, "mean": -0.05562084913253784, "std": 0.3417550325393677, "min": -0.8294029235839844, "p10": -0.4112281799316406, "median": -0.1107025146484375, "p90": 0.40740814208984377, "max": 0.8552398681640625, "pos_frac": 0.40625, "sample": [0.08346176147460938, 0.12808990478515625, 0.26904296875, 0.2689399719238281, -0.15833282470703125, -0.25435447692871094, -0.07747268676757812, -0.1976776123046875, -0.1647491455078125, -0.15473175048828125, 0.09366607666015625, -0.00171661376953125, 0.6890029907226562, -0.6452369689941406, -0.14612579345703125, -0.13991546630859375, -0.1699371337890625, -0.18349456787109375, -0.39068603515625, 0.4080810546875, 0.59234619140625, -0.79656982421875, 0.457427978515625, -0.6602401733398438, 0.0960845947265625, 0.0667877197265625, -0.8294029235839844, -0.32634735107421875, 0.5481529235839844, -0.384490966796875, 0.04567718505859375, -0.3234443664550781, 0.00153350830078125, -0.008487701416015625, 0.248138427734375, -0.31414794921875, -0.19110107421875, -0.203765869140625, 0.4058380126953125, -0.1224212646484375, -0.4079437255859375, 0.3651084899902344, -0.308563232421875, 0.16851425170898438, -0.27262115478515625, 0.3885993957519531, -0.05704498291015625, 0.05204010009765625, -0.3319129943847656, -0.3250732421875, 0.16688919067382812, 0.08236503601074219, 0.043109893798828125, 0.10204696655273438, -0.0989837646484375, -0.17232704162597656, -0.28699493408203125, 0.439605712890625, 0.8552398681640625, -0.029752731323242188, -0.4135322570800781, -0.41263580322265625, -0.5029449462890625, -0.16034317016601562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000016.npy"}
{"epoch": 0.02418745275888133, "step": 17, "batch_size": 64, "mean": -0.050705909729003906, "std": 0.40225985646247864, "min": -1.3983154296875, "p10": -0.5257904052734375, "median": -0.020704269409179688, "p90": 0.41272583007812524, "max": 0.8307952880859375, "pos_frac": 0.484375, "sample": [-0.27008819580078125, -0.0310516357421875, -0.10427284240722656, -0.19936752319335938, 0.314849853515625, 0.0988616943359375, -0.47434234619140625, 0.2519054412841797, -0.05224609375, -0.02184295654296875, 0.5140533447265625, 0.09516143798828125, 0.27063560485839844, -0.019565582275390625, 0.08558273315429688, -0.26068878173828125, 0.16278076171875, -0.5284042358398438, 0.18833351135253906, -0.08185577392578125, 0.0104217529296875, 0.34909820556640625, -0.937469482421875, 0.8307952880859375, -0.5595932006835938, 0.10101699829101562, -0.37277984619140625, 0.034000396728515625, 0.3098793029785156, 0.10786819458007812, 0.646148681640625, 0.2459564208984375, -0.5362091064453125, 0.43999481201171875, 0.46117401123046875, -0.897247314453125, -0.5196914672851562, -0.20796966552734375, -0.3686676025390625, -1.3983154296875, 0.5908966064453125, 0.025203704833984375, 0.24456024169921875, -0.19591903686523438, -0.027372360229492188, -0.07877349853515625, -0.0930633544921875, -0.33929443359375, 0.024810791015625, 0.149566650390625, -0.6777877807617188, -0.5164794921875, 0.11974334716796875, 0.1816539764404297, -0.38694000244140625, -0.25057220458984375, 0.167388916015625, 0.2842254638671875, -0.4223175048828125, -0.46392059326171875, -0.09334945678710938, 0.6555404663085938, -0.046047210693359375, 0.22621917724609375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000017.npy"}
{"epoch": 0.025699168556311415, "step": 18, "batch_size": 64, "mean": 0.10842183232307434, "std": 0.40685081481933594, "min": -0.814178466796875, "p10": -0.3362598419189453, "median": 0.061229705810546875, "p90": 0.5872707366943362, "max": 1.52117919921875, "pos_frac": 0.5625, "sample": [0.10570526123046875, 0.13938140869140625, 0.045650482177734375, -0.0696258544921875, 0.37493133544921875, -0.494720458984375, -0.3013648986816406, 0.2081756591796875, 0.5338897705078125, -0.4214210510253906, -0.053150177001953125, -0.1385498046875, -0.6048660278320312, -0.14238739013671875, 0.0221405029296875, -0.22025299072265625, -0.034149169921875, 0.046142578125, 0.13393402099609375, 1.52117919921875, -0.3274345397949219, -0.17787933349609375, 0.16675567626953125, 0.15981292724609375, 0.3266944885253906, 0.28513336181640625, 0.2503509521484375, -0.13017654418945312, -0.11275482177734375, -0.09547805786132812, -0.07906341552734375, -0.3145904541015625, -0.1275634765625, -0.2714691162109375, 0.398223876953125, 0.8292083740234375, 0.10771560668945312, -0.060546875, 1.1346588134765625, -0.814178466796875, 0.2556915283203125, 0.2063751220703125, 0.727447509765625, 0.9337158203125, 0.20377731323242188, 0.5320968627929688, 0.25412559509277344, -0.0045166015625, 0.5244979858398438, 0.13431549072265625, -0.07781982421875, 0.6059303283691406, -0.06995391845703125, -0.07822036743164062, -0.4043426513671875, -0.3400421142578125, 0.07631683349609375, -0.4054832458496094, 0.07805633544921875, 0.0222015380859375, 0.6083984375, 0.4975738525390625, 0.3170623779296875, 0.543731689453125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000018.npy"}
{"epoch": 0.027210884353741496, "step": 19, "batch_size": 64, "mean": -0.026145905256271362, "std": 0.36883682012557983, "min": -1.4538650512695312, "p10": -0.38705978393554685, "median": 0.006293296813964844, "p90": 0.37961502075195314, "max": 1.0499267578125, "pos_frac": 0.515625, "sample": [0.118316650390625, 0.16803741455078125, 0.089141845703125, 0.57470703125, -0.02387237548828125, -0.00261688232421875, -0.307037353515625, -0.01723480224609375, 0.14690399169921875, -0.06404876708984375, 0.6308364868164062, 0.38187408447265625, 0.5052833557128906, 0.15446090698242188, 0.0652923583984375, 0.05274200439453125, -0.4406852722167969, 0.059795379638671875, -0.01615142822265625, 0.16035079956054688, -1.4538650512695312, -1.028656005859375, -0.3646392822265625, 0.3743438720703125, -0.1946868896484375, 0.10010528564453125, 0.162628173828125, -0.0113067626953125, 0.06906890869140625, 0.05027198791503906, -0.22417449951171875, -0.10736846923828125, 0.1561126708984375, -0.5551033020019531, 1.0499267578125, -0.3617095947265625, -0.059589385986328125, 0.018619537353515625, 0.0338897705078125, 0.17887496948242188, 0.1299591064453125, -0.29559326171875, 0.5378570556640625, 0.4188079833984375, 0.0955352783203125, 0.13220596313476562, -0.37076568603515625, -0.216278076171875, 0.010068893432617188, -0.06258392333984375, 0.0514068603515625, -0.1323108673095703, 0.24190521240234375, -0.024921417236328125, 0.0025177001953125, -0.4729347229003906, -0.08026123046875, -0.11931228637695312, -0.03675079345703125, -0.39404296875, 0.0135955810546875, -0.780853271484375, -0.0300750732421875, -0.35935211181640625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000019.npy"}
{"epoch": 0.02872260015117158, "step": 20, "batch_size": 64, "mean": -0.03781440854072571, "std": 0.3980479836463928, "min": -1.0927505493164062, "p10": -0.5274978637695312, "median": -0.034847259521484375, "p90": 0.386461639404297, "max": 1.101715087890625, "pos_frac": 0.453125, "sample": [0.0663909912109375, 0.6823348999023438, 0.003238677978515625, 0.060352325439453125, -0.6316413879394531, 0.40053558349609375, 0.7043914794921875, 1.101715087890625, 0.3536224365234375, -0.7012290954589844, -0.0431060791015625, 0.13996505737304688, 0.017669677734375, -0.09924697875976562, 0.143646240234375, 0.5431747436523438, -0.26324462890625, -0.11220359802246094, -0.14539718627929688, -0.6537399291992188, -0.3747596740722656, -0.03565216064453125, 0.20941162109375, -0.14675140380859375, -0.15978240966796875, -0.34597015380859375, 0.335662841796875, -0.127105712890625, 0.288909912109375, -0.0340423583984375, -0.34967041015625, -0.06672286987304688, -0.2054901123046875, 0.024967193603515625, -0.18560791015625, 0.0714263916015625, 0.14906692504882812, -0.9264984130859375, -0.44500732421875, -0.3987922668457031, -0.29775428771972656, -0.043949127197265625, -0.00473785400390625, -0.61090087890625, 0.16007232666015625, 0.015625, 0.03066253662109375, -0.5501022338867188, -0.2393798828125, -0.12140655517578125, -1.0927505493164062, 0.1673583984375, 0.31172943115234375, 0.2977142333984375, -0.0315704345703125, 0.05377197265625, -0.47475433349609375, -0.14910888671875, 1.0817413330078125, 0.15388870239257812, -0.20827102661132812, 0.44089508056640625, 0.005741119384765625, -0.1594562530517578], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000020.npy"}
{"epoch": 0.030234315948601664, "step": 21, "batch_size": 64, "mean": 0.07584935426712036, "std": 0.40153610706329346, "min": -0.9348297119140625, "p10": -0.35763435363769525, "median": 0.02008056640625, "p90": 0.5984283447265626, "max": 1.09326171875, "pos_frac": 0.546875, "sample": [0.06034088134765625, -0.1686553955078125, -0.0578765869140625, 1.09326171875, 0.00046539306640625, 0.1240997314453125, -0.13208770751953125, 0.294464111328125, -0.3771934509277344, 0.2944774627685547, 0.5360984802246094, 0.2230224609375, 1.09222412109375, -0.4947357177734375, 0.586273193359375, -0.3119964599609375, 0.097686767578125, -0.01165771484375, -0.27046966552734375, 0.4623260498046875, -0.2419586181640625, 0.12025070190429688, -0.07184982299804688, -0.5757980346679688, 0.335693359375, -0.1291046142578125, -0.10375785827636719, -0.17816162109375, 0.9075927734375, -0.3856029510498047, 0.217987060546875, -0.29583740234375, -0.0750732421875, -0.15744781494140625, 0.4268341064453125, 0.22847747802734375, -0.9348297119140625, -0.0205535888671875, 0.7993316650390625, -0.04928398132324219, -0.1119842529296875, 0.010162353515625, 0.6036376953125, 0.16831207275390625, 0.21771240234375, 0.22570037841796875, -0.2400970458984375, 0.5459709167480469, -0.204345703125, 0.43717193603515625, -0.5006217956542969, -0.065460205078125, -0.22503280639648438, 0.141937255859375, 0.13785362243652344, 0.0250701904296875, 0.31627655029296875, -0.26968955993652344, 0.18925857543945312, 0.0150909423828125, -0.7909774780273438, 0.6283721923828125, 0.673828125, 0.06923675537109375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000021.npy"}
{"epoch": 0.031746031746031744, "step": 22, "batch_size": 64, "mean": -0.0592007040977478, "std": 0.4054179787635803, "min": -1.391021728515625, "p10": -0.5186279296874999, "median": -0.03862476348876953, "p90": 0.3722923278808596, "max": 1.0284423828125, "pos_frac": 0.46875, "sample": [0.099212646484375, 0.0943450927734375, 0.15659332275390625, -0.20528030395507812, -0.27788352966308594, -0.711212158203125, -0.0001678466796875, -0.9202651977539062, 0.04078865051269531, 0.2540779113769531, 0.0852508544921875, -0.504486083984375, 0.13067626953125, 0.07143020629882812, 0.47426605224609375, -0.35974884033203125, 0.32218170166015625, 0.24133682250976562, 0.29509735107421875, 0.039031982421875, -0.19106292724609375, 0.393768310546875, 0.027557373046875, -0.03680610656738281, -0.20756912231445312, 1.0284423828125, -0.0889892578125, -0.3484077453613281, 0.06042289733886719, 0.041900634765625, 0.6707763671875, -0.3315887451171875, -0.24797439575195312, 0.14910507202148438, 0.006702423095703125, -0.19257736206054688, -0.18889236450195312, -0.31926727294921875, -0.1942901611328125, 0.9556961059570312, 0.007244110107421875, 0.85150146484375, -0.536834716796875, 0.46634674072265625, -0.12335586547851562, -1.391021728515625, 0.2788124084472656, -0.6441650390625, 0.10650634765625, 0.07806396484375, -0.23226547241210938, -0.524688720703125, -0.09195327758789062, -0.37162017822265625, -0.5741348266601562, 0.0856781005859375, -0.19796371459960938, -0.43663787841796875, -0.448822021484375, -0.04247283935546875, -0.04044342041015625, 0.15438079833984375, -0.2293701171875, -0.2438201904296875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000022.npy"}
{"epoch": 0.03325774754346183, "step": 23, "batch_size": 64, "mean": 0.03171861171722412, "std": 0.38152167201042175, "min": -0.86761474609375, "p10": -0.40219001770019525, "median": 0.043914794921875, "p90": 0.46609497070312506, "max": 1.199493408203125, "pos_frac": 0.5625, "sample": [0.3728790283203125, -0.09173583984375, -0.22552490234375, -0.30963134765625, -0.25949668884277344, -0.027292251586914062, -0.076141357421875, 1.199493408203125, -0.23078155517578125, -0.02410888671875, -0.5186004638671875, 0.2266063690185547, 0.6137542724609375, 0.2930908203125, 0.45290374755859375, -0.63067626953125, 0.0764923095703125, 0.3304939270019531, 0.032703399658203125, 0.5888595581054688, 0.37851715087890625, 0.06213951110839844, -0.24913787841796875, 0.5570068359375, 0.7591819763183594, 0.581146240234375, -0.35231781005859375, 0.4436683654785156, -0.7599945068359375, 0.3193817138671875, 0.18182373046875, 0.04221343994140625, -0.2528076171875, -0.1823577880859375, 0.19427490234375, -0.27779388427734375, 0.29779052734375, 0.058502197265625, -0.3645896911621094, 0.1537322998046875, -0.418304443359375, 0.212921142578125, 0.11935806274414062, -0.86761474609375, -0.2469024658203125, -0.6699676513671875, -0.02016448974609375, 0.08300399780273438, 0.04561614990234375, -0.16033935546875, 0.4006805419921875, -0.029537200927734375, 0.185333251953125, 0.2848663330078125, 0.47174835205078125, 0.23985671997070312, 0.02272796630859375, 0.13240814208984375, -0.2347259521484375, 0.009796142578125, -0.7020111083984375, 0.0811004638671875, -0.012298583984375, -0.28122711181640625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000023.npy"}
{"epoch": 0.03476946334089191, "step": 24, "batch_size": 64, "mean": 0.030714750289916992, "std": 0.4012909233570099, "min": -1.0073776245117188, "p10": -0.5576995849609374, "median": 0.03994274139404297, "p90": 0.5239486694335938, "max": 0.9244461059570312, "pos_frac": 0.625, "sample": [0.16971588134765625, 0.10335159301757812, 0.41059112548828125, 0.1534442901611328, 0.30170440673828125, 0.5274505615234375, 0.028242111206054688, 0.6052017211914062, -1.0073776245117188, 0.9244461059570312, -0.02074432373046875, 0.15562820434570312, 0.020229339599609375, -0.4966888427734375, 0.0380859375, 0.128265380859375, -0.11960220336914062, -0.1355438232421875, 0.1392059326171875, -0.25969696044921875, 0.05591583251953125, 0.109344482421875, -0.15460205078125, 0.37540435791015625, -0.8597869873046875, 0.023677825927734375, 0.4211273193359375, 0.515777587890625, 0.038543701171875, 0.22090911865234375, 0.0610809326171875, 0.04134178161621094, -0.7246780395507812, 0.033428192138671875, 0.3323211669921875, -0.5838470458984375, -0.15972518920898438, 0.5540103912353516, 0.04576873779296875, -0.31320953369140625, -0.7984695434570312, -0.09658050537109375, 0.03832054138183594, -0.03015899658203125, -0.69903564453125, 0.692535400390625, -0.23159027099609375, 0.04509735107421875, -0.046901702880859375, 0.7487640380859375, -0.12691497802734375, 0.287750244140625, 0.15727996826171875, -0.1517791748046875, 0.263458251953125, -5.340576171875e-05, 0.019437789916992188, -0.15643310546875, 0.40631866455078125, 0.08200836181640625, 0.77947998046875, -0.9757537841796875, -0.17083740234375, 0.23109054565429688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000024.npy"}
{"epoch": 0.036281179138321996, "step": 25, "batch_size": 64, "mean": 0.028328508138656616, "std": 0.4574028253555298, "min": -1.122283935546875, "p10": -0.47253723144531246, "median": -0.0016527175903320312, "p90": 0.575604248046875, "max": 1.4410247802734375, "pos_frac": 0.5, "sample": [0.5041923522949219, -0.182403564453125, 0.38988494873046875, -0.13727569580078125, 0.37701416015625, -0.16960525512695312, 0.5540924072265625, -0.0048961639404296875, 1.22833251953125, -0.40850067138671875, -0.0875091552734375, -0.6194305419921875, 0.3945159912109375, -0.1161346435546875, 0.01291656494140625, 0.64495849609375, 1.4410247802734375, 0.079010009765625, -0.3208732604980469, 0.2411651611328125, -0.5402297973632812, 0.2992839813232422, -0.56207275390625, 0.8477325439453125, -0.43199920654296875, 0.2852210998535156, 0.16014480590820312, -1.122283935546875, -0.6797256469726562, -0.19040679931640625, -0.3331756591796875, 0.29732513427734375, 0.2867431640625, 0.038097381591796875, 0.0068511962890625, 0.2692527770996094, -0.433807373046875, 0.001590728759765625, -0.04279327392578125, 0.30035400390625, -0.4158935546875, -0.20351791381835938, 0.010700225830078125, -0.18730545043945312, -0.18140411376953125, -0.2761497497558594, 0.5848236083984375, 0.18651199340820312, -0.38732147216796875, 0.311309814453125, -0.21666526794433594, -0.021234512329101562, -0.4891357421875, 0.12225341796875, -0.3154296875, 0.4172210693359375, -0.72637939453125, 0.3910179138183594, -0.40639495849609375, 0.24610519409179688, -0.20230865478515625, 0.787384033203125, -0.10050773620605469, 0.6087646484375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000025.npy"}
{"epoch": 0.03779289493575208, "step": 26, "batch_size": 64, "mean": 0.09531095623970032, "std": 0.41913196444511414, "min": -0.9992218017578125, "p10": -0.35019569396972655, "median": 0.05176544189453125, "p90": 0.6143257141113282, "max": 1.42047119140625, "pos_frac": 0.515625, "sample": [0.09148406982421875, 0.2395172119140625, 0.33606719970703125, -0.1185150146484375, 0.3347816467285156, -0.07396316528320312, -0.2047119140625, -0.08346176147460938, -0.021015167236328125, 0.3323860168457031, 0.6204299926757812, 1.3558349609375, 0.157501220703125, -0.508636474609375, -0.43439483642578125, -0.9992218017578125, 0.13454437255859375, -0.33953857421875, -0.19783782958984375, -0.06509017944335938, 0.5274200439453125, -0.09891319274902344, 0.6981124877929688, -0.3547630310058594, 0.28401947021484375, 0.236907958984375, 0.3363761901855469, -0.11364555358886719, -0.10315704345703125, 0.165283203125, -0.09832382202148438, 0.26686859130859375, 0.03871917724609375, -0.00254058837890625, 0.6000823974609375, 0.3758544921875, -0.355072021484375, 0.09584617614746094, 0.2830657958984375, 0.639923095703125, 0.6773262023925781, 0.333221435546875, 0.43503761291503906, 0.0916748046875, 0.215087890625, 0.1817474365234375, -0.9137725830078125, 0.3642120361328125, 1.42047119140625, -0.15530014038085938, 0.06481170654296875, -0.298187255859375, -0.03498268127441406, -0.05637359619140625, -0.042819976806640625, 0.8125152587890625, -0.15741729736328125, -0.16677093505859375, -0.0566253662109375, -0.004878997802734375, -0.4477386474609375, -0.15652847290039062, 0.3053779602050781, -0.2884101867675781], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000026.npy"}
{"epoch": 0.039304610733182165, "step": 27, "batch_size": 64, "mean": 0.00870358943939209, "std": 0.33296385407447815, "min": -0.807647705078125, "p10": -0.36244735717773435, "median": -0.030771255493164062, "p90": 0.4071830749511719, "max": 0.81195068359375, "pos_frac": 0.46875, "sample": [0.01001739501953125, 0.15075302124023438, 0.35982513427734375, -0.17113876342773438, 0.26905059814453125, -0.5152740478515625, -0.030712127685546875, -0.0542144775390625, 0.096649169921875, -0.03643798828125, 0.4908294677734375, 0.6683807373046875, 0.32769203186035156, 0.1509075164794922, 0.3762359619140625, -0.41481781005859375, -0.14960098266601562, 0.5418701171875, -0.08814239501953125, -0.33423614501953125, -0.13271331787109375, -0.807647705078125, 0.25726318359375, 0.0042572021484375, -0.203369140625, 0.27532958984375, -0.7010345458984375, -0.20785903930664062, -0.15188217163085938, -0.03992462158203125, -0.289825439453125, 0.2822303771972656, -0.1588153839111328, 0.378753662109375, 0.44545745849609375, -0.258941650390625, -0.3517875671386719, -0.0379180908203125, 0.407501220703125, -0.12387847900390625, -0.2586822509765625, 0.275299072265625, 0.11650848388671875, 0.40644073486328125, 0.2851753234863281, -0.25757598876953125, -0.7187271118164062, 0.07047271728515625, 0.81195068359375, 0.5037193298339844, 0.13239288330078125, -0.17827606201171875, -0.2372760772705078, -0.402313232421875, 0.020416259765625, -0.02581024169921875, -0.03083038330078125, 0.35422515869140625, 0.365234375, -0.3670158386230469, 0.0423431396484375, -0.18997955322265625, -0.186798095703125, -0.206695556640625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000027.npy"}
{"epoch": 0.04081632653061224, "step": 28, "batch_size": 64, "mean": 0.017391175031661987, "std": 0.3534112870693207, "min": -0.9298171997070312, "p10": -0.45860252380371086, "median": -0.06082344055175781, "p90": 0.37886734008789064, "max": 0.9901809692382812, "pos_frac": 0.453125, "sample": [-0.22975921630859375, 0.26324462890625, -0.120086669921875, 0.37567901611328125, 0.28777313232421875, -0.058216094970703125, -0.5042266845703125, 0.7463836669921875, -0.06660842895507812, -0.256103515625, -0.10825347900390625, 0.4600868225097656, 0.9901809692382812, 0.17212677001953125, 0.3802337646484375, -0.3326835632324219, -0.001705169677734375, -0.07071304321289062, -0.6422920227050781, 0.3526611328125, -0.07720947265625, 0.3695526123046875, 0.2812461853027344, 0.088470458984375, 0.22642898559570312, 0.29608917236328125, 0.0875701904296875, 0.6571197509765625, -0.019245147705078125, 0.07283401489257812, -0.3143768310546875, 0.0658111572265625, 0.11909103393554688, 0.46764373779296875, -0.9298171997070312, 0.7025222778320312, -0.08687400817871094, -0.14113616943359375, -0.5230255126953125, 0.3392486572265625, 0.06476211547851562, -0.6108474731445312, -0.138397216796875, -0.25858306884765625, -0.16527557373046875, -0.10236549377441406, 0.3332633972167969, -0.647857666015625, 0.2516345977783203, -0.11487960815429688, 0.32578277587890625, -0.0662078857421875, 0.03285026550292969, -0.1857757568359375, -0.08056640625, -0.0634307861328125, -0.08546829223632812, -0.06676483154296875, -0.4940643310546875, -0.3758583068847656, -0.10363006591796875, 0.1702423095703125, 0.3065814971923828, -0.13177490234375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000028.npy"}
{"epoch": 0.042328042328042326, "step": 29, "batch_size": 64, "mean": 0.040889084339141846, "std": 0.4930531084537506, "min": -1.061309814453125, "p10": -0.691082763671875, "median": 0.08262252807617188, "p90": 0.5747711181640626, "max": 1.4513092041015625, "pos_frac": 0.578125, "sample": [-0.6622734069824219, 0.2997093200683594, 0.01561737060546875, 0.5548934936523438, -0.7402191162109375, 0.30817604064941406, -0.177947998046875, -0.12685775756835938, -0.6884765625, 1.4513092041015625, -0.3787841796875, -0.733551025390625, -0.01825714111328125, 0.2574653625488281, 0.2939910888671875, 0.7443008422851562, 0.1413116455078125, 0.38491058349609375, -0.828033447265625, 0.5030746459960938, -0.015539169311523438, -0.188201904296875, 0.3669776916503906, 0.09776687622070312, 0.56744384765625, 1.0009689331054688, 0.6384048461914062, -0.0372161865234375, 0.577911376953125, 0.12257003784179688, -0.23812103271484375, 0.0154876708984375, -0.1930694580078125, -0.4107398986816406, -0.69219970703125, -0.1551380157470703, 0.03927803039550781, 0.5275802612304688, 0.6324920654296875, -0.06927490234375, 0.14568328857421875, 0.23166656494140625, -0.31354522705078125, 0.42899322509765625, 0.21744537353515625, 0.05432891845703125, -0.8555870056152344, -1.061309814453125, 0.28208160400390625, 0.24322128295898438, 0.195892333984375, 0.37500762939453125, 0.29381561279296875, -0.2689056396484375, -0.45900726318359375, 0.5013542175292969, -0.837158203125, -0.3899993896484375, 0.2096099853515625, 0.9288177490234375, 0.0966796875, -0.050731658935546875, 0.06856536865234375, -0.607757568359375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000029.npy"}
{"epoch": 0.04383975812547241, "step": 30, "batch_size": 64, "mean": 0.06612381339073181, "std": 0.4809889793395996, "min": -2.226348876953125, "p10": -0.4032661437988281, "median": 0.11514759063720703, "p90": 0.5485584259033205, "max": 1.1069183349609375, "pos_frac": 0.609375, "sample": [0.6539306640625, -0.3467597961425781, 0.118896484375, 0.25720977783203125, 0.47371673583984375, -0.27875518798828125, 0.14763832092285156, 0.029296875, 0.3110809326171875, -0.41058349609375, 0.3084831237792969, 0.23291015625, 0.37255096435546875, -0.016998291015625, -0.054840087890625, 1.0072784423828125, 0.6702804565429688, -0.6617431640625, -0.15649795532226562, -0.03673553466796875, 0.217926025390625, 0.16883468627929688, -0.20824432373046875, 0.06890869140625, 1.1069183349609375, 0.22893524169921875, -0.098541259765625, 0.2310791015625, -0.23003387451171875, 0.4290351867675781, 0.8104705810546875, 0.5613574981689453, 0.11749267578125, -0.2996978759765625, 1.017059326171875, -0.3111114501953125, -0.26763153076171875, 0.11710548400878906, 0.37642669677734375, 0.22045516967773438, 0.016176223754882812, 0.49082183837890625, -2.226348876953125, 0.3467864990234375, -0.07532119750976562, -0.37422943115234375, -0.33260345458984375, -0.4206085205078125, -0.258636474609375, 0.113189697265625, 0.4163246154785156, 0.42317962646484375, 0.025783538818359375, -0.470428466796875, -0.563751220703125, 0.5186939239501953, -0.4266777038574219, 0.031463623046875, -0.134613037109375, 0.14709091186523438, 0.057094573974609375, -0.38619232177734375, 0.19332122802734375, 0.24430465698242188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000030.npy"}
{"epoch": 0.045351473922902494, "step": 31, "batch_size": 64, "mean": 0.0026363134384155273, "std": 0.42944827675819397, "min": -1.1973114013671875, "p10": -0.46548271179199213, "median": 0.05004119873046875, "p90": 0.5660255432128907, "max": 0.88787841796875, "pos_frac": 0.53125, "sample": [0.6062469482421875, 0.0489044189453125, -1.1973114013671875, -0.3328819274902344, 0.051177978515625, 0.88787841796875, -0.25870513916015625, -0.286895751953125, -0.8950042724609375, 0.5711593627929688, 0.05936431884765625, 0.065460205078125, -0.4056282043457031, -0.9141693115234375, -0.4911346435546875, -0.5506591796875, 0.05496025085449219, 0.250457763671875, 0.6290817260742188, 0.5279006958007812, -0.714263916015625, -0.19084548950195312, -0.350250244140625, 0.16480255126953125, -0.00513458251953125, -0.9867706298828125, -0.019317626953125, -0.05367279052734375, -0.29706573486328125, 0.36187744140625, -0.22479248046875, 0.449249267578125, -0.16652679443359375, 0.7439460754394531, 0.07302474975585938, -0.06875991821289062, -0.3859100341796875, 0.7575225830078125, -0.2660636901855469, 0.07592582702636719, 0.31575775146484375, 0.13482666015625, 0.059535980224609375, 0.15499114990234375, 0.29913330078125, 0.028106689453125, 0.20952606201171875, -0.000263214111328125, 0.2269134521484375, -0.39009857177734375, -0.163604736328125, 0.3233604431152344, 0.614501953125, 0.46131134033203125, 0.252960205078125, -0.13396835327148438, 0.3469963073730469, 0.554046630859375, -0.10499763488769531, 0.2604255676269531, -0.37586212158203125, 0.07273292541503906, -0.22000885009765625, -0.07477569580078125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000031.npy"}
{"epoch": 0.04686318972033258, "step": 32, "batch_size": 64, "mean": -0.011246174573898315, "std": 0.4154817759990692, "min": -1.4238815307617188, "p10": -0.4377487182617187, "median": 0.05916309356689453, "p90": 0.41459732055664067, "max": 0.6750106811523438, "pos_frac": 0.53125, "sample": [0.39827728271484375, -0.13236236572265625, -0.2893524169921875, -0.04632568359375, -0.177154541015625, 0.13498306274414062, -0.2645721435546875, 0.5230026245117188, -1.3683624267578125, 0.35707855224609375, -0.462677001953125, 0.2501373291015625, 0.19343185424804688, -0.46221160888671875, 0.3668975830078125, -0.2832794189453125, -0.4588623046875, 0.5706787109375, 0.4197540283203125, -0.22180938720703125, 0.07571601867675781, 0.11431121826171875, 0.0893096923828125, 0.35874176025390625, -0.03256988525390625, -0.2933502197265625, 0.40256500244140625, -0.176666259765625, -0.3894805908203125, 0.04261016845703125, 0.2538471221923828, 0.17893600463867188, 0.2533454895019531, -1.4238815307617188, -0.00257110595703125, -0.29657745361328125, -0.12442779541015625, 0.1024017333984375, -0.1802215576171875, 0.343109130859375, 0.15208053588867188, 0.6668472290039062, 0.20526885986328125, -0.2892875671386719, 0.151092529296875, 0.6750106811523438, -0.051631927490234375, 0.5379104614257812, 0.5506591796875, -0.13126754760742188, 0.087921142578125, -0.25936126708984375, -0.20806121826171875, 0.13261985778808594, 0.23035430908203125, -0.08066940307617188, 0.25601959228515625, 0.116119384765625, 0.2857398986816406, -0.15697479248046875, 0.02285003662109375, -0.2695465087890625, -0.45843505859375, -1.2274322509765625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000032.npy"}
{"epoch": 0.04837490551776266, "step": 33, "batch_size": 64, "mean": 0.009763926267623901, "std": 0.3930913209915161, "min": -1.1291351318359375, "p10": -0.42530250549316406, "median": 0.019324302673339844, "p90": 0.516424560546875, "max": 0.7682876586914062, "pos_frac": 0.53125, "sample": [0.7682876586914062, 0.28851318359375, -0.153594970703125, -1.1291351318359375, 0.020017623901367188, 0.105682373046875, -0.7415237426757812, 0.0186309814453125, 0.6047821044921875, 0.627532958984375, 0.5238037109375, -0.18974685668945312, -0.21480751037597656, -0.2655487060546875, 0.0414581298828125, 0.0318603515625, -0.6069488525390625, 0.24313735961914062, 0.02114105224609375, 0.1003875732421875, 0.00496673583984375, 0.0445556640625, -0.13342857360839844, -0.28131866455078125, 0.7556304931640625, 0.307769775390625, 0.25794219970703125, 0.256500244140625, 0.1236724853515625, -0.241790771484375, 0.141326904296875, 0.147613525390625, -0.1615447998046875, -0.1345996856689453, 0.16156768798828125, -0.4018440246582031, -0.10586738586425781, 0.49920654296875, 0.4275360107421875, -0.25682830810546875, -0.1607818603515625, 0.4829750061035156, -0.006591796875, -0.1114654541015625, 0.6531143188476562, 0.48895263671875, 0.39180755615234375, 0.02689361572265625, 0.2765388488769531, -0.0059356689453125, -0.3019866943359375, 0.702117919921875, -0.06332778930664062, 0.0911712646484375, -0.07448577880859375, -0.43535614013671875, -0.05054473876953125, 0.0644989013671875, -1.0994873046875, -0.4722442626953125, -0.6327095031738281, -0.111602783203125, -0.14632415771484375, -0.3853302001953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000033.npy"}
{"epoch": 0.049886621315192746, "step": 34, "batch_size": 64, "mean": 0.010291993618011475, "std": 0.4063844084739685, "min": -1.014404296875, "p10": -0.5277915954589842, "median": 0.018770217895507812, "p90": 0.5540725708007813, "max": 0.986602783203125, "pos_frac": 0.515625, "sample": [-0.0973052978515625, 0.24945831298828125, 0.47725677490234375, 0.18000411987304688, -0.18089675903320312, -0.22956085205078125, 0.209442138671875, 0.5565032958984375, -0.04931068420410156, 0.5645942687988281, -0.601959228515625, -0.430267333984375, -0.6053009033203125, -0.0433349609375, -0.1455230712890625, -0.04776763916015625, 0.096038818359375, -0.3703765869140625, -0.14355087280273438, 0.15312957763671875, 0.08206558227539062, -0.12857437133789062, -0.3404083251953125, 0.06313323974609375, -0.30571746826171875, -0.9817733764648438, 0.54840087890625, 0.014797210693359375, 0.48233795166015625, -0.10846328735351562, 0.15019607543945312, 0.6776199340820312, -0.1390819549560547, 0.02274322509765625, 0.12697601318359375, -1.014404296875, 0.02776336669921875, 0.044116973876953125, 0.24195098876953125, 0.244110107421875, -0.5695877075195312, -0.33159637451171875, -0.19384193420410156, 0.15107345581054688, -0.1487274169921875, 0.986602783203125, -0.28556060791015625, -0.6592941284179688, -0.21834564208984375, 0.09716415405273438, -0.37259864807128906, -0.2377166748046875, 0.975006103515625, -0.5990753173828125, 0.5574455261230469, 0.36003875732421875, 0.43921661376953125, 0.10991477966308594, -0.05554008483886719, 0.060306549072265625, 0.6093063354492188, 0.234130859375, 0.5411911010742188, -0.039886474609375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000034.npy"}
{"epoch": 0.05139833711262283, "step": 35, "batch_size": 64, "mean": 0.08489382266998291, "std": 0.38659369945526123, "min": -0.804412841796875, "p10": -0.3123424530029297, "median": -0.006701469421386719, "p90": 0.6453582763671876, "max": 1.511932373046875, "pos_frac": 0.5, "sample": [0.884521484375, 0.275543212890625, -0.055393218994140625, 0.620513916015625, -0.0364532470703125, 0.66632080078125, -0.02109527587890625, -0.057926177978515625, 0.068939208984375, -0.14693450927734375, 0.6983795166015625, 0.03304862976074219, 0.5226287841796875, -0.06814193725585938, 0.048828125, -0.3954505920410156, 0.2501373291015625, 0.2523956298828125, -0.385894775390625, -0.3588409423828125, -0.08831405639648438, -0.059173583984375, -0.15073394775390625, -0.18082427978515625, 0.19079208374023438, -0.30908203125, 1.0941314697265625, -0.2402496337890625, 0.19178009033203125, 0.089599609375, 0.656005859375, -0.27734375, 0.3723297119140625, -0.06747055053710938, 0.05908966064453125, -0.08420562744140625, 0.06719589233398438, -0.1503753662109375, -0.11831283569335938, -0.804412841796875, -0.1660919189453125, 0.0076923370361328125, 0.19532394409179688, 0.1985626220703125, 0.2574462890625, -0.088165283203125, 0.45442962646484375, 0.7553787231445312, -0.043117523193359375, -0.3137397766113281, 0.4162254333496094, 0.19405364990234375, -0.14352035522460938, -0.39815521240234375, 0.22048187255859375, 0.01769256591796875, 1.511932373046875, -0.286712646484375, -0.22698211669921875, 0.088287353515625, -0.07024765014648438, -0.07434844970703125, -0.3395042419433594, 0.280731201171875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000035.npy"}
{"epoch": 0.05291005291005291, "step": 36, "batch_size": 64, "mean": -0.05526161193847656, "std": 0.3993348479270935, "min": -0.9736251831054688, "p10": -0.5254753112792969, "median": -0.05037117004394531, "p90": 0.4703929901123047, "max": 0.9684066772460938, "pos_frac": 0.40625, "sample": [-0.269378662109375, -0.2458038330078125, 0.803924560546875, -0.2598876953125, -0.29833984375, -0.9736251831054688, -0.15491485595703125, 0.24036407470703125, -0.13477706909179688, -0.6851043701171875, -0.5290298461914062, 0.04441642761230469, -0.37836456298828125, -0.517181396484375, -0.20586585998535156, 0.0477294921875, -0.259796142578125, -0.4098625183105469, 0.22309494018554688, -0.808197021484375, 0.4709663391113281, 0.5006942749023438, 0.8087310791015625, 0.04720115661621094, -0.10596847534179688, 0.4292755126953125, 0.17336654663085938, -0.3863525390625, 0.3609275817871094, 0.46905517578125, 0.6062774658203125, -0.1375904083251953, -0.4586372375488281, -0.5357894897460938, -0.2368621826171875, -0.03952789306640625, -0.01445770263671875, 0.34722900390625, -0.025196075439453125, -0.26734161376953125, 0.2896842956542969, -0.78643798828125, 0.0995025634765625, -0.011400222778320312, -0.061214447021484375, 0.03737068176269531, 0.00020599365234375, 0.24052810668945312, -0.42401123046875, -0.32601165771484375, -0.19807815551757812, -0.038127899169921875, -0.4337577819824219, -0.6412734985351562, -0.21384048461914062, 0.33913421630859375, 0.9684066772460938, 0.15467453002929688, 0.0533599853515625, 0.5994987487792969, -0.00095367431640625, 0.0126953125, -0.16809844970703125, -0.26399993896484375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000036.npy"}
{"epoch": 0.05442176870748299, "step": 37, "batch_size": 64, "mean": -0.05224302411079407, "std": 0.36881133913993835, "min": -0.9727325439453125, "p10": -0.5679634094238281, "median": -0.0327606201171875, "p90": 0.3679336547851565, "max": 1.033660888671875, "pos_frac": 0.484375, "sample": [0.26169586181640625, 0.3130645751953125, 0.1569671630859375, -0.48055267333984375, -0.0989227294921875, 0.4188232421875, -0.14943313598632812, -0.0367584228515625, -0.09861946105957031, -0.1729583740234375, -0.687347412109375, -0.173126220703125, 0.31201171875, -0.9727325439453125, -0.537933349609375, -0.09008026123046875, -0.1791858673095703, 0.24630355834960938, 0.13560867309570312, -0.136627197265625, 0.4538993835449219, -0.22505950927734375, 0.115142822265625, 0.14031219482421875, -0.8852958679199219, -0.0287628173828125, 0.49853515625, -0.3092613220214844, -0.298370361328125, 0.037841796875, 0.0012054443359375, -0.2539520263671875, 0.21506500244140625, -0.05405426025390625, 0.165008544921875, -0.6481475830078125, 0.0018463134765625, 0.11794281005859375, 0.2712554931640625, 0.189483642578125, 0.4273681640625, 0.046123504638671875, 0.031742095947265625, -0.29920196533203125, 0.391448974609375, -0.24261474609375, -0.4019622802734375, -0.205291748046875, -0.5808334350585938, -0.40447998046875, -0.6640892028808594, 0.2557544708251953, -0.2233867645263672, 0.2451763153076172, 0.5231399536132812, 0.1726226806640625, -0.29993438720703125, 0.1302947998046875, 1.033660888671875, 0.1895294189453125, -0.1954193115234375, -0.48030853271484375, 0.2817115783691406, -0.60943603515625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000037.npy"}
{"epoch": 0.055933484504913075, "step": 38, "batch_size": 64, "mean": 0.0010806024074554443, "std": 0.4315952956676483, "min": -1.314666748046875, "p10": -0.43553543090820307, "median": -0.046057701110839844, "p90": 0.5702854156494142, "max": 1.3202285766601562, "pos_frac": 0.421875, "sample": [0.05556488037109375, -0.1495819091796875, 0.5606002807617188, -0.11718368530273438, 0.07798004150390625, 0.0327301025390625, 0.3100624084472656, 0.3168983459472656, 0.47754669189453125, -0.3245391845703125, -0.11178970336914062, -0.16126251220703125, 0.06856346130371094, -0.056537628173828125, -0.351287841796875, 0.5565185546875, -0.546600341796875, -0.05013847351074219, -0.19385528564453125, 0.9513015747070312, 0.1387176513671875, 0.6770477294921875, -0.5519638061523438, 0.7787628173828125, -0.001697540283203125, -0.1449432373046875, -0.34063720703125, -0.13570594787597656, 0.3406982421875, -0.28913116455078125, -0.23600387573242188, -0.7760467529296875, -0.14116287231445312, 0.476654052734375, -0.16709136962890625, -0.6554489135742188, -0.2946014404296875, -0.17333030700683594, -0.3630104064941406, 0.01163482666015625, -0.36394500732421875, -1.314666748046875, -0.08232688903808594, -0.0419769287109375, 0.4490966796875, -0.243927001953125, -0.0206146240234375, 0.043834686279296875, 0.39386749267578125, 0.602569580078125, -0.466217041015625, -0.5900650024414062, 1.3202285766601562, -0.18605804443359375, 0.08402061462402344, -0.17858314514160156, -0.03232574462890625, 0.5744361877441406, 0.00278472900390625, 0.10724639892578125, -0.3560218811035156, 0.6160888671875, -0.011142730712890625, 0.2651252746582031], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000038.npy"}
{"epoch": 0.05744520030234316, "step": 39, "batch_size": 64, "mean": 0.014002442359924316, "std": 0.32431039214134216, "min": -0.8652496337890625, "p10": -0.38655433654785154, "median": -0.020891189575195312, "p90": 0.4554737091064454, "max": 0.6943359375, "pos_frac": 0.484375, "sample": [0.6078414916992188, -0.3905448913574219, 0.26432037353515625, 0.12788772583007812, -0.08946990966796875, 0.56756591796875, -0.3772430419921875, 0.38903045654296875, -0.26163482666015625, 0.18896865844726562, -0.302825927734375, -0.21828460693359375, 0.04427146911621094, -0.03472328186035156, 0.2034149169921875, 0.1707611083984375, 0.1144866943359375, -0.0518341064453125, -0.1878509521484375, 0.01287841796875, -0.0572662353515625, 0.072845458984375, -0.1852741241455078, 0.027133941650390625, 0.468109130859375, 0.48122406005859375, -0.18804931640625, -0.7003955841064453, -0.1381683349609375, 0.23785781860351562, -0.08179855346679688, 0.1777801513671875, 0.6943359375, -0.3996315002441406, -0.4186363220214844, -0.07205581665039062, -0.6224632263183594, -0.519775390625, 0.4259910583496094, 0.25982666015625, 0.5249176025390625, -0.038787841796875, -0.23618507385253906, 0.3563079833984375, 0.0030517578125, -0.110076904296875, 0.3476715087890625, 0.0844879150390625, -0.022365570068359375, 0.14188766479492188, 0.6326065063476562, -0.01941680908203125, -0.10871124267578125, -0.15258407592773438, -0.11163711547851562, -0.8652496337890625, -0.22296524047851562, 0.381256103515625, -0.1774139404296875, 0.30890655517578125, -0.17858505249023438, 0.2007160186767578, 0.14098358154296875, -0.22126388549804688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000039.npy"}
{"epoch": 0.05895691609977324, "step": 40, "batch_size": 64, "mean": 0.07039055228233337, "std": 0.32655420899391174, "min": -0.6977195739746094, "p10": -0.30328369140625, "median": 0.05231666564941406, "p90": 0.45368423461914065, "max": 0.940826416015625, "pos_frac": 0.578125, "sample": [-0.5525283813476562, 0.43979644775390625, 0.45096588134765625, -0.37835693359375, 0.2989463806152344, 0.20431900024414062, 0.67730712890625, 0.44427490234375, 0.0312042236328125, -0.6977195739746094, 0.5665626525878906, -0.018917083740234375, 0.17945098876953125, 0.06716156005859375, -0.00928497314453125, 0.27307891845703125, 0.23516845703125, 0.940826416015625, -0.06382560729980469, -0.056186676025390625, 0.11411285400390625, 0.269775390625, -0.5909576416015625, -0.16498565673828125, 0.5890960693359375, 0.469879150390625, 0.32886505126953125, 0.21334075927734375, 0.6182632446289062, 0.31961822509765625, 0.06158638000488281, 0.024322509765625, -0.03656005859375, -0.29628753662109375, -0.13432693481445312, 0.015569686889648438, -0.13206863403320312, -0.22008132934570312, -0.22716522216796875, 0.3884735107421875, -0.11510086059570312, -0.103057861328125, 0.15073776245117188, 0.170440673828125, -0.24094390869140625, 0.09088897705078125, 0.04304695129394531, -0.27728271484375, -0.4969673156738281, 0.3820648193359375, 0.4548492431640625, -0.0468597412109375, -0.23436355590820312, 0.1035614013671875, -0.14263916015625, 0.00128173828125, 0.3016815185546875, 0.22245025634765625, -0.29156494140625, -0.3799591064453125, 0.27570533752441406, 0.3731346130371094, -0.072540283203125, -0.30628204345703125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000040.npy"}
{"epoch": 0.06046863189720333, "step": 41, "batch_size": 64, "mean": 0.08602365851402283, "std": 0.3597239553928375, "min": -0.6973953247070312, "p10": -0.3466476440429687, "median": 0.06833839416503906, "p90": 0.543267822265625, "max": 1.0122146606445312, "pos_frac": 0.59375, "sample": [-0.0108642578125, 0.27254486083984375, 0.1324291229248047, 0.0547943115234375, 0.0396270751953125, 0.0794677734375, -0.00496673583984375, 0.23467254638671875, 0.256805419921875, 0.07410430908203125, -0.01764678955078125, -0.05152320861816406, -0.01062774658203125, 0.3055267333984375, 0.251953125, 0.6456756591796875, 0.20277786254882812, 0.045196533203125, 0.107269287109375, 0.06257247924804688, 0.1174774169921875, -0.0623779296875, -0.24858856201171875, -0.188201904296875, 0.32265663146972656, 0.15545654296875, -0.44770050048828125, -0.11276435852050781, -0.24604034423828125, 0.2504730224609375, -0.4989814758300781, -0.6408538818359375, 0.6435699462890625, 0.38149261474609375, 0.23891067504882812, 0.0900115966796875, 0.26671600341796875, -0.4021415710449219, -0.2985801696777344, 0.00775146484375, 0.8623237609863281, 0.05414581298828125, -0.316558837890625, -0.2087078094482422, 0.536773681640625, -0.29077911376953125, 1.0122146606445312, 0.46575927734375, 0.498291015625, -0.1847381591796875, -0.3595428466796875, 0.546051025390625, -0.1325531005859375, -0.6973953247070312, 0.48770904541015625, -0.30682373046875, 0.2188873291015625, -0.061275482177734375, 0.08408737182617188, -0.5545654296875, 0.7517852783203125, 0.762359619140625, 0.3743934631347656, -0.03440093994140625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000041.npy"}
{"epoch": 0.06198034769463341, "step": 42, "batch_size": 64, "mean": 0.07593908905982971, "std": 0.40648043155670166, "min": -0.6785240173339844, "p10": -0.4313508987426758, "median": 0.09351158142089844, "p90": 0.5818695068359376, "max": 1.11572265625, "pos_frac": 0.578125, "sample": [0.4966278076171875, 0.27398681640625, -0.05548095703125, 0.20972442626953125, 0.14745712280273438, 0.3786773681640625, -0.4451751708984375, 0.13478660583496094, -0.02867889404296875, -0.3515663146972656, -0.12728118896484375, 0.29294586181640625, 0.11288833618164062, -0.4990692138671875, -0.00933837890625, 0.13363265991210938, 0.44488525390625, 0.1820068359375, -0.09798812866210938, -0.29312896728515625, 0.2271270751953125, 0.26892852783203125, 0.5292205810546875, -0.22841262817382812, -0.4324798583984375, 1.1004486083984375, -0.14128494262695312, -0.6785240173339844, 0.12369918823242188, -0.2528953552246094, 0.0688018798828125, 0.20267486572265625, -0.42871665954589844, -0.33495330810546875, -0.4511528015136719, -0.074462890625, 0.0067291259765625, -0.22979736328125, 0.1466827392578125, 1.11572265625, 0.12831497192382812, 0.2861785888671875, 0.53955078125, 0.6794204711914062, -0.247406005859375, 0.0024261474609375, 0.343414306640625, 1.0974273681640625, -0.2426605224609375, -0.41594696044921875, -0.3427886962890625, -0.4863128662109375, 0.600006103515625, 0.15131378173828125, -0.3977241516113281, 0.7165069580078125, 0.07413482666015625, 0.1286487579345703, -0.5070419311523438, 0.2602081298828125, 0.046260833740234375, 0.2803230285644531, 0.868621826171875, -0.14004135131835938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000042.npy"}
{"epoch": 0.06349206349206349, "step": 43, "batch_size": 64, "mean": 0.010380983352661133, "std": 0.4331362545490265, "min": -1.403564453125, "p10": -0.43509521484374997, "median": 0.022979736328125, "p90": 0.43646335601806646, "max": 1.5010528564453125, "pos_frac": 0.515625, "sample": [0.0588226318359375, 0.24249649047851562, -0.45977783203125, -0.07136917114257812, -0.3771781921386719, 0.3652229309082031, -0.2995948791503906, 0.039890289306640625, 0.0319061279296875, -0.15984344482421875, 0.14484024047851562, -0.14066314697265625, -0.112274169921875, 0.06252288818359375, 0.12493896484375, 0.224517822265625, 0.40090179443359375, -0.3582916259765625, -0.0919952392578125, 0.06447601318359375, 0.050159454345703125, 0.403472900390625, -0.5011138916015625, -0.305999755859375, -0.1993560791015625, 0.6473503112792969, -0.04520606994628906, 0.049102783203125, 0.4295539855957031, -0.300445556640625, 1.5010528564453125, -0.729339599609375, -0.3018646240234375, 0.3882598876953125, -0.24234771728515625, 1.150421142578125, -0.0440216064453125, 0.32364654541015625, 0.4056968688964844, 0.4565582275390625, 0.4685783386230469, -0.06705093383789062, 0.071746826171875, -0.8511695861816406, -0.1419506072998047, -0.190185546875, -0.6825790405273438, -0.2621288299560547, -0.454376220703125, 0.23997879028320312, 0.4394245147705078, 0.27520751953125, -0.07418060302734375, 0.23932647705078125, 0.324066162109375, -0.2030029296875, 0.19368743896484375, 0.5141754150390625, 0.0140533447265625, -1.403564453125, -0.21905517578125, 0.10076522827148438, -0.10240554809570312, -0.390106201171875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000043.npy"}
{"epoch": 0.06500377928949358, "step": 44, "batch_size": 64, "mean": -0.007240712642669678, "std": 0.44536638259887695, "min": -1.28131103515625, "p10": -0.5095191955566406, "median": 0.027545928955078125, "p90": 0.5425643920898438, "max": 1.0883560180664062, "pos_frac": 0.515625, "sample": [0.14850234985351562, 0.5759468078613281, -0.38741302490234375, 0.2327728271484375, -0.31256103515625, -1.28131103515625, -0.07865142822265625, -0.06533050537109375, -0.5009384155273438, 0.165985107421875, 0.3364295959472656, -0.178070068359375, 0.219970703125, 0.29209327697753906, -0.280792236328125, 0.19775390625, -0.026409149169921875, -0.6141357421875, -0.41727447509765625, -0.38079261779785156, 0.04929351806640625, -0.2958946228027344, 0.09549331665039062, 0.0664825439453125, 0.06740188598632812, 0.198028564453125, 0.3211822509765625, 0.19594573974609375, -0.6986083984375, 0.15024566650390625, 0.727020263671875, -0.3465118408203125, -0.5023040771484375, -0.41922760009765625, 0.08875083923339844, 0.28495025634765625, -0.5126113891601562, 0.3387336730957031, 0.335968017578125, 0.5863571166992188, -0.2505531311035156, -0.27396392822265625, -0.8431396484375, -0.05904388427734375, 0.9301986694335938, 0.1166839599609375, 0.539093017578125, 0.22037506103515625, -0.0957794189453125, -0.25783348083496094, -0.16126251220703125, -0.206085205078125, -0.42462921142578125, -0.0093841552734375, -0.8618011474609375, 0.4537200927734375, 1.0883560180664062, 1.0014801025390625, 0.5440521240234375, -0.5388870239257812, 0.00579833984375, 0.112579345703125, 0.265106201171875, -0.13495635986328125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000044.npy"}
{"epoch": 0.06651549508692366, "step": 45, "batch_size": 64, "mean": 0.05561518669128418, "std": 0.44790127873420715, "min": -1.10693359375, "p10": -0.4658470153808594, "median": 0.05726432800292969, "p90": 0.5446853637695315, "max": 1.4740219116210938, "pos_frac": 0.5625, "sample": [0.07899093627929688, 0.2994537353515625, -0.46416473388671875, 0.1488494873046875, -0.3816547393798828, 0.9432220458984375, 0.5838432312011719, 1.4740219116210938, -0.4625244140625, -0.23865509033203125, -0.2546539306640625, 0.26731109619140625, -0.6173782348632812, 0.13022613525390625, -0.2804546356201172, -0.5738372802734375, -0.09525680541992188, 0.259521484375, -0.260589599609375, 0.04943084716796875, 1.005767822265625, 0.2342987060546875, -0.15216445922851562, 0.15214157104492188, 0.15005874633789062, 0.01434326171875, -0.35601806640625, 0.044498443603515625, -0.20705413818359375, -0.319976806640625, 0.22723770141601562, -1.10693359375, -0.608612060546875, -0.040802001953125, 0.06509780883789062, 0.0172882080078125, -0.1095733642578125, 0.14540863037109375, 0.20465850830078125, 0.436798095703125, 0.47733306884765625, 0.4948577880859375, -0.06414031982421875, 0.713714599609375, 0.14469146728515625, 0.22431564331054688, -0.016937255859375, 0.2983856201171875, -0.766021728515625, 0.40261077880859375, -0.4665679931640625, -0.0189666748046875, 0.14073944091796875, 0.3674468994140625, 0.46856689453125, -0.035968780517578125, 0.79949951171875, 0.267303466796875, -0.1707439422607422, -0.44792747497558594, -0.0287017822265625, 0.5660400390625, 0.4939422607421875, -0.6862640380859375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000045.npy"}
{"epoch": 0.06802721088435375, "step": 46, "batch_size": 64, "mean": 0.01664254069328308, "std": 0.35603371262550354, "min": -0.823638916015625, "p10": -0.42491607666015624, "median": -0.000919342041015625, "p90": 0.43422927856445315, "max": 0.93121337890625, "pos_frac": 0.5, "sample": [-0.13927459716796875, -0.385498046875, -0.1770477294921875, -0.823638916015625, 0.612030029296875, 0.76019287109375, -0.00199127197265625, 0.23233413696289062, 0.27182769775390625, 0.1167449951171875, 0.13760757446289062, -0.07585906982421875, 0.259735107421875, 0.20578765869140625, 0.0912628173828125, -0.019500732421875, -0.5061569213867188, 0.34476470947265625, -0.041294097900390625, 0.0281524658203125, 0.04556846618652344, 0.43316650390625, -0.2363739013671875, -0.1738433837890625, -0.26023101806640625, -0.13436126708984375, 0.93121337890625, -0.13700103759765625, 0.3912200927734375, -0.4586181640625, 0.02587890625, 0.5754165649414062, 0.076507568359375, -0.32602691650390625, -0.20934295654296875, -0.21142578125, -0.002147674560546875, -0.3893585205078125, -0.483734130859375, 0.0513916015625, -0.27443695068359375, 0.3695526123046875, 0.43944549560546875, 0.8018798828125, 0.2850227355957031, 0.3302726745605469, 0.2717094421386719, -0.24304962158203125, -0.11487960815429688, 0.43468475341796875, 0.052852630615234375, 0.35605621337890625, -0.29955482482910156, -0.644073486328125, 0.000152587890625, -0.26013946533203125, -0.6346969604492188, -0.440155029296875, -0.14241790771484375, 0.28253173828125, 0.32513427734375, 0.0019321441650390625, -0.08295822143554688, -0.14781951904296875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000046.npy"}
{"epoch": 0.06953892668178382, "step": 47, "batch_size": 64, "mean": 0.0484640896320343, "std": 0.32644423842430115, "min": -0.580902099609375, "p10": -0.3609046936035156, "median": 0.041207313537597656, "p90": 0.4852874755859376, "max": 1.0329437255859375, "pos_frac": 0.515625, "sample": [0.3083972930908203, -0.1619873046875, -0.3635520935058594, 0.14225006103515625, 0.49359130859375, 0.24863052368164062, -0.23692893981933594, -0.06501007080078125, -0.30310630798339844, 0.209228515625, -0.06661796569824219, -0.10933303833007812, 0.095977783203125, 0.5299758911132812, -0.28899383544921875, 0.13257980346679688, 0.24017333984375, -0.2487945556640625, 0.3485870361328125, 0.22168350219726562, -0.10137557983398438, -0.04718780517578125, 0.5538330078125, 0.3499641418457031, 0.05610847473144531, -0.0157623291015625, 0.3070487976074219, -0.09039497375488281, -0.0214691162109375, -0.13609886169433594, -0.40174102783203125, 0.3905448913574219, 0.22186279296875, -0.23818206787109375, 0.34906005859375, -0.354339599609375, 0.185791015625, 0.506011962890625, -0.114013671875, -0.48847198486328125, 0.3145294189453125, 0.465911865234375, 0.27793121337890625, 0.02630615234375, 0.12591934204101562, -0.3756828308105469, -0.3988037109375, -0.3064231872558594, 0.3634490966796875, -0.09722900390625, -0.3613739013671875, -0.35980987548828125, 1.0329437255859375, 0.5668411254882812, 0.08274078369140625, 0.08458709716796875, 0.09505844116210938, -0.10644912719726562, -0.3543243408203125, -0.18746185302734375, -0.26351165771484375, 0.5650634765625, -0.580902099609375, 0.4544525146484375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000047.npy"}
{"epoch": 0.0710506424792139, "step": 48, "batch_size": 64, "mean": 0.07357525825500488, "std": 0.3125082552433014, "min": -0.5700263977050781, "p10": -0.3271953582763672, "median": 0.09590530395507812, "p90": 0.4513046264648438, "max": 0.9339599609375, "pos_frac": 0.625, "sample": [0.034664154052734375, -0.1966094970703125, 0.09458160400390625, 0.123443603515625, 0.043212890625, 0.5574493408203125, -0.5489654541015625, -0.3218841552734375, -0.20212936401367188, 0.3121681213378906, -0.10916900634765625, 0.09966278076171875, 0.4410400390625, 0.2352752685546875, 0.9339599609375, 0.6238861083984375, -0.2368316650390625, 0.09722900390625, -0.222686767578125, 0.24393463134765625, -0.09414291381835938, 0.12530517578125, 0.12848663330078125, 0.3037567138671875, -0.28556060791015625, 0.48407745361328125, 0.16677093505859375, 0.3489227294921875, -0.5700263977050781, -0.20941162109375, -0.40444183349609375, 0.2231292724609375, -0.33899688720703125, 0.6655464172363281, 0.18801116943359375, 0.18938446044921875, -0.24161911010742188, 0.17824554443359375, -0.05161285400390625, 0.09229278564453125, 0.047031402587890625, 0.33124542236328125, -0.1611785888671875, 0.12098121643066406, 0.9154586791992188, 0.045177459716796875, 0.27822113037109375, 0.4557037353515625, 0.0453033447265625, -0.0136871337890625, 0.2036285400390625, -0.0505828857421875, -0.3294715881347656, 0.2618141174316406, -0.15330886840820312, 0.21056747436523438, -0.0665740966796875, -0.08710479736328125, -0.3891143798828125, -0.4649162292480469, 0.2352275848388672, 0.05594635009765625, 0.15308380126953125, 0.16501617431640625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000048.npy"}
{"epoch": 0.07256235827664399, "step": 49, "batch_size": 64, "mean": -0.005153149366378784, "std": 0.3522629737854004, "min": -0.7711105346679688, "p10": -0.43271274566650386, "median": -0.00278472900390625, "p90": 0.4378768920898438, "max": 0.9855880737304688, "pos_frac": 0.484375, "sample": [0.44545745849609375, -0.15590667724609375, -0.12203788757324219, -0.3984222412109375, 0.22259521484375, 0.2793235778808594, 0.20079421997070312, -0.29915618896484375, 0.15167236328125, 0.39910125732421875, 0.18507003784179688, 0.00568389892578125, -0.5947723388671875, 0.3671417236328125, -0.2131366729736328, -0.44740867614746094, 0.029376983642578125, -0.69427490234375, -0.00070953369140625, -0.32098388671875, -0.10031318664550781, 0.022125244140625, -0.023956298828125, -0.6718368530273438, -0.14479827880859375, -0.248321533203125, 0.444244384765625, 0.024059295654296875, -0.034961700439453125, -0.09844207763671875, -0.032352447509765625, -0.23218154907226562, 0.13305282592773438, 0.4230194091796875, 0.3743324279785156, 0.836669921875, 0.0377960205078125, -0.7264480590820312, -0.176361083984375, -0.1360015869140625, -0.2853546142578125, 0.4709434509277344, 0.49765777587890625, 0.5679931640625, 0.1578826904296875, 0.00350189208984375, -0.17099761962890625, -0.035404205322265625, 0.20545578002929688, 0.10859870910644531, 0.9855880737304688, -0.08294677734375, 0.137115478515625, 0.18930816650390625, -0.350982666015625, -0.7711105346679688, -0.6001319885253906, 0.0321807861328125, -0.21982574462890625, -0.00485992431640625, -0.01885986328125, 0.2653961181640625, -0.29160308837890625, 0.1719207763671875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000049.npy"}
{"epoch": 0.07407407407407407, "step": 50, "batch_size": 64, "mean": 0.008985787630081177, "std": 0.3481454849243164, "min": -0.948883056640625, "p10": -0.35544433593749997, "median": 0.007982254028320312, "p90": 0.3846111297607425, "max": 1.4087982177734375, "pos_frac": 0.5, "sample": [-0.1202850341796875, -0.22552490234375, -0.1471099853515625, -0.22716140747070312, -0.12431716918945312, -0.47846221923828125, -0.034290313720703125, -0.5573654174804688, -0.2034454345703125, 0.2582855224609375, 0.017398834228515625, 0.8104019165039062, -0.6251678466796875, -0.049785614013671875, -0.2748260498046875, -0.22597694396972656, 0.48355865478515625, 0.12984848022460938, 0.2046661376953125, 0.4169464111328125, -0.2897491455078125, -0.20436477661132812, 0.09746932983398438, 1.4087982177734375, 0.0818328857421875, -0.948883056640625, -0.27758026123046875, 0.09862136840820312, -0.23383712768554688, -0.0202484130859375, 0.4891319274902344, -0.001434326171875, 0.4407176971435547, -0.00229644775390625, 0.26287841796875, -0.3076019287109375, -0.33587646484375, 0.12592315673828125, -0.0074634552001953125, 0.032073974609375, 0.114166259765625, -0.21318817138671875, 0.26453399658203125, 0.12863922119140625, -0.014875411987304688, 0.1726245880126953, 0.23522186279296875, 0.0654449462890625, 0.0751190185546875, -0.3950347900390625, 0.25875091552734375, 0.06919097900390625, 0.2619819641113281, 0.3091621398925781, 0.25043296813964844, -0.15471649169921875, -0.11471939086914062, 0.5571479797363281, 0.074554443359375, 0.04747962951660156, -0.36383056640625, -0.16316986083984375, 0.11835479736328125, -0.4436798095703125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000050.npy"}
{"epoch": 0.07558578987150416, "step": 51, "batch_size": 64, "mean": -0.026766330003738403, "std": 0.420632541179657, "min": -1.2001495361328125, "p10": -0.5586616516113281, "median": 0.006587028503417969, "p90": 0.4804901123046876, "max": 0.9294815063476562, "pos_frac": 0.515625, "sample": [0.4876861572265625, 0.3694648742675781, 0.22721290588378906, -0.4915657043457031, -0.39926910400390625, 0.9294815063476562, 0.3530693054199219, 0.050319671630859375, 0.2584705352783203, -0.0086822509765625, 0.0049800872802734375, -0.512481689453125, 0.31072998046875, -0.033660888671875, -0.10990142822265625, -0.441802978515625, 0.531707763671875, 0.2130279541015625, 0.2918968200683594, -0.01007080078125, -0.41620635986328125, -0.23722457885742188, 0.05600738525390625, -0.969390869140625, -0.7389755249023438, 0.0737762451171875, -0.25316619873046875, -0.6893672943115234, 0.25237464904785156, -0.4863739013671875, 0.5736160278320312, -0.2034149169921875, 0.06821441650390625, -1.2001495361328125, -0.6048355102539062, 0.3300437927246094, 0.5876846313476562, -0.3679161071777344, -0.1457977294921875, -0.29351806640625, 0.41802215576171875, -0.41448211669921875, -0.12750244140625, 0.37127685546875, 0.25031280517578125, 0.2507781982421875, 0.5442581176757812, -0.509002685546875, -0.268707275390625, 0.0081939697265625, 0.24062728881835938, 0.1922149658203125, 0.1997528076171875, 0.1748046875, -0.1912689208984375, -0.16048431396484375, -0.614532470703125, 0.19945144653320312, 0.4636993408203125, -0.5784530639648438, -0.11981964111328125, 0.4633026123046875, -0.36351776123046875, 0.5020370483398438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000051.npy"}
{"epoch": 0.07709750566893424, "step": 52, "batch_size": 64, "mean": 0.023045867681503296, "std": 0.46363016963005066, "min": -1.2257461547851562, "p10": -0.5340370178222656, "median": 0.08414077758789062, "p90": 0.5120662689208985, "max": 1.386199951171875, "pos_frac": 0.59375, "sample": [0.14293670654296875, -0.9937744140625, 0.4801025390625, 0.6403656005859375, -0.4474639892578125, 0.09818267822265625, -0.0589447021484375, 0.047924041748046875, 0.17535400390625, 0.9262466430664062, 0.069000244140625, 0.30730438232421875, -0.45989227294921875, 0.30817413330078125, 0.2570648193359375, -0.6202545166015625, 0.287750244140625, -0.8760223388671875, -0.2687263488769531, 0.17291259765625, -0.2772674560546875, -0.19894790649414062, 0.49953460693359375, 0.08354568481445312, 1.386199951171875, -0.5439682006835938, 0.3880767822265625, 0.14238929748535156, -0.181396484375, 0.026073455810546875, 0.04076385498046875, 0.3411712646484375, 0.97210693359375, 0.5174369812011719, 0.37945556640625, 0.1176605224609375, -0.09362030029296875, 0.012510299682617188, -0.5254974365234375, 0.8241424560546875, 0.1958332061767578, -0.22631454467773438, -0.19991302490234375, 0.3958740234375, -0.3547706604003906, 0.15035247802734375, 0.08473587036132812, -0.43036651611328125, -0.17529296875, -1.2257461547851562, -0.37448883056640625, -0.342620849609375, 0.16454315185546875, 0.23201751708984375, -0.5376968383789062, 0.1277008056640625, 0.67706298828125, 0.24691009521484375, 0.09246826171875, 0.24694252014160156, -0.615447998046875, -0.3459053039550781, -0.25969505310058594, -0.1498565673828125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000052.npy"}
{"epoch": 0.07860922146636433, "step": 53, "batch_size": 64, "mean": 0.1007724404335022, "std": 0.46119046211242676, "min": -0.9241180419921875, "p10": -0.533771514892578, "median": 0.10037612915039062, "p90": 0.6437255859375, "max": 1.16943359375, "pos_frac": 0.625, "sample": [0.20113372802734375, 0.5813446044921875, 0.514373779296875, -0.309234619140625, 0.14883995056152344, -0.1596527099609375, -0.02858734130859375, 0.0564117431640625, -0.18285751342773438, -0.2860527038574219, -0.5995101928710938, 0.6441192626953125, 0.5620956420898438, 0.176177978515625, -0.131134033203125, 0.16621017456054688, 0.0660247802734375, 0.0027313232421875, 0.06222724914550781, 0.5484542846679688, 0.06357192993164062, 0.14241409301757812, 0.42780303955078125, 0.09647369384765625, -0.35915374755859375, 0.6488571166992188, -0.020566940307617188, 0.104278564453125, 0.2541046142578125, 0.5069541931152344, -0.1162872314453125, -0.5836563110351562, 0.012176513671875, -0.9241180419921875, 0.11194419860839844, 1.0807037353515625, 0.11853408813476562, 0.4529457092285156, -0.6843643188476562, 0.5436439514160156, -0.041545867919921875, 0.5198135375976562, 0.17364883422851562, 0.6428070068359375, 0.53472900390625, 0.822357177734375, -0.4173736572265625, -0.7206497192382812, -0.2698097229003906, -0.7269134521484375, 1.16943359375, -0.3150978088378906, 0.3118877410888672, 0.034027099609375, -0.7739715576171875, -0.395721435546875, 1.1546630859375, 0.7390327453613281, 0.23150634765625, -0.19834136962890625, -0.14635848999023438, 0.241485595703125, -0.19469833374023438, 0.16515159606933594], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000053.npy"}
{"epoch": 0.0801209372637944, "step": 54, "batch_size": 64, "mean": 0.03820416331291199, "std": 0.4379441738128662, "min": -1.273162841796875, "p10": -0.5264419555664062, "median": 0.13230514526367188, "p90": 0.536385726928711, "max": 0.7484893798828125, "pos_frac": 0.609375, "sample": [-0.9993133544921875, 0.2851715087890625, 0.143890380859375, -0.08470726013183594, 0.5148963928222656, 0.3058204650878906, 0.22198104858398438, 0.32653045654296875, 0.1370391845703125, 0.11721038818359375, 0.5260543823242188, -0.09257125854492188, 0.201385498046875, -0.10166168212890625, 0.2660560607910156, -0.09346199035644531, 0.5082550048828125, 0.2667827606201172, -0.33246612548828125, 0.0844573974609375, 0.25516510009765625, 0.5976409912109375, -0.16999244689941406, 0.4087333679199219, -0.7407684326171875, 0.5468597412109375, 0.037303924560546875, -0.10524559020996094, -0.363037109375, 0.33875274658203125, -0.82415771484375, 0.5408134460449219, -0.20738983154296875, 0.2540569305419922, 0.5478916168212891, 0.01044464111328125, 0.2438201904296875, 0.7484893798828125, -0.03397369384765625, 0.4834442138671875, -0.23537826538085938, -0.11885833740234375, -0.44305419921875, 0.138458251953125, 0.20647430419921875, 0.5255546569824219, 0.645233154296875, 0.15104293823242188, -0.03645515441894531, 0.3553123474121094, 0.12757110595703125, -0.491973876953125, 0.14188385009765625, -0.03095245361328125, 0.1805706024169922, -0.14541053771972656, -0.48075294494628906, 0.060146331787109375, 0.6985855102539062, -0.5412139892578125, -0.92864990234375, -1.273162841796875, -0.9366378784179688, 0.10653305053710938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000054.npy"}
{"epoch": 0.08163265306122448, "step": 55, "batch_size": 64, "mean": 0.20346179604530334, "std": 0.468463271856308, "min": -0.8526763916015625, "p10": -0.3287948608398437, "median": 0.1666851043701172, "p90": 0.7205886840820315, "max": 1.70989990234375, "pos_frac": 0.609375, "sample": [0.7478256225585938, -0.15264892578125, 0.45587921142578125, -0.30794525146484375, 0.227813720703125, -0.340118408203125, -0.09026718139648438, -0.1968841552734375, -0.10406494140625, -0.33773040771484375, 0.24466705322265625, 1.07989501953125, 0.0619049072265625, 0.624847412109375, 0.39378929138183594, -0.0552520751953125, 0.3882408142089844, 0.9994430541992188, 0.620391845703125, -0.10955047607421875, -0.2510223388671875, -0.034759521484375, 0.62518310546875, 0.00217437744140625, 1.70989990234375, 0.28168487548828125, -0.08074569702148438, 0.5545806884765625, 0.586212158203125, 0.608306884765625, -0.01165771484375, 0.4455108642578125, 0.05220794677734375, 0.39624786376953125, -0.8526763916015625, 0.16364288330078125, -0.042034149169921875, -0.11504936218261719, 0.6570358276367188, 0.21087265014648438, 0.196929931640625, 0.16972732543945312, -0.028682708740234375, 0.639617919921875, 0.548919677734375, 0.08260536193847656, 0.9270706176757812, -0.215911865234375, 0.054595947265625, -0.745697021484375, -0.12604522705078125, 0.20345306396484375, -0.191986083984375, -0.555328369140625, 0.08386993408203125, 0.7964515686035156, 0.38720703125, -0.103729248046875, 1.2362022399902344, -0.3949737548828125, 0.5981597900390625, 0.58941650390625, 0.21818161010742188, -0.40435028076171875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000055.npy"}
{"epoch": 0.08314436885865457, "step": 56, "batch_size": 64, "mean": 0.08395865559577942, "std": 0.44815248250961304, "min": -1.2030181884765625, "p10": -0.39792385101318356, "median": 0.07049942016601562, "p90": 0.6230934143066409, "max": 1.5000762939453125, "pos_frac": 0.5625, "sample": [-0.6327743530273438, 0.0764923095703125, 0.04483985900878906, 0.20719146728515625, 0.333648681640625, -0.06948089599609375, 0.21373748779296875, -0.7066497802734375, -0.12168121337890625, -0.06900978088378906, -0.011199951171875, -0.1251373291015625, 1.5000762939453125, -0.53009033203125, -0.08350372314453125, 0.3002815246582031, 0.5660400390625, 0.34468841552734375, -0.09403228759765625, 0.3500709533691406, -0.3481884002685547, 0.9373703002929688, -0.0036525726318359375, 0.8792572021484375, -0.14541244506835938, -0.13409805297851562, 0.15183258056640625, 0.2809906005859375, 0.14028167724609375, -0.23239898681640625, 0.06450653076171875, 0.47760772705078125, -0.04328155517578125, 0.4116630554199219, 0.1157379150390625, -0.1692352294921875, -1.1370391845703125, 0.2922630310058594, -0.41050148010253906, 0.169097900390625, 0.16158676147460938, 0.05675506591796875, -0.1484527587890625, -0.4109039306640625, 0.017293930053710938, 0.4767303466796875, -0.3685760498046875, 0.6475448608398438, 0.19744300842285156, -0.0170135498046875, 0.32601165771484375, 0.8610916137695312, 0.21463775634765625, -0.23352432250976562, -1.2030181884765625, 0.20441818237304688, -0.24211502075195312, 0.20798492431640625, 0.15714263916015625, 0.6671142578125, -0.16717147827148438, 0.7463836669921875, -0.11577224731445312, 0.547454833984375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000056.npy"}
{"epoch": 0.08465608465608465, "step": 57, "batch_size": 64, "mean": 0.25117942690849304, "std": 0.5275972485542297, "min": -1.0581817626953125, "p10": -0.3738838195800781, "median": 0.19232749938964844, "p90": 0.7260597229003907, "max": 1.8155059814453125, "pos_frac": 0.71875, "sample": [0.2869720458984375, -0.20426559448242188, 0.06772613525390625, 1.8155059814453125, -0.14014434814453125, 0.48431396484375, 0.19797134399414062, 0.6688232421875, 0.6121482849121094, 0.299591064453125, 0.6866607666015625, 0.4696769714355469, 0.28985595703125, 0.24292755126953125, 0.7323379516601562, 0.052730560302734375, -0.9161376953125, 0.6726264953613281, 1.3180999755859375, 0.24584388732910156, -0.48734283447265625, 0.1584625244140625, 0.34307098388671875, 0.180084228515625, 0.11371231079101562, -0.042247772216796875, 0.1844024658203125, 1.0796737670898438, -0.06999969482421875, 0.33681488037109375, 0.69403076171875, 0.4911689758300781, 0.24065780639648438, 0.141876220703125, -0.6023597717285156, 1.4423751831054688, 0.47119140625, 0.66302490234375, -0.742340087890625, 0.384552001953125, -0.024707794189453125, 0.16048431396484375, -1.0581817626953125, 0.5901336669921875, 0.7114105224609375, -0.035125732421875, -0.33008575439453125, 0.18668365478515625, 0.00933074951171875, -0.08002090454101562, -0.11217117309570312, 0.13423919677734375, 0.6981048583984375, 1.04901123046875, 0.2505645751953125, 0.5859375, 1.308074951171875, -0.3926544189453125, 0.18318557739257812, -0.141021728515625, 0.037830352783203125, -0.10968780517578125, 0.123748779296875, -0.5336723327636719], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000057.npy"}
{"epoch": 0.08616780045351474, "step": 58, "batch_size": 64, "mean": 0.059114038944244385, "std": 0.4708560109138489, "min": -1.1473464965820312, "p10": -0.5062278747558594, "median": 0.09383773803710938, "p90": 0.6575256347656251, "max": 1.6656646728515625, "pos_frac": 0.578125, "sample": [0.4832763671875, 0.17113113403320312, -0.3633747100830078, -0.10894012451171875, 0.011871337890625, 0.10375213623046875, 0.6939163208007812, 1.6656646728515625, -0.4296875, 0.3204193115234375, 0.6869277954101562, -0.19312667846679688, 0.144989013671875, 0.08392333984375, -0.538909912109375, 0.4815635681152344, 0.677581787109375, -0.5391464233398438, -0.27685546875, 0.2260284423828125, -0.01099395751953125, 0.13115310668945312, -0.035404205322265625, 0.6647491455078125, -1.1473464965820312, -0.267578125, -0.5555191040039062, 0.6406707763671875, -0.35141563415527344, -0.17786788940429688, -0.2883148193359375, 0.33554840087890625, -0.01146697998046875, 0.8586273193359375, -0.80517578125, -0.11679840087890625, 0.7736663818359375, 0.507781982421875, 0.5455474853515625, -0.32343292236328125, 0.245513916015625, 0.06590080261230469, 0.4696502685546875, -0.4899444580078125, -0.40476226806640625, 0.2226409912109375, -0.0463104248046875, 0.110137939453125, 0.2994041442871094, 0.15269851684570312, 0.3645172119140625, 0.20377349853515625, 0.194091796875, -0.1309051513671875, 0.13909149169921875, 0.247467041015625, -0.14789962768554688, -0.4337921142578125, -0.5132064819335938, 0.364227294921875, -1.0625152587890625, 0.1412353515625, 0.064971923828125, 0.05987739562988281], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000058.npy"}
{"epoch": 0.08767951625094482, "step": 59, "batch_size": 64, "mean": 0.14015719294548035, "std": 0.47749724984169006, "min": -0.968170166015625, "p10": -0.48761367797851557, "median": 0.14632415771484375, "p90": 0.7902667999267579, "max": 1.295806884765625, "pos_frac": 0.625, "sample": [-0.7492752075195312, -0.968170166015625, 0.2664794921875, -0.07142257690429688, 0.35408782958984375, 0.02719879150390625, 0.24254608154296875, -0.5634307861328125, 0.050662994384765625, 0.80029296875, 1.295806884765625, -0.11942291259765625, 0.6180572509765625, -0.10343170166015625, 1.160573959350586, -0.045963287353515625, -0.4448699951171875, -0.09447479248046875, 0.681427001953125, -0.038299560546875, 0.23350906372070312, -0.0807647705078125, -0.4454498291015625, 0.28951263427734375, -0.33953094482421875, 0.7848167419433594, -0.10826873779296875, 0.351593017578125, 0.7926025390625, 0.31302642822265625, -0.3090858459472656, 0.8510665893554688, -0.02594757080078125, 0.17881011962890625, -0.8074188232421875, 0.064300537109375, -0.5056838989257812, -0.100982666015625, -0.10827255249023438, 0.17003631591796875, 1.1374359130859375, 0.71551513671875, 0.6143722534179688, 0.0408477783203125, 0.0543365478515625, 0.16613006591796875, 0.27289581298828125, 0.38530540466308594, -0.28586387634277344, 1.067291259765625, -0.5566940307617188, 0.24409103393554688, 0.2018890380859375, 0.22272491455078125, 0.5389175415039062, -0.5275650024414062, -0.3281707763671875, 0.06243896484375, 0.22517013549804688, 0.12651824951171875, 0.039615631103515625, 0.433258056640625, 0.41310882568359375, 0.2102508544921875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000059.npy"}
{"epoch": 0.08919123204837491, "step": 60, "batch_size": 64, "mean": 0.030887097120285034, "std": 0.4905214011669159, "min": -1.33453369140625, "p10": -0.5674865722656249, "median": 0.0052814483642578125, "p90": 0.6230567932128908, "max": 1.146026611328125, "pos_frac": 0.515625, "sample": [1.1079559326171875, 0.0422210693359375, 0.5324935913085938, 0.38344573974609375, -0.5733871459960938, -0.016246795654296875, -1.33453369140625, -0.5032882690429688, -0.7821197509765625, -0.609100341796875, 0.3055572509765625, 0.603729248046875, -0.27234458923339844, -0.2646675109863281, -0.5537185668945312, 0.46463775634765625, 0.15634536743164062, 0.09282302856445312, 0.0022754669189453125, 0.2314300537109375, -0.6683197021484375, 0.21574020385742188, 0.39434814453125, -0.17999267578125, -0.24843215942382812, 0.02579498291015625, -0.13434982299804688, 0.008287429809570312, 0.2071533203125, -0.5380096435546875, -0.3520469665527344, 0.0826568603515625, 0.35364532470703125, 0.05962371826171875, -0.09503173828125, -0.176361083984375, -0.05765533447265625, -0.00911712646484375, 0.530303955078125, -0.3718223571777344, -0.8017654418945312, 0.027812957763671875, -0.12841415405273438, -0.1832733154296875, -0.17496490478515625, -0.238372802734375, 0.16571807861328125, 1.146026611328125, 1.1224441528320312, 0.35005950927734375, -0.2043609619140625, 0.0778961181640625, 0.36340904235839844, -0.9305343627929688, 0.9755859375, -0.10946846008300781, -0.1737518310546875, -0.08849716186523438, 0.8545303344726562, 0.369049072265625, 0.73077392578125, -0.11450386047363281, 0.6313400268554688, 0.25011253356933594], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000060.npy"}
{"epoch": 0.09070294784580499, "step": 61, "batch_size": 64, "mean": 0.03356604278087616, "std": 0.47732895612716675, "min": -1.1315727233886719, "p10": -0.4237174987792969, "median": 0.012554168701171875, "p90": 0.5259132385253907, "max": 1.3675994873046875, "pos_frac": 0.53125, "sample": [-0.2466278076171875, 0.979217529296875, 0.5810546875, 0.8883514404296875, 0.3821868896484375, 0.047061920166015625, 0.19709014892578125, 0.47579193115234375, -0.43379974365234375, 0.11417770385742188, -0.39751434326171875, -0.16765975952148438, -0.1147613525390625, 0.3297576904296875, 0.00444793701171875, 0.1528034210205078, 0.002712249755859375, -0.03708839416503906, -0.7680130004882812, 0.1710662841796875, -0.3929901123046875, 0.48889923095703125, 0.361785888671875, -0.22224903106689453, -0.0056819915771484375, 0.08193397521972656, 0.1401214599609375, -0.021392822265625, -0.4001922607421875, -0.2569236755371094, -0.063446044921875, 0.22086143493652344, -0.3170051574707031, -0.1853485107421875, 0.534759521484375, 0.020660400390625, -0.2866325378417969, -0.1974964141845703, -0.3950653076171875, -0.05622100830078125, 0.4491119384765625, -0.6799697875976562, 1.1346282958984375, -0.9168930053710938, -0.0656585693359375, 0.6940078735351562, -0.3402519226074219, 0.1352081298828125, -0.25804710388183594, 0.464263916015625, 0.2201080322265625, -1.0253753662109375, 0.3307228088378906, 0.5052719116210938, 0.18860626220703125, -0.7858657836914062, 1.3675994873046875, 0.13001060485839844, -0.2315216064453125, 0.16997528076171875, 0.35984039306640625, -0.154449462890625, -1.1315727233886719, 0.37984466552734375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000061.npy"}
{"epoch": 0.09221466364323508, "step": 62, "batch_size": 64, "mean": 0.13860544562339783, "std": 0.548408567905426, "min": -1.249664306640625, "p10": -0.4355522155761718, "median": 0.12518310546875, "p90": 0.8367095947265631, "max": 1.460693359375, "pos_frac": 0.640625, "sample": [0.4039764404296875, -0.7880401611328125, 0.20457077026367188, 0.21841812133789062, -0.3446044921875, 0.4454841613769531, -0.472564697265625, 0.3178596496582031, -0.0107574462890625, 0.6670074462890625, 0.32660675048828125, 0.33817291259765625, 0.0904541015625, -1.2076263427734375, 0.6705856323242188, 1.2227706909179688, 0.12713623046875, 0.05194091796875, 0.4454689025878906, 0.2841644287109375, -0.6370773315429688, -0.3440399169921875, -1.249664306640625, 0.2910270690917969, -0.34918975830078125, -0.18973541259765625, -0.165557861328125, 0.12322998046875, 0.41298675537109375, 0.9079055786132812, -0.1210479736328125, -0.23769187927246094, 0.5851058959960938, -0.2659454345703125, 0.4453163146972656, -0.104705810546875, 0.40045928955078125, 1.1436767578125, -0.19370269775390625, -0.9174652099609375, 0.11255645751953125, 0.15680694580078125, -0.32207489013671875, 0.15615081787109375, 0.5769577026367188, 0.016510009765625, 1.460693359375, 0.09534454345703125, 0.12734222412109375, 0.9112701416015625, -0.5884246826171875, 0.2971076965332031, 0.4968070983886719, 0.10553932189941406, -0.2179412841796875, 0.00487518310546875, 1.1993865966796875, 1.4540252685546875, 0.5383319854736328, -0.2100067138671875, 0.14675140380859375, -0.0982818603515625, -0.1621246337890625, 0.08823776245117188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000062.npy"}
{"epoch": 0.09372637944066516, "step": 63, "batch_size": 64, "mean": 0.19448035955429077, "std": 0.5842654705047607, "min": -1.7519073486328125, "p10": -0.40021820068359376, "median": 0.13508224487304688, "p90": 0.7480056762695315, "max": 2.1090240478515625, "pos_frac": 0.671875, "sample": [0.442047119140625, -0.01699066162109375, 0.026885986328125, -0.06508636474609375, 2.0628662109375, -0.472869873046875, 0.17346763610839844, -0.9202880859375, 0.6755142211914062, 0.12923049926757812, -1.7519073486328125, -0.45545196533203125, 0.3236045837402344, -0.41292572021484375, 0.49196624755859375, 0.073516845703125, 2.1090240478515625, 0.116668701171875, -0.0265350341796875, -0.10442352294921875, 1.0801315307617188, 0.135162353515625, 0.5352897644042969, 0.863800048828125, 1.0675811767578125, 0.28115081787109375, -0.123626708984375, 0.6629829406738281, 0.4276084899902344, -0.920562744140625, 0.5967178344726562, 0.4388008117675781, 0.269256591796875, -0.33101654052734375, 0.1625823974609375, 0.69610595703125, -0.4015045166015625, -0.05780029296875, 0.3382987976074219, 0.13500213623046875, 0.2464122772216797, 0.58905029296875, -0.2550506591796875, 0.45704078674316406, 0.01806640625, 0.5394210815429688, -0.397216796875, 0.0140228271484375, 0.340484619140625, 0.01993560791015625, -0.06929397583007812, 0.34899139404296875, -0.37810516357421875, 0.6948623657226562, 0.02075958251953125, 0.1410064697265625, 0.4580078125, -0.05636024475097656, -0.0593109130859375, 0.07315826416015625, -0.20061492919921875, 0.7702484130859375, 0.7889251708984375, 0.0880279541015625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000063.npy"}
{"epoch": 0.09523809523809523, "step": 64, "batch_size": 64, "mean": 0.021149873733520508, "std": 0.5517740845680237, "min": -2.1276397705078125, "p10": -0.5809513092041015, "median": 0.076629638671875, "p90": 0.6762214660644534, "max": 1.59906005859375, "pos_frac": 0.59375, "sample": [0.5110702514648438, 1.59906005859375, 0.09675216674804688, 0.06314849853515625, -0.5912361145019531, 0.0679931640625, 0.705474853515625, 0.9329643249511719, 0.1735687255859375, 0.7093353271484375, 0.016042709350585938, 0.3341712951660156, 0.8892974853515625, 0.0996551513671875, 0.5428085327148438, 0.29143714904785156, 0.15418243408203125, -0.11318206787109375, -0.218597412109375, 0.12635040283203125, 0.1844329833984375, 0.26955413818359375, -0.04830169677734375, -0.04331207275390625, -0.61383056640625, -0.18828582763671875, 0.2437744140625, -0.579010009765625, 0.054073333740234375, 0.31585693359375, 0.1751842498779297, 0.4434661865234375, -0.03940582275390625, -0.8140792846679688, 0.6079635620117188, 0.827667236328125, 0.4901142120361328, -2.1276397705078125, -0.4360084533691406, 0.22313690185546875, 0.19020843505859375, 0.0244903564453125, -0.0914154052734375, -0.28227996826171875, -0.443939208984375, -0.5467529296875, -0.06008148193359375, 0.216094970703125, -1.0906143188476562, 0.1640625, -0.22759246826171875, 0.1303272247314453, -0.5817832946777344, -0.15630340576171875, 0.3538398742675781, 0.15903472900390625, 0.08526611328125, 0.010343551635742188, -0.3124504089355469, -0.4481544494628906, 0.752166748046875, -1.227508544921875, -0.42755126953125, -0.1714630126953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000064.npy"}
{"epoch": 0.09674981103552532, "step": 65, "batch_size": 64, "mean": 0.171383798122406, "std": 0.5902786254882812, "min": -1.0070953369140625, "p10": -0.6341247558593749, "median": 0.12885665893554688, "p90": 0.9988052368164063, "max": 1.4393539428710938, "pos_frac": 0.609375, "sample": [0.9635105133056641, 0.3098564147949219, 1.047821044921875, -0.27783203125, 0.2775917053222656, 0.397247314453125, -0.981719970703125, 0.0347900390625, -0.2909088134765625, 0.3400459289550781, -0.3307037353515625, 0.3336296081542969, 0.8560447692871094, -0.372283935546875, 0.18622589111328125, -0.46556854248046875, 0.6107959747314453, 0.7330245971679688, 0.23586273193359375, 0.94012451171875, 1.132659912109375, 0.42122650146484375, -1.0070953369140625, 0.5669097900390625, 1.4393539428710938, -0.14976119995117188, -0.9363632202148438, -0.04119873046875, 0.12525177001953125, -0.381378173828125, -0.03396034240722656, -0.04343414306640625, 0.24979400634765625, -0.85003662109375, 0.4317779541015625, 1.0100250244140625, 0.1324615478515625, 0.105194091796875, 1.2941741943359375, 0.3924293518066406, 0.972625732421875, -0.02333831787109375, 0.318756103515625, 0.08748626708984375, -0.8237762451171875, 0.10726165771484375, -0.029689788818359375, 0.5810317993164062, -0.3110466003417969, -0.691619873046875, 0.4507904052734375, -0.1354827880859375, -0.154754638671875, 0.10748672485351562, -0.0076904296875, -0.05142974853515625, 0.1831817626953125, 0.13396453857421875, 0.8362274169921875, -0.499969482421875, 1.3420944213867188, -0.917022705078125, 1.018585205078125, 0.06930732727050781], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000065.npy"}
{"epoch": 0.0982615268329554, "step": 66, "batch_size": 64, "mean": 0.18604576587677002, "std": 0.634008526802063, "min": -1.523284912109375, "p10": -0.4681629180908203, "median": 0.2288837432861328, "p90": 0.8834808349609377, "max": 2.8885345458984375, "pos_frac": 0.65625, "sample": [0.30287933349609375, 2.8885345458984375, -0.056171417236328125, 0.22529220581054688, 0.07850074768066406, 0.21895599365234375, -0.01262664794921875, 0.91717529296875, -0.47710418701171875, 0.27959442138671875, -0.16925048828125, -0.9331283569335938, -0.423248291015625, -0.2097320556640625, 0.2897186279296875, 0.42230224609375, 0.41223907470703125, 0.6111106872558594, 0.8942337036132812, 0.9940643310546875, 0.45198822021484375, 0.4400482177734375, -0.05304718017578125, 0.14430999755859375, -0.8878860473632812, 0.6288490295410156, 0.2907142639160156, 0.930999755859375, 0.446929931640625, -0.14002227783203125, 0.6543312072753906, 0.1677093505859375, 0.01387786865234375, 0.3246917724609375, 0.23247528076171875, -0.11722946166992188, -0.30460357666015625, 0.15972137451171875, -0.24576568603515625, 0.8340988159179688, -1.523284912109375, -0.9822463989257812, 0.42966461181640625, 0.3669891357421875, -0.26548004150390625, -0.2835540771484375, 0.6130828857421875, 0.2657470703125, 0.14703369140625, 0.11344528198242188, -0.020120620727539062, 0.06497955322265625, 0.39649200439453125, 1.053436279296875, 0.8583908081054688, -0.4472999572753906, 0.3102264404296875, 0.662628173828125, -1.0118255615234375, 1.2751312255859375, 0.23677825927734375, -0.3535728454589844, 0.34461212158203125, -0.569854736328125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000066.npy"}
{"epoch": 0.09977324263038549, "step": 67, "batch_size": 64, "mean": 0.1952143907546997, "std": 0.5142912268638611, "min": -0.7031021118164062, "p10": -0.439449119567871, "median": 0.160614013671875, "p90": 0.8077705383300782, "max": 1.7320404052734375, "pos_frac": 0.625, "sample": [-0.384735107421875, 0.2569427490234375, -0.12514495849609375, -0.07488059997558594, -0.130126953125, 0.32402801513671875, 0.9737014770507812, 0.5081386566162109, 0.50823974609375, -0.528167724609375, 1.566131591796875, 0.0032558441162109375, 0.13614654541015625, -0.455657958984375, 0.36464691162109375, 0.21206092834472656, -0.24724578857421875, 0.31954193115234375, 0.035129547119140625, -0.04788970947265625, -0.4877166748046875, 0.571075439453125, -0.5537567138671875, 1.471343994140625, -0.3045940399169922, 0.322418212890625, 0.5134391784667969, 0.7338104248046875, 0.47447967529296875, 0.043766021728515625, -0.6276512145996094, -0.4016284942626953, 0.13509368896484375, 0.7951240539550781, 0.8443679809570312, 0.2085590362548828, 0.4755516052246094, 0.42467498779296875, -0.052219390869140625, 0.40938568115234375, 0.6676177978515625, 0.15058135986328125, 0.17064666748046875, 1.7320404052734375, 0.504608154296875, 0.1279296875, 0.23870849609375, -0.5905342102050781, 0.4099273681640625, -0.0881195068359375, 0.1717529296875, -0.20916748046875, -0.17328834533691406, -0.03281402587890625, 0.401458740234375, -0.2772216796875, 0.12702178955078125, 0.4244880676269531, -0.37900543212890625, -0.7031021118164062, 0.8131904602050781, -0.260009765625, 1.088714599609375, -0.031341552734375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000067.npy"}
{"epoch": 0.10128495842781557, "step": 68, "batch_size": 64, "mean": 0.05804276466369629, "std": 0.6317670941352844, "min": -1.9453125, "p10": -0.6211982727050781, "median": 0.10622596740722656, "p90": 0.8635540008544925, "max": 1.4172286987304688, "pos_frac": 0.5625, "sample": [0.7545013427734375, 0.997100830078125, -0.020999908447265625, 0.156524658203125, -0.5704498291015625, 1.4172286987304688, 0.01369476318359375, 0.25991058349609375, 1.3607158660888672, -1.0923843383789062, 0.30348968505859375, -1.9453125, 0.073577880859375, -0.0137786865234375, 0.333740234375, 0.5090599060058594, 0.20168304443359375, -0.5011138916015625, 0.448486328125, 0.1127777099609375, 1.1365432739257812, -0.34709930419921875, 0.14459228515625, 0.478790283203125, -0.2932777404785156, 0.2659111022949219, -0.9613113403320312, -0.4512748718261719, 0.9623031616210938, -0.7756195068359375, 0.7826576232910156, 0.29204368591308594, 0.2309417724609375, -0.1031341552734375, 0.12319564819335938, -0.0098114013671875, 0.898223876953125, 0.1689910888671875, -0.54254150390625, -0.2093963623046875, -1.9432373046875, -0.5890426635742188, -0.319244384765625, 1.0272331237792969, -0.16112518310546875, 0.37081146240234375, -0.07201385498046875, 0.10803604125976562, -0.1826019287109375, -0.00702667236328125, 0.1044158935546875, 0.08530807495117188, 0.33095550537109375, 0.492584228515625, 0.27138519287109375, -0.018218994140625, -0.634979248046875, -0.6595382690429688, 0.6941986083984375, 0.21840286254882812, -0.077362060546875, -0.1753387451171875, 0.41782379150390625, -0.1558685302734375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000068.npy"}
{"epoch": 0.10279667422524566, "step": 69, "batch_size": 64, "mean": 0.18140414357185364, "std": 0.6294430494308472, "min": -1.28973388671875, "p10": -0.6775413513183594, "median": 0.2251911163330078, "p90": 0.8693405151367188, "max": 1.65753173828125, "pos_frac": 0.640625, "sample": [0.4378623962402344, -0.097137451171875, 0.6600341796875, -0.4304351806640625, 0.38199615478515625, 0.3949165344238281, 0.2293243408203125, -0.057109832763671875, 0.635833740234375, 1.3970870971679688, 0.09768486022949219, 0.0464935302734375, -0.2738075256347656, 0.74835205078125, -0.12262725830078125, 0.27591705322265625, 0.22105789184570312, 1.35845947265625, 0.8895263671875, -0.8095245361328125, -0.036285400390625, -0.987945556640625, 0.1334991455078125, 0.6738395690917969, 0.624267578125, -0.135345458984375, 0.5614852905273438, -0.2663726806640625, 0.8547821044921875, -0.2323455810546875, 0.6540260314941406, -1.28973388671875, 0.11011505126953125, 0.2583465576171875, 0.08782958984375, -1.1804351806640625, 1.3316726684570312, 0.5752944946289062, -0.61932373046875, 0.3806915283203125, 0.51220703125, 0.004215240478515625, -0.0031280517578125, -0.2971649169921875, 0.8019447326660156, -0.7024917602539062, 0.8818206787109375, 0.5926265716552734, 0.019346237182617188, -0.36286163330078125, -0.3094673156738281, 0.44287109375, -1.0651092529296875, -0.8493690490722656, 0.875579833984375, 0.4646759033203125, 0.6137237548828125, 0.8318748474121094, 0.5243072509765625, -0.4720268249511719, 1.65753173828125, 0.43816375732421875, -0.48126220703125, 0.009891510009765625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000069.npy"}
{"epoch": 0.10430839002267574, "step": 70, "batch_size": 64, "mean": 0.2574295997619629, "std": 0.6950369477272034, "min": -1.156585693359375, "p10": -0.5896541595458984, "median": 0.24650192260742188, "p90": 1.1451683044433594, "max": 1.875396728515625, "pos_frac": 0.640625, "sample": [0.272979736328125, 0.08517074584960938, 1.875396728515625, -0.5496826171875, 0.13895797729492188, -0.21767616271972656, -0.3376312255859375, 0.8742446899414062, 0.8776473999023438, -0.700592041015625, 1.508544921875, -0.2641258239746094, -0.841552734375, 1.1472854614257812, 0.005496978759765625, 0.05549430847167969, 0.5129928588867188, -0.3771209716796875, -0.43527984619140625, 0.016748428344726562, 0.8543701171875, 0.7263717651367188, 0.8432884216308594, 0.2414398193359375, -0.8080291748046875, 0.6389617919921875, -0.6067848205566406, 1.2798080444335938, -0.23482513427734375, 1.140228271484375, 0.108123779296875, -0.07125473022460938, -0.44618988037109375, -0.26042938232421875, -0.2238006591796875, 1.2987823486328125, 0.15855789184570312, -0.240936279296875, 0.3080425262451172, 0.45452117919921875, 0.3454399108886719, 0.2695770263671875, -0.2622337341308594, 1.0432281494140625, 0.9105682373046875, 1.7079849243164062, 0.4104766845703125, -1.156585693359375, -0.3551788330078125, 0.563934326171875, 0.9666824340820312, 1.7677841186523438, 0.6306877136230469, 0.7158203125, -0.38132476806640625, 0.2840576171875, 0.25156402587890625, -0.6077995300292969, 0.27503204345703125, 1.098184585571289, -1.0837249755859375, -0.06627655029296875, 0.03692054748535156, 0.303131103515625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000070.npy"}
{"epoch": 0.10582010582010581, "step": 71, "batch_size": 64, "mean": 0.18517541885375977, "std": 0.6167566776275635, "min": -1.3022918701171875, "p10": -0.7478919982910156, "median": 0.1938800811767578, "p90": 0.9401302337646484, "max": 1.3309173583984375, "pos_frac": 0.65625, "sample": [0.528167724609375, 0.200836181640625, 0.4828834533691406, 0.4936408996582031, 0.9398231506347656, 0.378570556640625, 0.5143833160400391, -0.56060791015625, 0.6285629272460938, 0.078826904296875, 0.07197952270507812, -0.7682418823242188, -0.0576629638671875, -0.10460662841796875, -0.9009170532226562, 0.13686370849609375, -0.6148605346679688, 0.3128662109375, 0.833160400390625, -0.3069610595703125, 0.3668365478515625, 0.8631362915039062, -0.22133255004882812, 0.12457275390625, 0.43173980712890625, 1.2097015380859375, 0.8837661743164062, -0.1812591552734375, 0.17198562622070312, 0.0579681396484375, -0.9376220703125, -1.077239990234375, -0.700408935546875, -0.7706069946289062, -0.095611572265625, 0.496795654296875, 0.39833831787109375, -0.2610931396484375, 0.5718421936035156, -0.5345916748046875, -0.0164642333984375, -1.3022918701171875, 0.7733154296875, 0.9402618408203125, -0.4597949981689453, 0.0028247833251953125, -0.01953125, 1.1319122314453125, 0.09795379638671875, 0.19392776489257812, 0.3074989318847656, 1.0958786010742188, -0.7726974487304688, 1.3309173583984375, 0.8573989868164062, 0.5087203979492188, 0.9567031860351562, 0.1938323974609375, 0.27410125732421875, 0.6823196411132812, 0.8994522094726562, 0.13021278381347656, 1.25970458984375, -0.298553466796875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000071.npy"}
{"epoch": 0.1073318216175359, "step": 72, "batch_size": 64, "mean": 0.3179430365562439, "std": 0.7393075823783875, "min": -1.8384857177734375, "p10": -0.3582992553710937, "median": 0.2949867248535156, "p90": 1.3966485977172858, "max": 2.5281295776367188, "pos_frac": 0.6875, "sample": [-0.31418609619140625, 1.8464202880859375, 0.29715728759765625, 0.7585220336914062, -0.6282501220703125, 0.04663848876953125, 0.11230850219726562, 1.7399330139160156, -0.3034400939941406, 0.4670524597167969, -0.15340042114257812, 0.5075302124023438, -0.2847442626953125, -0.0254058837890625, 0.27956390380859375, 0.3742256164550781, 0.3092823028564453, 0.043987274169921875, 0.24288177490234375, 0.414276123046875, 0.7307281494140625, 0.3246269226074219, 0.8053150177001953, 0.18791961669921875, 0.9110870361328125, 0.42230987548828125, 0.24435806274414062, -0.04113006591796875, -0.37720489501953125, 0.43289947509765625, 1.675872802734375, 0.6129589080810547, -0.173553466796875, -0.22698211669921875, 1.6259765625, -0.2699546813964844, 0.6182651519775391, 0.22357177734375, -0.2284698486328125, 0.5648689270019531, 0.6338920593261719, 1.2474422454833984, 1.0972213745117188, -0.14367103576660156, 0.1028900146484375, 0.591033935546875, 0.243804931640625, 0.0083770751953125, 0.35706329345703125, 0.42469024658203125, -1.6439056396484375, -0.42193603515625, 2.5281295776367188, -0.055522918701171875, 1.598602294921875, -0.297088623046875, 0.560882568359375, 0.30345916748046875, -1.8384857177734375, -0.6773262023925781, 1.4605941772460938, -0.5433082580566406, 0.292816162109375, 0.724884033203125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000072.npy"}
{"epoch": 0.10884353741496598, "step": 73, "batch_size": 64, "mean": 0.031659454107284546, "std": 0.7540120482444763, "min": -2.354217529296875, "p10": -0.8535598754882812, "median": -0.018251419067382812, "p90": 0.945510864257813, "max": 1.984527587890625, "pos_frac": 0.484375, "sample": [0.49205970764160156, -0.1172637939453125, 0.8173904418945312, -0.80584716796875, -0.2628440856933594, -1.1246261596679688, -0.9389266967773438, -0.02001953125, -0.3666534423828125, 1.0004196166992188, -0.7576522827148438, 0.49585723876953125, 0.8120536804199219, -0.42993927001953125, -0.19971466064453125, 0.6382980346679688, -0.790252685546875, 0.16483306884765625, -0.9068984985351562, 0.7889175415039062, 1.241973876953125, 0.05492401123046875, -0.016483306884765625, -0.39052581787109375, 1.0648193359375, 1.4458389282226562, -0.36684417724609375, -0.5645561218261719, -0.06309700012207031, 0.7378997802734375, 0.06047821044921875, -0.3161354064941406, 0.21981430053710938, -0.21642303466796875, 0.33184814453125, 0.3211517333984375, 0.3231925964355469, 0.2660369873046875, -0.896087646484375, -0.40808868408203125, -1.321258544921875, 1.984527587890625, -2.354217529296875, -0.13016510009765625, -0.46009063720703125, 0.585845947265625, 0.26587677001953125, -0.17426300048828125, 0.4568061828613281, 0.009674072265625, -0.442474365234375, 0.6452140808105469, 0.024580001831054688, -0.5141716003417969, 0.388671875, 0.5846023559570312, -0.4595489501953125, 0.36749267578125, -0.20996856689453125, -0.8740081787109375, -0.14849472045898438, 1.7130889892578125, 1.37261962890625, -0.6030616760253906], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000073.npy"}
{"epoch": 0.11035525321239607, "step": 74, "batch_size": 64, "mean": 0.31990575790405273, "std": 0.673071026802063, "min": -1.3017120361328125, "p10": -0.5679611206054687, "median": 0.3635997772216797, "p90": 1.041006088256836, "max": 2.126983642578125, "pos_frac": 0.703125, "sample": [-0.37436676025390625, 0.678314208984375, -1.3017120361328125, 0.4595794677734375, -0.12477874755859375, 0.7550888061523438, 0.2976226806640625, 1.59844970703125, 0.5768585205078125, 0.7783279418945312, 0.3647041320800781, 0.42458343505859375, 0.9129638671875, 0.6963729858398438, -0.2046356201171875, 1.0422706604003906, 0.737213134765625, 1.5404052734375, -0.12500953674316406, -0.9679107666015625, 1.038055419921875, 0.44483184814453125, 0.36249542236328125, 0.2816886901855469, 0.7805862426757812, -0.07086181640625, 0.015903472900390625, 0.7448310852050781, 1.3559417724609375, -0.7578964233398438, -0.24295425415039062, 0.6475257873535156, 0.6426277160644531, -0.5023040771484375, -0.4398841857910156, 0.7147369384765625, -0.08956146240234375, -0.6922454833984375, -0.7900772094726562, 0.10913467407226562, 0.5139808654785156, -0.33341217041015625, 0.7054519653320312, 0.8158035278320312, 0.3923492431640625, 2.126983642578125, 0.8198699951171875, 0.23161697387695312, 1.6327362060546875, -0.35230255126953125, -1.0788116455078125, 0.1563549041748047, -0.596099853515625, 0.060413360595703125, 0.522003173828125, -0.12345123291015625, 0.842620849609375, 0.152679443359375, 0.3212890625, 0.5966110229492188, 0.32183074951171875, 1.1230621337890625, 0.03105926513671875, 0.2744140625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000074.npy"}
{"epoch": 0.11186696900982615, "step": 75, "batch_size": 64, "mean": 0.26736900210380554, "std": 0.7214210629463196, "min": -2.1520538330078125, "p10": -0.297911262512207, "median": 0.20368194580078125, "p90": 1.0859832763671877, "max": 2.16912841796875, "pos_frac": 0.65625, "sample": [-0.2590827941894531, 0.8978424072265625, 0.4223899841308594, 0.7687301635742188, 1.5403861999511719, 0.56396484375, -0.20169448852539062, 0.22552871704101562, 0.2583465576171875, 1.1873207092285156, 0.205535888671875, -0.2941474914550781, -0.31081199645996094, 0.2018280029296875, 0.0923309326171875, 1.034393310546875, 0.675994873046875, -0.29952430725097656, -0.09210014343261719, -0.23192214965820312, 2.16912841796875, 0.8797950744628906, 0.695281982421875, -0.12181854248046875, -0.189697265625, -0.26385498046875, 0.5834808349609375, -1.5228729248046875, 0.9133872985839844, 0.06753921508789062, 0.9053726196289062, 1.3090705871582031, -0.28601837158203125, 0.5949726104736328, -1.20379638671875, -0.11136817932128906, 1.6546630859375, 0.585784912109375, 0.45795440673828125, 0.49304962158203125, -0.5261993408203125, 0.6768779754638672, -2.1520538330078125, -0.17738723754882812, 0.2656593322753906, -0.11495208740234375, -0.23719024658203125, 0.09349822998046875, -0.2273387908935547, 0.14881515502929688, 0.3387889862060547, 0.04085540771484375, 0.0925140380859375, 0.1904296875, 0.5524940490722656, 0.46063995361328125, 1.10809326171875, 0.2216167449951172, -0.44637298583984375, 0.1127471923828125, 0.0058422088623046875, -0.04047203063964844, 2.06048583984375, 0.6688613891601562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000075.npy"}
{"epoch": 0.11337868480725624, "step": 76, "batch_size": 64, "mean": 0.22301700711250305, "std": 0.7982138991355896, "min": -2.0301666259765625, "p10": -0.6745975494384765, "median": 0.16076278686523438, "p90": 1.2452972412109378, "max": 1.9461212158203125, "pos_frac": 0.609375, "sample": [0.00197601318359375, 0.1054229736328125, -0.3663787841796875, 0.2938385009765625, 0.62493896484375, -0.6100044250488281, -0.2953376770019531, 0.08933258056640625, 0.6203098297119141, 0.647796630859375, 1.1476287841796875, 0.715606689453125, -0.14291000366210938, -0.7848052978515625, 1.1684417724609375, -1.5475540161132812, -0.619842529296875, 0.6808223724365234, 0.2689208984375, 0.03794097900390625, 0.6603546142578125, 1.45928955078125, -0.9929275512695312, 1.5925216674804688, 1.201995849609375, 0.2923107147216797, -0.029352188110351562, 0.9005203247070312, -0.8228530883789062, 0.6535263061523438, -0.42072296142578125, -0.432159423828125, -0.6896476745605469, -0.5658493041992188, -0.3401165008544922, -0.28098297119140625, 1.73419189453125, 1.9461212158203125, -0.21422195434570312, -0.1459503173828125, -0.0817413330078125, 0.6481475830078125, -0.6394805908203125, 0.105499267578125, 0.26900482177734375, -0.10622215270996094, 0.21602630615234375, -2.0301666259765625, 1.0874366760253906, 0.6056289672851562, 0.5086288452148438, 1.26385498046875, 0.24684524536132812, 1.7979965209960938, 1.13836669921875, 0.08909225463867188, 0.2359161376953125, 0.9309234619140625, -0.8203811645507812, 0.4600067138671875, 1.38934326171875, 0.06524276733398438, -0.08414459228515625, -0.5649280548095703], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000076.npy"}
{"epoch": 0.11489040060468632, "step": 77, "batch_size": 64, "mean": 0.3992432653903961, "std": 0.7737142443656921, "min": -1.2285690307617188, "p10": -0.46333351135253903, "median": 0.3045177459716797, "p90": 1.6981819152832032, "max": 2.3232078552246094, "pos_frac": 0.703125, "sample": [0.029041290283203125, 0.29665374755859375, 0.5684623718261719, -0.98974609375, 0.9727802276611328, 0.10141754150390625, -0.02272796630859375, 0.5337295532226562, -1.2285690307617188, -0.10711669921875, 0.413818359375, -0.20911407470703125, -0.1127777099609375, 0.4623298645019531, -0.6047210693359375, 2.3232078552246094, 1.8628692626953125, 0.8172454833984375, 0.27379608154296875, 1.9461746215820312, 0.02498626708984375, 0.3961639404296875, 1.225555419921875, 1.097808837890625, 0.3123817443847656, 0.24434280395507812, -0.11021041870117188, 1.5547637939453125, 0.08869552612304688, 0.4569854736328125, -0.3978691101074219, -0.11322784423828125, 0.4374580383300781, 2.034027099609375, 0.23131179809570312, 1.028656005859375, 1.806640625, -0.21625900268554688, 0.2215709686279297, 1.9602508544921875, 0.6200942993164062, -0.6712722778320312, 0.9921073913574219, -0.4810981750488281, -0.5844345092773438, 0.2167205810546875, -0.23181533813476562, -0.3211078643798828, 0.8940582275390625, 0.5268478393554688, 0.575164794921875, 0.35492706298828125, 0.19185256958007812, 0.6126174926757812, 1.6701812744140625, 0.33928680419921875, -0.42188262939453125, 0.384002685546875, 1.7101821899414062, 0.009960174560546875, -0.7963485717773438, 0.545501708984375, 0.013702392578125, -0.208465576171875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000077.npy"}
{"epoch": 0.1164021164021164, "step": 78, "batch_size": 64, "mean": 0.3344375789165497, "std": 0.7805081605911255, "min": -1.5887832641601562, "p10": -0.56845703125, "median": 0.30043792724609375, "p90": 1.3567993164062502, "max": 2.2531280517578125, "pos_frac": 0.671875, "sample": [0.6442947387695312, 0.944122314453125, 0.4980812072753906, 1.017059326171875, 0.3191680908203125, 0.844390869140625, 2.2531280517578125, 0.406585693359375, -0.38451385498046875, 0.641937255859375, 0.7052001953125, 0.5893096923828125, 0.15246009826660156, 1.2984848022460938, 0.6568450927734375, -0.27371978759765625, -0.60235595703125, 1.9110984802246094, -0.13727569580078125, -0.2357177734375, -1.074462890625, -0.92425537109375, 0.05126953125, -0.37783050537109375, -0.41943359375, 0.16400146484375, -0.062023162841796875, -0.5175628662109375, 1.379852294921875, -0.5487823486328125, 0.8127059936523438, 0.32421875, -0.14042282104492188, 1.5338821411132812, 1.938720703125, -0.5768890380859375, -1.5887832641601562, 0.1443328857421875, 0.9327201843261719, 0.8252429962158203, -0.36338043212890625, 0.18488311767578125, 1.303009033203125, 0.9540786743164062, 0.11058807373046875, 0.18817901611328125, -1.0688514709472656, -0.09082412719726562, 0.4917144775390625, 0.8474273681640625, 0.06034660339355469, 0.6958770751953125, 0.40421295166015625, -0.52239990234375, 1.9490203857421875, 0.281707763671875, -0.05840301513671875, -0.6593704223632812, 0.6713180541992188, 0.6300811767578125, 0.4977569580078125, 0.24446487426757812, 1.5177268981933594, 0.00975799560546875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000078.npy"}
{"epoch": 0.11791383219954649, "step": 79, "batch_size": 64, "mean": 0.46014824509620667, "std": 0.9755588173866272, "min": -1.76727294921875, "p10": -0.6425567626953124, "median": 0.3349781036376953, "p90": 1.5724868774414067, "max": 3.407684326171875, "pos_frac": 0.6875, "sample": [0.10518455505371094, -0.03379058837890625, -1.76727294921875, 0.5524559020996094, -0.7080535888671875, 0.8911056518554688, -0.7642936706542969, 0.060993194580078125, 0.21099853515625, -0.4755401611328125, -0.29706382751464844, -0.0475921630859375, 0.407684326171875, 3.407684326171875, -0.3357086181640625, -0.11487579345703125, 0.740203857421875, 0.6012153625488281, -0.4897308349609375, -0.071685791015625, 1.0322017669677734, 2.727325439453125, -0.2406005859375, 0.470733642578125, 1.1361427307128906, -0.23095703125, 2.82794189453125, 1.6562042236328125, 0.24773406982421875, 1.459320068359375, -1.2628097534179688, 0.3010406494140625, 0.8638687133789062, 1.1104736328125, 0.3243141174316406, 0.10708236694335938, 0.34564208984375, 0.5025253295898438, -0.1038055419921875, 1.38348388671875, -0.7321243286132812, 0.9630050659179688, 1.0789413452148438, 0.04128265380859375, 0.3461456298828125, 0.2038116455078125, 0.954925537109375, 0.22877883911132812, 0.5957431793212891, -0.1235198974609375, 0.7658176422119141, -1.1831207275390625, 1.6209869384765625, 0.9106292724609375, 1.2269134521484375, 3.083160400390625, -0.728302001953125, 0.4141082763671875, -0.39532470703125, 0.49468994140625, 0.0124359130859375, 1.2416534423828125, 1.830718994140625, 0.06835174560546875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000079.npy"}
{"epoch": 0.11942554799697656, "step": 80, "batch_size": 64, "mean": 0.2801358699798584, "std": 0.8778076171875, "min": -1.5652008056640625, "p10": -0.7703018188476562, "median": 0.2859992980957031, "p90": 1.3749488830566408, "max": 3.430023193359375, "pos_frac": 0.65625, "sample": [0.5261611938476562, 1.5386428833007812, -0.0965423583984375, 0.65045166015625, 1.4574527740478516, 0.05926513671875, -0.6771316528320312, -0.22620010375976562, -0.7346267700195312, -1.5077743530273438, 0.6723823547363281, 1.829071044921875, 0.8087997436523438, 0.0879669189453125, 0.5715446472167969, 1.6717681884765625, 0.33460235595703125, -0.595550537109375, -0.003635406494140625, 0.9004592895507812, 0.0107879638671875, 0.14727783203125, -0.4859352111816406, 1.3926620483398438, 0.9297027587890625, 0.7172164916992188, 0.310394287109375, 0.5924453735351562, 1.063323974609375, 0.4580078125, -0.7855911254882812, 0.28626251220703125, 0.6307411193847656, -0.29543113708496094, 0.25183868408203125, 1.30535888671875, 0.13552093505859375, 1.3336181640625, 0.5652389526367188, -0.522186279296875, 0.78411865234375, 0.6644210815429688, -1.5652008056640625, 1.0926475524902344, -0.5666313171386719, 0.40360260009765625, -0.9664306640625, 0.2197551727294922, -1.3987197875976562, -0.544158935546875, 3.430023193359375, 0.285736083984375, 0.2926788330078125, -0.5229415893554688, -0.0255279541015625, -0.9889678955078125, 0.48841094970703125, 0.5662727355957031, 0.19689369201660156, -0.92681884765625, 0.0754547119140625, -0.11407470703125, 1.779510498046875, -0.0397186279296875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000080.npy"}
{"epoch": 0.12093726379440665, "step": 81, "batch_size": 64, "mean": 0.3710387945175171, "std": 0.8975710272789001, "min": -1.3420486450195312, "p10": -0.7618804931640624, "median": 0.22093582153320312, "p90": 1.5608760833740238, "max": 2.812896728515625, "pos_frac": 0.640625, "sample": [0.32131004333496094, 0.9909286499023438, 0.4806976318359375, -0.6094284057617188, 0.8746414184570312, -0.4724884033203125, -0.0336456298828125, 0.00698089599609375, -0.2754840850830078, 0.00028228759765625, 1.8119659423828125, -1.187713623046875, -0.2693634033203125, 0.13482666015625, 0.262115478515625, 1.0424118041992188, 0.02275371551513672, -0.08008575439453125, 0.4911956787109375, 0.8233814239501953, 0.35765838623046875, -0.3500404357910156, 1.0574111938476562, -0.19219970703125, 1.07208251953125, 0.4664421081542969, -0.5546112060546875, 2.5038986206054688, -0.01349639892578125, -0.3211212158203125, 0.10894775390625, -1.3420486450195312, 1.798309326171875, 0.021343231201171875, 0.6244735717773438, 1.3041229248046875, 0.917083740234375, 1.868072509765625, -0.932159423828125, 2.812896728515625, -1.1724700927734375, 0.9638576507568359, 0.1607818603515625, 0.7786941528320312, -0.8272171020507812, 0.1530475616455078, 1.0330085754394531, 1.0183486938476562, 1.8914337158203125, 1.595458984375, 1.2503395080566406, -0.5590744018554688, 1.4801826477050781, -0.36675262451171875, 0.3537178039550781, -0.31589508056640625, -1.1325969696044922, 0.17975616455078125, -0.0918121337890625, 1.1222114562988281, -0.09935951232910156, 0.7339019775390625, 0.927154541015625, -0.8725814819335938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000081.npy"}
{"epoch": 0.12244897959183673, "step": 82, "batch_size": 64, "mean": 0.6147127151489258, "std": 1.055301547050476, "min": -1.562469482421875, "p10": -0.4882904052734375, "median": 0.4528160095214844, "p90": 1.6269317626953126, "max": 4.127899169921875, "pos_frac": 0.796875, "sample": [0.282318115234375, 0.3152732849121094, 0.17757034301757812, 0.012334823608398438, 1.3408279418945312, 0.27129364013671875, 1.3262271881103516, 1.2475090026855469, 1.70184326171875, 1.2922477722167969, 0.7406005859375, 0.5669898986816406, 3.5455780029296875, 1.0099563598632812, 0.569610595703125, 0.21649169921875, 1.293609619140625, -0.9557304382324219, 0.9253730773925781, 1.979736328125, 0.5692501068115234, 0.4695320129394531, 0.3532829284667969, -0.37693023681640625, -0.0171051025390625, 1.5303421020507812, 0.20725250244140625, 0.7585525512695312, 0.7774238586425781, 0.2937431335449219, -0.9144001007080078, 0.4361000061035156, 4.127899169921875, -0.49393463134765625, 1.6392364501953125, -1.562469482421875, 1.5982208251953125, 1.26287841796875, 0.77020263671875, 1.871124267578125, 0.02785491943359375, -0.12747955322265625, 0.4022979736328125, -0.12601470947265625, 0.3350677490234375, 0.4360389709472656, 0.3930244445800781, 0.07734298706054688, 0.7645339965820312, 0.41595458984375, -1.3801956176757812, 0.41650390625, 1.223236083984375, 3.9157333374023438, 0.8450927734375, -0.8390388488769531, 1.2653465270996094, 0.8517532348632812, 0.6923370361328125, 0.19907188415527344, -0.47342681884765625, -1.3043365478515625, -0.47512054443359375, 0.6461715698242188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000082.npy"}
{"epoch": 0.12396069538926682, "step": 83, "batch_size": 64, "mean": 0.5431385636329651, "std": 0.8882945775985718, "min": -1.1162109375, "p10": -0.47535057067871084, "median": 0.5095205307006836, "p90": 1.7294860839843753, "max": 3.3176116943359375, "pos_frac": 0.71875, "sample": [1.76422119140625, 0.55474853515625, -1.0441436767578125, 0.405853271484375, 0.6138267517089844, 1.052398681640625, 0.96337890625, 1.5506057739257812, 0.45223236083984375, 1.6484375, 1.0757293701171875, 0.7306327819824219, -0.5859527587890625, 0.578369140625, 0.4027843475341797, 0.4774303436279297, 0.9059524536132812, 2.0519790649414062, 0.03787994384765625, -0.21026992797851562, 0.7947006225585938, -0.5252113342285156, 2.3096237182617188, 0.6738243103027344, -0.3564300537109375, 0.044647216796875, 2.0718231201171875, 0.8766632080078125, -0.14278411865234375, -0.2966022491455078, -0.975677490234375, -0.2761077880859375, 0.11237716674804688, -1.1162109375, 0.08304595947265625, 1.0705528259277344, 0.4180908203125, 1.9860763549804688, -0.20259857177734375, -0.647613525390625, -0.32721710205078125, -0.29172515869140625, 1.5155181884765625, 0.4751472473144531, -0.08607292175292969, -0.3321361541748047, 0.3849830627441406, 0.778167724609375, 0.92645263671875, 1.0651397705078125, 3.3176116943359375, 0.38501739501953125, 1.1928596496582031, 1.2597885131835938, 0.6974868774414062, 0.9700679779052734, 0.5416107177734375, 0.8180580139160156, 2.1966094970703125, -0.3590087890625, 0.2457733154296875, 0.4169464111328125, 0.7508964538574219, -1.1093902587890625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000083.npy"}
{"epoch": 0.1254724111866969, "step": 84, "batch_size": 64, "mean": 0.5130621194839478, "std": 0.9534111618995667, "min": -1.977325439453125, "p10": -0.5199398040771485, "median": 0.5094203948974609, "p90": 1.6675018310546876, "max": 2.990020751953125, "pos_frac": 0.671875, "sample": [-0.186737060546875, 0.9629325866699219, -0.17119979858398438, 1.3554420471191406, 2.1631317138671875, 0.5313453674316406, 1.7693328857421875, 0.051639556884765625, 0.07296371459960938, -1.7720947265625, 1.0193328857421875, 0.8998794555664062, 1.5274276733398438, -0.159149169921875, 1.380157470703125, 0.9772415161132812, -0.09853363037109375, -1.0624313354492188, 1.5607452392578125, 0.11814498901367188, 0.0488739013671875, 1.0053024291992188, -0.1928863525390625, 1.1853141784667969, 1.1706981658935547, 0.737457275390625, -0.5232276916503906, 0.9888916015625, -0.000274658203125, 1.8436813354492188, 0.47480010986328125, -0.0710296630859375, 0.036251068115234375, 2.103240966796875, 0.12140655517578125, 0.9300765991210938, 0.674560546875, 1.182668685913086, 0.4627799987792969, -0.9928054809570312, -1.977325439453125, 1.8052520751953125, -0.14417266845703125, -0.3348884582519531, 0.48749542236328125, 0.00702667236328125, -0.049571990966796875, 0.952423095703125, 1.3083953857421875, -0.5803298950195312, -0.51226806640625, 1.1214752197265625, 0.9793853759765625, 1.2404098510742188, 0.46149444580078125, 1.6762542724609375, 1.6470794677734375, -0.22115516662597656, -0.45073699951171875, -1.1302947998046875, 1.1512336730957031, 0.69049072265625, -0.40706634521484375, 2.990020751953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000084.npy"}
{"epoch": 0.12698412698412698, "step": 85, "batch_size": 64, "mean": 0.5673837661743164, "std": 0.9147955179214478, "min": -1.4821929931640625, "p10": -0.8138179779052734, "median": 0.6224536895751953, "p90": 1.5750301361083985, "max": 2.3821868896484375, "pos_frac": 0.765625, "sample": [0.9241523742675781, -0.9438552856445312, 0.6029624938964844, 0.5640239715576172, -0.0911407470703125, 0.9577789306640625, 1.0091552734375, 0.6971912384033203, -0.047138214111328125, 0.4285430908203125, 1.4452590942382812, -1.09576416015625, 1.5141105651855469, -1.3001556396484375, 0.42697715759277344, 0.787109375, 1.2731361389160156, 1.47271728515625, 2.0231399536132812, 1.5426521301269531, -1.2817916870117188, 0.3160400390625, -0.8287811279296875, 0.9524688720703125, 0.940673828125, 2.1510848999023438, 0.007343292236328125, 1.3073310852050781, 1.89996337890625, 0.28436851501464844, -0.56134033203125, 0.9925498962402344, 1.0721893310546875, 0.465667724609375, 0.89361572265625, 0.2704429626464844, 2.1769790649414062, 1.1792182922363281, 0.3483009338378906, 0.13473129272460938, -0.7789039611816406, 0.902130126953125, 1.5480308532714844, 1.5866012573242188, 2.3821868896484375, 0.6419448852539062, 0.734771728515625, 0.13294029235839844, 0.34757232666015625, 1.5455245971679688, 0.24518966674804688, -0.4465484619140625, -0.06945037841796875, -0.25855255126953125, 0.3945960998535156, 0.5537796020507812, -0.7402572631835938, -1.4821929931640625, 0.7631797790527344, 0.8840713500976562, 1.9101181030273438, -1.0027923583984375, 1.2449798583984375, 0.3617286682128906], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000085.npy"}
{"epoch": 0.12849584278155707, "step": 86, "batch_size": 64, "mean": 0.3417322337627411, "std": 0.8305058479309082, "min": -1.58709716796875, "p10": -0.6412975311279296, "median": 0.28076744079589844, "p90": 1.5496006011962893, "max": 2.4152069091796875, "pos_frac": 0.671875, "sample": [1.894805908203125, 0.259979248046875, -0.398834228515625, 0.12335205078125, 1.4986915588378906, 0.38161659240722656, -0.5739059448242188, 0.4608917236328125, -1.58709716796875, -0.010936737060546875, 0.3015556335449219, -0.3912925720214844, 1.7815704345703125, 0.488616943359375, -0.8762359619140625, 0.227386474609375, 0.4705162048339844, -0.6817970275878906, -1.088897705078125, 0.4848175048828125, 2.344512939453125, 1.166015625, 0.49019622802734375, 1.1385383605957031, 1.617401123046875, 0.491302490234375, -0.328704833984375, 0.5087547302246094, 0.5767669677734375, 0.21743392944335938, 1.386444091796875, 1.711151123046875, -0.1091156005859375, -1.3130035400390625, 0.1345062255859375, 0.8607101440429688, 0.008220672607421875, -0.07056045532226562, 0.544586181640625, 0.2225189208984375, -0.43184661865234375, 1.0791206359863281, 0.7993240356445312, -0.6590805053710938, 0.0693206787109375, -0.03725433349609375, 1.4806022644042969, -0.0808258056640625, 0.39694976806640625, -0.5998039245605469, 2.4152069091796875, 0.871856689453125, 0.127838134765625, -0.12845230102539062, 0.6375732421875, 0.2289886474609375, -0.6813545227050781, 1.5714187622070312, 0.00384521484375, -0.4067649841308594, 0.31098175048828125, -0.4012870788574219, 0.5804290771484375, 0.3615989685058594], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000086.npy"}
{"epoch": 0.13000755857898716, "step": 87, "batch_size": 64, "mean": 0.7085880041122437, "std": 1.1770966053009033, "min": -1.862060546875, "p10": -0.8481918334960937, "median": 0.8929252624511719, "p90": 2.2230674743652346, "max": 3.147613525390625, "pos_frac": 0.6875, "sample": [-0.63812255859375, 2.173309326171875, 2.2380752563476562, -0.4128150939941406, 2.18804931640625, -0.777130126953125, 2.3990478515625, -0.7634811401367188, 2.589202880859375, 1.9529914855957031, 0.33853912353515625, -0.5485706329345703, 1.1468582153320312, -1.1349105834960938, -1.862060546875, 2.6995086669921875, -0.915802001953125, 1.6496810913085938, 1.0019454956054688, 0.6259117126464844, 0.9088134765625, 1.54364013671875, 0.6315460205078125, -0.41019439697265625, 1.391845703125, -1.4724884033203125, -0.2578620910644531, -0.29215049743652344, 0.8877944946289062, -0.8786468505859375, 0.31292724609375, -0.21866989135742188, 0.7357673645019531, -0.4125099182128906, 0.43944549560546875, -1.3392181396484375, 1.149749755859375, 0.9999427795410156, 1.0856132507324219, 0.0165557861328125, 0.5045547485351562, 1.7078781127929688, 2.114704132080078, -0.31714630126953125, 1.8587112426757812, -0.43988800048828125, -0.17138290405273438, 1.5073738098144531, -1.0434532165527344, 1.635345458984375, 2.3695755004882812, 0.8980560302734375, 3.147613525390625, 0.9463882446289062, 0.9594345092773438, 0.22971343994140625, 0.6823978424072266, 0.09401321411132812, 2.375335693359375, 1.584197998046875, 1.860748291015625, 0.9216423034667969, 1.688201904296875, 1.4634857177734375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000087.npy"}
{"epoch": 0.13151927437641722, "step": 88, "batch_size": 64, "mean": 0.3276621699333191, "std": 1.206264615058899, "min": -3.6255035400390625, "p10": -1.1637516021728513, "median": 0.3669872283935547, "p90": 1.7195686340332041, "max": 3.8233489990234375, "pos_frac": 0.625, "sample": [2.0369033813476562, 0.19202804565429688, -0.09365272521972656, 0.5842056274414062, 1.2043266296386719, 1.8262481689453125, 1.3493881225585938, -0.590484619140625, -1.5945205688476562, 0.24564743041992188, 1.1100540161132812, 0.6810684204101562, 1.2169265747070312, -0.014589309692382812, -0.4178657531738281, 0.7445831298828125, -0.3027801513671875, -1.2144432067871094, -1.04547119140625, 0.7350997924804688, 0.6638202667236328, 1.02838134765625, 0.3759880065917969, 0.5958023071289062, -0.8173446655273438, 2.0561447143554688, 0.845977783203125, -0.02093505859375, -0.27007293701171875, -0.6474685668945312, -1.39752197265625, -1.42901611328125, 3.8233489990234375, 0.3579864501953125, -0.25774383544921875, 0.5840606689453125, -0.9084548950195312, 0.7763099670410156, 1.1656951904296875, 0.378021240234375, 0.7750587463378906, 1.9016265869140625, 0.3908538818359375, 1.3183631896972656, -1.2583160400390625, 0.921783447265625, 0.2623939514160156, -0.2710990905761719, 0.14886474609375, 0.34246063232421875, 1.3622169494628906, 1.1616668701171875, 2.8596649169921875, 2.9614410400390625, 1.4706497192382812, -0.9652462005615234, 0.085174560546875, 0.56719970703125, -0.5143051147460938, 0.23289108276367188, -0.9270706176757812, -3.6255035400390625, -1.2818183898925781, -0.5042228698730469], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000088.npy"}
{"epoch": 0.1330309901738473, "step": 89, "batch_size": 64, "mean": 0.7864429354667664, "std": 1.3663591146469116, "min": -2.0570640563964844, "p10": -0.5997827529907226, "median": 0.5552225112915039, "p90": 2.4484792709350596, "max": 5.804290771484375, "pos_frac": 0.6875, "sample": [0.3003196716308594, 0.691680908203125, 0.3793449401855469, -0.27950286865234375, 0.480255126953125, 2.6716880798339844, -0.6997909545898438, 0.5282974243164062, 0.010852813720703125, 1.3194351196289062, 0.9584274291992188, 1.9310302734375, 1.30474853515625, 5.804290771484375, 0.4456596374511719, 0.0915679931640625, 1.0708084106445312, 0.39404296875, -0.4347038269042969, -0.42401123046875, -0.48919677734375, -0.5573558807373047, 0.6130905151367188, 2.0246810913085938, -0.6965141296386719, 4.764495849609375, -0.02063751220703125, 1.3941688537597656, -0.42452049255371094, -0.6179656982421875, 0.6702117919921875, -0.16741943359375, 0.7113151550292969, -0.28961181640625, 2.0741729736328125, 0.5821475982666016, 0.7149505615234375, 1.858154296875, 2.59710693359375, 1.6874847412109375, 1.1812210083007812, 1.2684097290039062, 1.73260498046875, 3.542510986328125, 0.26538658142089844, 0.011547088623046875, 1.5466461181640625, 2.77484130859375, 1.7729568481445312, -0.7204666137695312, 0.168426513671875, 1.4733333587646484, -0.19594573974609375, -0.1774444580078125, 1.412567138671875, -1.3929901123046875, 1.0528564453125, -0.9092674255371094, -0.4522514343261719, -2.0570640563964844, 2.5449981689453125, 0.3470115661621094, 2.223268508911133, -0.05400848388671875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000089.npy"}
{"epoch": 0.1345427059712774, "step": 90, "batch_size": 64, "mean": 0.5627224445343018, "std": 1.1510341167449951, "min": -3.0887451171875, "p10": -0.7984184265136719, "median": 0.5384330749511719, "p90": 1.8114295959472657, "max": 3.8298797607421875, "pos_frac": 0.703125, "sample": [-0.07997322082519531, 1.22564697265625, 0.4889068603515625, 0.8548355102539062, 0.3980255126953125, 0.24237823486328125, 1.5369224548339844, 3.2417373657226562, 0.011383056640625, -3.0887451171875, 1.7146263122558594, 1.375701904296875, 1.0176963806152344, 0.4756793975830078, 1.8967437744140625, -0.8107986450195312, 0.86212158203125, 1.4927215576171875, -0.8654136657714844, 0.97869873046875, 1.4162330627441406, -2.1235809326171875, 1.925567626953125, -0.17021942138671875, 1.6131591796875, 0.9056892395019531, 0.34894561767578125, 1.767578125, 0.3616046905517578, 2.505401611328125, -0.40738677978515625, -0.8729171752929688, 0.8308486938476562, 1.820404052734375, 0.195098876953125, 0.76763916015625, -0.5840530395507812, -1.0971832275390625, 3.8298797607421875, 1.1426563262939453, -0.76953125, 1.1809463500976562, 1.8673553466796875, -0.1914215087890625, 0.1723785400390625, 0.6866111755371094, 0.5665550231933594, 0.5103111267089844, 0.48126983642578125, 1.3697776794433594, -1.0423812866210938, 0.8252182006835938, 0.31032562255859375, -0.10526657104492188, 0.4212837219238281, -0.056774139404296875, 1.3342514038085938, -0.37023353576660156, -0.5413360595703125, 0.6880760192871094, -0.6476898193359375, 1.7904891967773438, -0.6343593597412109, 1.0241203308105469], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000090.npy"}
{"epoch": 0.1360544217687075, "step": 91, "batch_size": 64, "mean": 0.5365312099456787, "std": 1.3683608770370483, "min": -2.6495132446289062, "p10": -1.0734870910644532, "median": 0.3691596984863281, "p90": 2.2688899993896485, "max": 4.95208740234375, "pos_frac": 0.671875, "sample": [2.12139892578125, 0.650238037109375, 0.9214553833007812, -0.1262493133544922, 1.18603515625, 2.3019790649414062, 0.14233779907226562, -0.5258369445800781, 1.3246231079101562, 3.403076171875, 2.281005859375, -1.1841659545898438, 0.7333393096923828, -0.21065521240234375, 0.8172731399536133, -0.3769721984863281, 0.3793792724609375, -0.46845245361328125, 0.3385887145996094, -2.6495132446289062, 1.1436119079589844, 2.8684234619140625, 1.6434249877929688, 0.04063987731933594, 0.833740234375, 0.3083782196044922, 1.0325241088867188, -0.6936435699462891, -0.9744186401367188, 0.35894012451171875, 0.3527374267578125, 2.616851806640625, -1.708221435546875, 0.09553146362304688, 2.240619659423828, -0.25263214111328125, -1.0465545654296875, 1.8121566772460938, -1.0850296020507812, -1.2246932983398438, 0.756134033203125, -1.4251708984375, 0.5389862060546875, 0.17523193359375, 0.27971649169921875, 2.5051803588867188, 0.6376190185546875, 1.7700233459472656, 1.5159111022949219, 2.1469497680664062, -0.5011978149414062, -0.6414222717285156, -0.8819198608398438, -2.2106761932373047, 1.6648941040039062, 1.2004432678222656, 1.17047119140625, -0.30442047119140625, 4.95208740234375, 0.27832794189453125, -0.7557220458984375, 0.7003440856933594, 1.2413063049316406, 0.10363006591796875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000091.npy"}
{"epoch": 0.13756613756613756, "step": 92, "batch_size": 64, "mean": 0.6379896402359009, "std": 1.279119610786438, "min": -1.5874481201171875, "p10": -0.8071498870849609, "median": 0.29758644104003906, "p90": 2.4166580200195313, "max": 4.079185485839844, "pos_frac": 0.65625, "sample": [0.579345703125, 0.5308456420898438, -0.778350830078125, -0.28008270263671875, -1.5874481201171875, 0.24286270141601562, -0.5175247192382812, 1.0530433654785156, -0.6331787109375, -0.3135795593261719, -0.1087188720703125, 0.28000640869140625, 1.40924072265625, 1.6560745239257812, 1.4515037536621094, 1.4840068817138672, 0.4719390869140625, 0.1345081329345703, 2.3970794677734375, -0.001094818115234375, 1.5777740478515625, -0.7772140502929688, 0.6992340087890625, 0.73236083984375, 3.7412872314453125, 0.08952713012695312, -1.2261428833007812, -0.39644622802734375, -1.2306671142578125, 1.2078704833984375, -0.17557907104492188, 1.4979667663574219, 2.083251953125, -0.08202552795410156, 2.5661849975585938, 0.36574554443359375, 1.091033935546875, 1.6564178466796875, -1.3041229248046875, -0.036144256591796875, 0.22154998779296875, -0.2231597900390625, 4.0564727783203125, 3.0755996704101562, 0.3151664733886719, -1.0466079711914062, 0.19606781005859375, 1.4391708374023438, 0.21361923217773438, -0.8194923400878906, 0.8585624694824219, 2.425048828125, 0.8745880126953125, -0.09808540344238281, 0.2557220458984375, -0.90179443359375, 4.079185485839844, 0.989776611328125, 0.17946243286132812, 1.718109130859375, 3.0054931640625, 0.46303367614746094, -0.01497650146484375, 0.018032073974609375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000092.npy"}
{"epoch": 0.13907785336356765, "step": 93, "batch_size": 64, "mean": 0.4430374205112457, "std": 1.612220287322998, "min": -3.1591339111328125, "p10": -1.4936698913574218, "median": 0.3513774871826172, "p90": 2.469416046142578, "max": 6.0891571044921875, "pos_frac": 0.609375, "sample": [-0.47803688049316406, 1.3966140747070312, -1.7490234375, -0.09004974365234375, -3.002899169921875, 1.0198326110839844, -0.7750091552734375, -1.5531005859375, 0.5176906585693359, 0.4525108337402344, 3.5286712646484375, 2.865711212158203, 6.0891571044921875, -1.001190185546875, -0.7405128479003906, 0.09471321105957031, 0.7206649780273438, 1.6011199951171875, 0.14397048950195312, -1.3848876953125, 1.587371826171875, 3.3418121337890625, -0.39189910888671875, -0.21451568603515625, 0.3657493591308594, -0.00559234619140625, 2.1701812744140625, 0.07920074462890625, -0.7036590576171875, 0.6025810241699219, 0.337005615234375, 0.806915283203125, 0.227935791015625, 2.69195556640625, -3.1591339111328125, 0.2391204833984375, -0.2859153747558594, 0.8879375457763672, -1.5402908325195312, -0.5595855712890625, 1.8742256164550781, 0.40293121337890625, 1.660329818725586, 1.1364250183105469, -0.8379936218261719, -0.972747802734375, 0.9873580932617188, 1.0695457458496094, 3.5944366455078125, 0.7604217529296875, 1.3610954284667969, 0.7519969940185547, 2.48175048828125, 0.9837417602539062, -1.70599365234375, -0.6363964080810547, 2.4406356811523438, -1.9305953979492188, -1.1480865478515625, 1.863067626953125, 1.0999565124511719, -0.922149658203125, -0.33277130126953125, 0.24008941650390625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000093.npy"}
{"epoch": 0.14058956916099774, "step": 94, "batch_size": 64, "mean": 1.0405027866363525, "std": 1.6829571723937988, "min": -2.6143035888671875, "p10": -0.8775829315185545, "median": 0.8851957321166992, "p90": 3.3516105651855472, "max": 5.88702392578125, "pos_frac": 0.796875, "sample": [3.3687515258789062, 3.6128921508789062, -0.97698974609375, 1.4512271881103516, 0.5741958618164062, -2.1190948486328125, 0.5418338775634766, -0.0023193359375, 0.8015975952148438, 3.714324951171875, 0.4459362030029297, 1.8337364196777344, 1.672210693359375, -2.458599090576172, 0.8823795318603516, 1.3838272094726562, 0.7467575073242188, 0.6921291351318359, 0.99090576171875, 0.36435699462890625, 1.8103561401367188, 1.4200210571289062, 0.8880119323730469, 0.5975494384765625, 1.8321685791015625, 0.2139415740966797, 0.46230316162109375, 0.6400527954101562, -1.4311599731445312, 0.968963623046875, 1.1878490447998047, 0.49080657958984375, 0.83880615234375, -1.92498779296875, 0.9849319458007812, 1.2821846008300781, 2.3142471313476562, 3.2254791259765625, 0.32981109619140625, 2.4929046630859375, 2.3434600830078125, 1.2185745239257812, -0.5118865966796875, 5.88702392578125, 2.123809814453125, 1.2584152221679688, 4.2672119140625, 4.586456298828125, 0.5550994873046875, 1.913421630859375, 5.147979736328125, -0.6345405578613281, -0.30728912353515625, -0.6456336975097656, 1.4458141326904297, -0.18532371520996094, -2.6143035888671875, 0.9664688110351562, 3.311614990234375, 0.8988227844238281, 0.002109527587890625, 0.0303955078125, -1.202779769897461, 0.5929527282714844], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000094.npy"}
{"epoch": 0.1421012849584278, "step": 95, "batch_size": 64, "mean": 0.6995913982391357, "std": 1.567514181137085, "min": -2.35308837890625, "p10": -1.3268573760986326, "median": 0.5179023742675781, "p90": 2.994982147216797, "max": 3.8513946533203125, "pos_frac": 0.671875, "sample": [3.095001220703125, -1.2419929504394531, -0.6547012329101562, -1.1086349487304688, -2.1048660278320312, -0.22286224365234375, 3.8513946533203125, 1.5319595336914062, 3.3466415405273438, 1.913330078125, 0.6285171508789062, 0.4007568359375, 0.075897216796875, 0.29412841796875, -1.054779052734375, 1.8952865600585938, -1.97509765625, -0.5821685791015625, -1.8721427917480469, 0.2963294982910156, 1.273681640625, 3.102325439453125, 0.3658103942871094, 1.5071182250976562, -1.670562744140625, -0.3135986328125, 1.267578125, 2.514892578125, -1.3632278442382812, -2.35308837890625, 0.3185882568359375, 1.313385009765625, 1.0998916625976562, 1.36663818359375, -0.26486968994140625, 0.6896743774414062, 2.1759414672851562, 1.9251861572265625, -0.8135986328125, -0.14624404907226562, 0.40728759765625, -0.0355072021484375, 2.9940261840820312, 0.6399154663085938, -1.7018585205078125, 2.995391845703125, 2.97003173828125, 0.07566070556640625, 0.8963050842285156, 3.6078262329101562, 0.7678070068359375, 3.1573486328125, -0.9265632629394531, 0.05574798583984375, 0.8380393981933594, 2.41455078125, -0.3554363250732422, 2.4400482177734375, 0.019001007080078125, 1.030853271484375, 0.00823974609375, 2.2936363220214844, 2.222015380859375, -0.5480384826660156], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000095.npy"}
{"epoch": 0.1436130007558579, "step": 96, "batch_size": 64, "mean": 0.8266537189483643, "std": 1.5834839344024658, "min": -2.8252639770507812, "p10": -0.7659448623657227, "median": 0.543701171875, "p90": 2.9516471862792972, "max": 4.5447540283203125, "pos_frac": 0.671875, "sample": [0.7867927551269531, -0.33043670654296875, 2.2796630859375, -1.5478591918945312, 1.5724029541015625, -0.460235595703125, 1.2913894653320312, -0.4380512237548828, 1.179107666015625, 0.44588470458984375, 3.8894424438476562, -0.052032470703125, 0.6705398559570312, 3.276458740234375, 2.0870285034179688, 2.4275360107421875, -1.5291976928710938, 0.045680999755859375, -0.7768001556396484, -1.1938972473144531, 0.5157241821289062, -0.1787261962890625, 1.6069717407226562, -2.8252639770507812, 0.12729644775390625, 0.5818500518798828, 1.8776969909667969, 0.7154769897460938, 2.0733184814453125, -0.073638916015625, 2.8347549438476562, 2.56683349609375, 2.97998046875, 3.2440567016601562, -0.04835700988769531, 2.6490325927734375, -2.1714630126953125, 0.18407058715820312, 0.169921875, 1.0251197814941406, 2.313934326171875, 0.30709075927734375, 0.5770950317382812, -0.3387908935546875, 4.426910400390625, 0.02204132080078125, 2.352773666381836, -0.7406158447265625, -0.09201812744140625, -0.4488639831542969, 0.6268768310546875, -0.3740386962890625, 0.4342803955078125, 4.5447540283203125, 3.7059783935546875, 1.180023193359375, -0.45139312744140625, 0.5716781616210938, 0.40947723388671875, 1.334136962890625, 2.8855361938476562, 0.46260833740234375, -0.654296875, -1.6274147033691406], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000096.npy"}
{"epoch": 0.14512471655328799, "step": 97, "batch_size": 64, "mean": 0.3047705590724945, "std": 1.5680749416351318, "min": -3.4212608337402344, "p10": -1.562166976928711, "median": 0.3930511474609375, "p90": 2.063062286376953, "max": 3.9620513916015625, "pos_frac": 0.609375, "sample": [-2.4046173095703125, -0.35593414306640625, -0.5098285675048828, 2.852315902709961, 0.4349861145019531, 3.9620513916015625, -1.4007148742675781, -0.19608306884765625, -1.2307510375976562, -2.445831298828125, -2.150684356689453, 2.6830825805664062, 0.17490386962890625, 1.8659629821777344, -1.5673980712890625, -3.4212608337402344, 0.6303367614746094, 0.3557281494140625, 1.3429412841796875, 0.9071807861328125, 0.3619403839111328, 1.6197891235351562, 0.08036041259765625, -1.0405044555664062, 1.9814910888671875, 0.7451705932617188, -2.5949859619140625, 0.19446563720703125, 0.41228485107421875, 1.0061264038085938, 0.37662506103515625, -1.459228515625, 1.2511672973632812, 1.2664222717285156, -0.8157539367675781, 1.1262741088867188, 3.12396240234375, 1.8893241882324219, -0.39708709716796875, -0.3347930908203125, 1.2909774780273438, -1.157257080078125, 2.011627197265625, 1.1315765380859375, 0.6378936767578125, -0.546966552734375, 1.9835319519042969, 2.48016357421875, 0.981658935546875, -1.108551025390625, -0.6408729553222656, 0.40947723388671875, -2.633941650390625, -1.5499610900878906, 2.9267807006835938, 1.950439453125, 0.9668045043945312, -0.17646026611328125, 2.0851058959960938, 1.4478302001953125, -1.3379135131835938, 0.6729660034179688, 0.040283203125, -0.6793136596679688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000097.npy"}
{"epoch": 0.14663643235071808, "step": 98, "batch_size": 64, "mean": 0.42344963550567627, "std": 1.4205034971237183, "min": -4.289306640625, "p10": -1.2196895599365232, "median": 0.6013507843017578, "p90": 1.8525535583496098, "max": 4.286651611328125, "pos_frac": 0.65625, "sample": [-0.7963943481445312, 0.10751628875732422, 1.5559196472167969, 1.884613037109375, 1.5944442749023438, 0.7845306396484375, 1.16741943359375, 1.3629608154296875, -0.6709060668945312, 4.286651611328125, -2.756317138671875, -1.3934211730957031, 0.9888916015625, -0.11195755004882812, -2.5647735595703125, 1.6285018920898438, 0.9894638061523438, 1.942779541015625, 0.21864700317382812, -0.9968833923339844, 1.0717506408691406, 3.1244888305664062, 0.9458770751953125, 2.0958251953125, 0.9349632263183594, -0.9833469390869141, 3.158344268798828, 1.1698646545410156, 1.1367931365966797, 0.53253173828125, 1.7777481079101562, 1.3499832153320312, -0.7614212036132812, 0.63299560546875, -0.47402191162109375, 0.8623809814453125, 1.3291397094726562, 0.30773162841796875, 0.8680267333984375, -0.4270057678222656, 1.026153564453125, -0.21605682373046875, -2.005157470703125, 0.5697059631347656, 1.1713485717773438, -1.9111976623535156, -0.07106399536132812, 0.48630523681640625, -0.251708984375, 1.0561676025390625, -0.08650588989257812, 0.5138397216796875, -0.24910736083984375, -0.7386360168457031, -4.289306640625, 1.26287841796875, 0.9172096252441406, 0.33812904357910156, 0.3070831298828125, 0.1144561767578125, 2.4922332763671875, 0.6772842407226562, -0.5724334716796875, -1.3151779174804688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000098.npy"}
{"epoch": 0.14814814814814814, "step": 99, "batch_size": 64, "mean": 0.6521108150482178, "std": 1.9971574544906616, "min": -3.5272369384765625, "p10": -1.6933406829833983, "median": 0.5056514739990234, "p90": 3.144923782348633, "max": 5.9309844970703125, "pos_frac": 0.609375, "sample": [5.564109802246094, 3.154705047607422, -0.246368408203125, 4.76763916015625, 0.491912841796875, -1.2816162109375, -0.3169097900390625, -2.266946792602539, 0.6963958740234375, 2.733013153076172, 1.40380859375, 1.726470947265625, 0.5649528503417969, 1.9547309875488281, -0.06889915466308594, 2.294921875, -0.4494972229003906, 1.1346473693847656, 0.3536796569824219, 0.8354682922363281, 3.5145111083984375, -0.4271812438964844, -0.9185752868652344, -1.5315132141113281, -1.4375152587890625, -2.02447509765625, 4.927825927734375, -0.8791046142578125, 0.3323822021484375, 5.479881286621094, 0.8107490539550781, 2.0380859375, -1.3606796264648438, 0.9078598022460938, 0.12275314331054688, 1.59088134765625, -0.12518310546875, 0.49074745178222656, -1.3577537536621094, 1.8516502380371094, -0.04378318786621094, -3.5272369384765625, 3.122100830078125, 5.9309844970703125, -0.3275146484375, 1.9498519897460938, 1.9979705810546875, 1.3892803192138672, -1.7626953125, 0.8964443206787109, 0.5193901062011719, 1.2613105773925781, 0.6326103210449219, 0.6810722351074219, 0.466339111328125, -0.20342254638671875, -0.3632621765136719, 0.5994720458984375, 0.3399848937988281, -3.4301986694335938, -0.11035919189453125, -1.8447914123535156, 0.5578765869140625, -2.0478973388671875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000099.npy"}
{"epoch": 0.14965986394557823, "step": 100, "batch_size": 64, "mean": 0.7829455137252808, "std": 2.3182389736175537, "min": -4.6741943359375, "p10": -1.7133598327636719, "median": 0.7657985687255859, "p90": 3.1803970336914062, "max": 9.00396728515625, "pos_frac": 0.703125, "sample": [2.013153076171875, -1.002105712890625, -1.9051513671875, 1.6750106811523438, 2.4154205322265625, -0.6165847778320312, 0.46387290954589844, 0.7644309997558594, 3.2035369873046875, 0.5286331176757812, 5.36993408203125, 1.5780563354492188, 0.5524215698242188, 4.535614013671875, -0.40755462646484375, 4.53961181640625, -1.4118194580078125, 0.33785247802734375, 1.3137626647949219, 1.0282440185546875, 0.9112701416015625, -1.48333740234375, -1.0682106018066406, 0.2657508850097656, -1.7712936401367188, 2.3392715454101562, 1.3778057098388672, 2.1569747924804688, 0.11496543884277344, -4.6741943359375, 0.01300048828125, 0.36077880859375, -1.2773895263671875, 0.071685791015625, 0.3955650329589844, -1.5717353820800781, 2.4430084228515625, 0.8969345092773438, -4.5395050048828125, 2.2057723999023438, 0.790740966796875, 6.4942169189453125, -1.7194137573242188, -0.13324356079101562, -0.8072357177734375, 2.3693161010742188, 1.0750808715820312, 0.2731781005859375, -0.12122535705566406, 0.921722412109375, 2.26641845703125, 3.12640380859375, 0.9641189575195312, 1.6303138732910156, 1.2454986572265625, 0.9203929901123047, -2.392730712890625, 3.983734130859375, 1.9833183288574219, 0.7671661376953125, 9.00396728515625, 0.61163330078125, -3.589080810546875, -1.6992340087890625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000100.npy"}
{"epoch": 0.15117157974300832, "step": 101, "batch_size": 64, "mean": 0.9559941291809082, "std": 1.6888560056686401, "min": -4.6527557373046875, "p10": -0.8317363739013671, "median": 0.9230384826660156, "p90": 3.1329925537109378, "max": 4.86810302734375, "pos_frac": 0.765625, "sample": [0.13413238525390625, 0.4856109619140625, 1.22576904296875, 0.0139923095703125, 3.8685684204101562, 2.179962158203125, 0.9970169067382812, 0.5001430511474609, -0.2236175537109375, -0.23731613159179688, 3.7075271606445312, 1.7386550903320312, -4.6527557373046875, 0.15965652465820312, 2.9431419372558594, -0.02904510498046875, 1.4110488891601562, 1.0322856903076172, 0.5523300170898438, -1.1417808532714844, 1.9925270080566406, 2.758514404296875, 3.6171035766601562, 1.4342308044433594, 0.7442512512207031, -0.28031158447265625, 0.5276145935058594, 0.8606643676757812, 0.92388916015625, 2.476959228515625, -0.6530685424804688, 3.0336151123046875, 1.6485214233398438, 0.5051822662353516, -0.8731307983398438, -0.7067089080810547, 0.65045166015625, -2.3221282958984375, -1.2266674041748047, 1.49749755859375, -0.4687652587890625, 0.9221878051757812, 0.8106727600097656, 0.5517005920410156, 0.38895416259765625, 1.9552230834960938, 1.3700180053710938, -3.6202392578125, -0.7351493835449219, 0.707550048828125, 3.0023422241210938, 1.7551307678222656, 3.354602813720703, 1.2482242584228516, 1.312469482421875, 0.074615478515625, 3.1755828857421875, 2.1583328247070312, -1.1144866943359375, 4.86810302734375, 1.377716064453125, 2.255495071411133, 3.26092529296875, 1.2980880737304688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000101.npy"}
{"epoch": 0.15268329554043839, "step": 102, "batch_size": 64, "mean": 0.5958476662635803, "std": 2.022280216217041, "min": -4.633369445800781, "p10": -1.6086105346679687, "median": 0.3785114288330078, "p90": 3.336316680908203, "max": 7.32586669921875, "pos_frac": 0.609375, "sample": [1.8752593994140625, -0.7658004760742188, -0.8500556945800781, 0.36151123046875, -3.3769989013671875, -0.6771030426025391, -1.6234283447265625, 1.0376167297363281, -2.2567367553710938, 0.4198570251464844, 0.058441162109375, 1.0877761840820312, -1.57403564453125, 0.8677330017089844, 1.569091796875, -1.0656204223632812, -1.0407772064208984, 3.6086883544921875, 0.06237030029296875, 4.622161865234375, -1.0557403564453125, 0.5741500854492188, 1.3858642578125, 0.17975616455078125, 0.9642333984375, 0.22150421142578125, 0.9076080322265625, 2.2743301391601562, -4.633369445800781, -0.0036144256591796875, 4.65509033203125, 1.8923187255859375, -1.6762275695800781, 1.1509933471679688, -0.67962646484375, 3.2985000610351562, 2.160594940185547, 4.118267059326172, 1.9915313720703125, -1.5167808532714844, 1.6829452514648438, 3.7904205322265625, -1.1655349731445312, -0.37665557861328125, -0.6335372924804688, 1.0831413269042969, -1.095428466796875, 1.6395950317382812, 0.0749664306640625, -0.10403823852539062, 1.5264205932617188, 2.77056884765625, 0.3955116271972656, -1.9650135040283203, -0.5911521911621094, -1.627471923828125, -0.3575897216796875, 7.32586669921875, -0.8888320922851562, 0.2401580810546875, 2.2388992309570312, 3.3525238037109375, 1.2654495239257812, 1.0037040710449219], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000102.npy"}
{"epoch": 0.15419501133786848, "step": 103, "batch_size": 64, "mean": 1.2109458446502686, "std": 2.1130940914154053, "min": -2.87371826171875, "p10": -1.125636672973633, "median": 1.0452308654785156, "p90": 3.513968658447266, "max": 7.4670867919921875, "pos_frac": 0.75, "sample": [-2.3249893188476562, 2.8156814575195312, 3.21759033203125, -2.6249008178710938, 0.752349853515625, 1.8684806823730469, 0.631744384765625, 0.3672676086425781, 2.0212631225585938, 1.0104141235351562, -0.8124122619628906, -0.39470672607421875, -1.1303253173828125, 1.063446044921875, 3.3220748901367188, 3.6992034912109375, 0.9224281311035156, 1.140289306640625, 5.93310546875, 2.9170150756835938, 0.0580902099609375, 3.2645111083984375, 0.6562042236328125, 2.1535072326660156, 2.022296905517578, -0.9668502807617188, 0.707977294921875, 5.9386749267578125, -0.8327369689941406, 2.0040664672851562, 2.018951416015625, 7.4670867919921875, 2.07611083984375, 0.6519699096679688, 0.524200439453125, 0.396270751953125, -1.2705459594726562, 3.439483642578125, 1.1421966552734375, 0.13295745849609375, 3.2497520446777344, 1.0270156860351562, 1.1361732482910156, -1.1146965026855469, 1.2017650604248047, -2.03173828125, 3.3906021118164062, -0.5080490112304688, 3.7485885620117188, 0.09700393676757812, 2.741546630859375, 2.0120697021484375, -2.87371826171875, -0.8453788757324219, -0.5274200439453125, 5.4959869384765625, -2.6643142700195312, 0.2900123596191406, -0.9920387268066406, 0.6251754760742188, 1.4556159973144531, 1.3792457580566406, 3.5458908081054688, 1.6820068359375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000103.npy"}
{"epoch": 0.15570672713529857, "step": 104, "batch_size": 64, "mean": 1.1380459070205688, "std": 2.38462233543396, "min": -4.3871307373046875, "p10": -1.9678508758544921, "median": 1.1731834411621094, "p90": 3.7329772949218754, "max": 8.367218017578125, "pos_frac": 0.6875, "sample": [1.53460693359375, 2.51800537109375, 1.530120849609375, 1.2161941528320312, -2.158294677734375, 2.379669189453125, -0.5689849853515625, 1.2329254150390625, 0.05423736572265625, -1.222900390625, 0.34679412841796875, -4.3871307373046875, 8.367218017578125, 3.617279052734375, -1.2719573974609375, -0.11595916748046875, 6.0234375, -0.9397239685058594, 2.81695556640625, 3.4872283935546875, -2.913055419921875, -2.366180419921875, 2.401763916015625, 1.021392822265625, -1.00811767578125, 0.17284774780273438, -0.3062248229980469, -2.23773193359375, 3.443431854248047, -0.7722930908203125, 0.7879543304443359, 0.5470428466796875, -1.3070144653320312, 1.1301727294921875, 1.7736244201660156, 2.0756149291992188, 1.2705001831054688, -0.9860343933105469, 2.99261474609375, 1.6231765747070312, 0.1370697021484375, -1.9433937072753906, 5.363800048828125, 0.8307075500488281, 3.203266143798828, 6.006439208984375, 2.5232505798339844, 0.9051933288574219, -0.627349853515625, -2.0263290405273438, 1.9652519226074219, 1.552896499633789, 2.26318359375, 3.782562255859375, 0.37618255615234375, 2.168304443359375, -0.8263397216796875, 1.240234375, 0.7939853668212891, 2.4787750244140625, -1.97833251953125, 2.58843994140625, 5.477272033691406, 4.776664733886719], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000104.npy"}
{"epoch": 0.15721844293272866, "step": 105, "batch_size": 64, "mean": 0.9921829700469971, "std": 2.833014726638794, "min": -5.7185821533203125, "p10": -3.074233245849609, "median": 1.0794754028320312, "p90": 5.333284759521485, "max": 7.858917236328125, "pos_frac": 0.640625, "sample": [-0.32059478759765625, 3.0414352416992188, 1.998687744140625, 1.939291000366211, 0.4196205139160156, 2.234649658203125, -5.7185821533203125, 1.7939453125, -2.9793472290039062, 1.4724006652832031, 3.959075927734375, 1.4889755249023438, -0.7627830505371094, 2.52880859375, -0.02513885498046875, 6.932472229003906, -0.5453701019287109, -1.2011489868164062, -0.03961753845214844, 5.637016296386719, 2.1178436279296875, 1.1836128234863281, 0.882598876953125, -4.174995422363281, 1.0934829711914062, 0.685516357421875, 5.496009826660156, -1.1868438720703125, 1.34307861328125, -3.576385498046875, -3.114898681640625, 1.9785003662109375, -0.15688323974609375, 2.3887786865234375, -4.1153717041015625, 7.858917236328125, 0.18436431884765625, 0.0489349365234375, 1.1767654418945312, 0.3202934265136719, 1.0654678344726562, -0.06476593017578125, -0.7605743408203125, 1.6616554260253906, -3.251251220703125, 1.8866081237792969, 5.808258056640625, -1.0598335266113281, 0.6877593994140625, -1.4265518188476562, -3.776336669921875, 1.4832496643066406, -0.6907119750976562, 5.363311767578125, 7.1873931884765625, 1.3121337890625, 3.0090980529785156, 1.1240310668945312, -0.120391845703125, 2.8189525604248047, -1.973419189453125, 5.263221740722656, 0.7767410278320312, 4.8885498046875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000105.npy"}
{"epoch": 0.15873015873015872, "step": 106, "batch_size": 64, "mean": 0.8636230826377869, "std": 2.386996030807495, "min": -7.4319000244140625, "p10": -1.9668830871582028, "median": 0.5937986373901367, "p90": 3.4448108673095703, "max": 7.3133544921875, "pos_frac": 0.671875, "sample": [-0.5177154541015625, 1.3384857177734375, 0.3945732116699219, 3.3665847778320312, -1.6941452026367188, -0.6443099975585938, 4.786567687988281, 0.8306083679199219, 0.849945068359375, 5.677978515625, -0.14154052734375, 4.259815216064453, 5.428108215332031, 1.8233566284179688, 0.7957363128662109, -2.1103286743164062, -0.5254249572753906, 3.3605117797851562, 0.26959228515625, 0.4754486083984375, 1.0438461303710938, 0.4889984130859375, -4.42584228515625, -0.711669921875, -0.46024131774902344, 0.6048583984375, -0.6141853332519531, -2.083770751953125, 0.46710968017578125, 3.3001327514648438, 0.5203933715820312, -0.8742618560791016, 0.5810394287109375, 2.733551025390625, 2.9217147827148438, 0.7643013000488281, -0.08797645568847656, -0.30667686462402344, 1.1918907165527344, 7.3133544921875, 2.1552085876464844, 1.562835693359375, 1.7288322448730469, -0.30091285705566406, -2.7262649536132812, -0.25087738037109375, 2.3481369018554688, 2.0066299438476562, -7.4319000244140625, 0.3299064636230469, -2.9029693603515625, 0.6588973999023438, 3.0311813354492188, 3.4204978942871094, 2.3744277954101562, 3.7107696533203125, 0.9895858764648438, 0.21713638305664062, -0.20164871215820312, 3.455230712890625, -2.784393310546875, 0.17915916442871094, 2.7292556762695312, 0.5827388763427734], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000106.npy"}
{"epoch": 0.1602418745275888, "step": 107, "batch_size": 64, "mean": 0.6637719869613647, "std": 2.2557008266448975, "min": -6.901123046875, "p10": -1.810132598876953, "median": 0.7541618347167969, "p90": 3.2717353820800783, "max": 5.393867492675781, "pos_frac": 0.6875, "sample": [0.9819259643554688, -4.4051666259765625, 1.1172943115234375, 0.9763641357421875, 0.8242950439453125, 1.99176025390625, -3.974578857421875, 2.574981689453125, 0.48435211181640625, 1.7404518127441406, -0.157196044921875, 5.393867492675781, 1.1629581451416016, 1.2497406005859375, 0.7887191772460938, -1.1359786987304688, -3.546966552734375, 0.7196044921875, 2.876750946044922, 1.1859283447265625, -0.333892822265625, 3.7959442138671875, 4.8833465576171875, -1.4944534301757812, 0.9022903442382812, -1.0277938842773438, 2.682605743408203, -0.4670600891113281, -1.0330657958984375, 1.1063461303710938, 0.1107177734375, -1.8720474243164062, 0.4690093994140625, 2.7719345092773438, 0.0954132080078125, 3.776947021484375, 3.502504348754883, -1.0183219909667969, 0.04473876953125, 2.9807205200195312, 0.926605224609375, 2.903472900390625, -0.318389892578125, -0.23444366455078125, 1.578653335571289, 2.8265762329101562, 0.5317039489746094, 3.0914230346679688, 0.353607177734375, -3.0282745361328125, 4.072998046875, 3.2057418823242188, -0.7696647644042969, -1.9123878479003906, 0.2115631103515625, 1.1317214965820312, 2.8664703369140625, 3.300018310546875, 0.048000335693359375, 0.5035552978515625, -1.5807571411132812, -6.901123046875, 0.6150131225585938, -1.6656646728515625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000107.npy"}
{"epoch": 0.1617535903250189, "step": 108, "batch_size": 64, "mean": 0.8531872034072876, "std": 2.328357219696045, "min": -5.1143341064453125, "p10": -2.1083612442016597, "median": 0.8174152374267578, "p90": 3.8634323120117195, "max": 6.430419921875, "pos_frac": 0.65625, "sample": [-0.7080612182617188, -0.4214019775390625, 1.3225860595703125, 2.4698486328125, 0.33904075622558594, 2.5762405395507812, 1.8424835205078125, 1.7999916076660156, 6.430419921875, -0.6358642578125, 3.0524978637695312, 2.0367813110351562, 1.4519729614257812, -2.2611141204833984, 1.3575897216796875, 1.0191230773925781, -1.08782958984375, 0.16160011291503906, -0.0188751220703125, 0.16775131225585938, 4.44488525390625, -0.6253509521484375, -1.6122245788574219, 0.266510009765625, 0.7742996215820312, -2.3454742431640625, 1.751312255859375, -5.1143341064453125, 1.759735107421875, -0.824981689453125, -2.90997314453125, -0.4614448547363281, 4.321380615234375, 0.2606964111328125, 3.3577728271484375, 5.532215118408203, 0.8782234191894531, -0.443389892578125, -3.8860740661621094, 3.3750076293945312, -1.7519378662109375, 3.9535064697265625, 2.3077354431152344, 0.8605308532714844, -1.4022598266601562, 3.014080047607422, -0.6511077880859375, 1.2409744262695312, -0.0062103271484375, -3.0185775756835938, -3.4316139221191406, 1.7888145446777344, 0.423797607421875, 2.26068115234375, 4.3422393798828125, -0.4339714050292969, 1.278167724609375, 6.0251922607421875, 0.3746185302734375, 1.2347793579101562, 0.3756895065307617, 2.6616439819335938, 0.11037445068359375, 3.65325927734375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000108.npy"}
{"epoch": 0.16326530612244897, "step": 109, "batch_size": 64, "mean": 1.3385167121887207, "std": 2.7469823360443115, "min": -3.7001495361328125, "p10": -1.5858459472656248, "median": 1.0993633270263672, "p90": 4.774440765380859, "max": 9.356430053710938, "pos_frac": 0.640625, "sample": [-1.1943206787109375, 5.8578643798828125, 4.8123779296875, -0.705108642578125, 3.4177627563476562, 1.9761962890625, 4.032054901123047, -1.3466644287109375, 0.13590240478515625, 2.5270233154296875, 0.17135238647460938, 1.938201904296875, 4.685920715332031, 1.8656539916992188, -1.634552001953125, 0.423828125, -0.42745399475097656, 1.4345760345458984, -0.44964599609375, 1.1066436767578125, 0.898284912109375, 1.9289207458496094, -3.1832427978515625, -1.3487625122070312, -0.413787841796875, 9.356430053710938, -0.520477294921875, -1.0780601501464844, 1.3094844818115234, 6.201904296875, 0.7810039520263672, -2.368927001953125, 1.0920829772949219, -0.19445419311523438, -1.1508674621582031, 1.7845077514648438, -2.0072402954101562, 4.5111083984375, 0.22089385986328125, 2.358196258544922, -0.170806884765625, 2.3870677947998047, 1.5991649627685547, 7.0848541259765625, -1.4036598205566406, -2.184093475341797, -0.7471466064453125, 1.238372802734375, -3.7001495361328125, -2.1252212524414062, 1.0528297424316406, 1.8872833251953125, 2.041290283203125, 3.9392852783203125, 2.4659996032714844, 9.3067626953125, 4.974613189697266, 4.1120452880859375, 3.174072265625, 3.448150634765625, -0.514617919921875, 0.012950897216796875, -1.472198486328125, 2.4536056518554688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000109.npy"}
{"epoch": 0.16477702191987906, "step": 110, "batch_size": 64, "mean": 0.2416144609451294, "std": 2.4553651809692383, "min": -7.534149169921875, "p10": -2.5034172058105466, "median": 0.3393135070800781, "p90": 3.04330062866211, "max": 5.515327453613281, "pos_frac": 0.53125, "sample": [1.503570556640625, -0.334930419921875, -1.4777679443359375, -1.4603118896484375, -3.0093631744384766, 4.8556976318359375, 2.3437728881835938, 3.4108810424804688, -4.958709716796875, 1.1598663330078125, 2.480194091796875, 0.5457763671875, 1.5401382446289062, 3.084716796875, 1.0518875122070312, -0.39237213134765625, 1.8030853271484375, 5.515327453613281, -0.83538818359375, -3.7689590454101562, 1.8724288940429688, -2.5523910522460938, 1.9806957244873047, 1.759613037109375, 0.29541778564453125, -0.19005584716796875, -0.00539398193359375, 2.9466629028320312, -0.44443511962890625, -0.7368850708007812, -1.8138885498046875, 2.428874969482422, 0.383209228515625, -0.010677337646484375, -2.1214065551757812, 1.286163330078125, -1.142730712890625, 2.8816375732421875, -1.877410888671875, 0.5415725708007812, 4.1620941162109375, -0.14875030517578125, -1.0504989624023438, 1.9919204711914062, -3.0528221130371094, 1.3449859619140625, 2.4946975708007812, -2.1364288330078125, -7.534149169921875, -5.2045440673828125, -0.36249542236328125, 1.5494461059570312, -1.0707244873046875, 0.9212265014648438, 0.24350738525390625, 4.28192138671875, 3.1174087524414062, 1.3472328186035156, 0.581207275390625, -0.7897186279296875, -2.279491424560547, -1.6199111938476562, 2.5282440185546875, -2.3891448974609375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000110.npy"}
{"epoch": 0.16628873771730915, "step": 111, "batch_size": 64, "mean": 0.9300180673599243, "std": 2.500761032104492, "min": -4.330024719238281, "p10": -1.8684600830078124, "median": 0.32891082763671875, "p90": 4.209584426879883, "max": 7.655059814453125, "pos_frac": 0.5625, "sample": [-1.808197021484375, -0.460784912109375, 2.8187255859375, 1.633514404296875, 4.673862457275391, -0.7887649536132812, 0.3450927734375, 2.62579345703125, 1.7161483764648438, 0.034786224365234375, 3.3413619995117188, -0.8454666137695312, -1.7735595703125, 3.4350738525390625, -4.330024719238281, -0.38958740234375, -0.9237480163574219, 0.1505451202392578, -3.4505958557128906, -0.171234130859375, 5.5957489013671875, 1.8959808349609375, 4.113323211669922, 5.509605407714844, 4.1065826416015625, 1.9923439025878906, 0.611114501953125, 2.318511962890625, 1.2820854187011719, -1.9679031372070312, -0.1686382293701172, -0.8964767456054688, -1.894287109375, -3.5923919677734375, 0.8590545654296875, -0.29091644287109375, -0.32158660888671875, 2.105499267578125, -0.6109809875488281, 2.5700149536132812, 2.7060089111328125, 3.4315948486328125, 5.254875183105469, -0.21407127380371094, 4.648136138916016, -1.2084674835205078, 1.9547958374023438, 4.2508392333984375, -1.7165069580078125, -0.3508033752441406, 3.6443252563476562, -0.112701416015625, 2.235902786254883, -1.2765350341796875, -3.3800506591796875, 1.162689208984375, -2.9250640869140625, 2.9724197387695312, -0.1864013671875, 0.0038604736328125, 7.655059814453125, -0.11576461791992188, 1.7246589660644531, 0.3127288818359375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000111.npy"}
{"epoch": 0.16780045351473924, "step": 112, "batch_size": 64, "mean": 1.2315394878387451, "std": 2.667114496231079, "min": -5.575101852416992, "p10": -2.1768554687499995, "median": 1.1737041473388672, "p90": 5.094884490966799, "max": 7.121673583984375, "pos_frac": 0.6875, "sample": [3.23443603515625, 1.14288330078125, 4.327117919921875, 0.9661788940429688, 0.5960044860839844, -0.01064300537109375, 3.7744064331054688, 1.8229141235351562, 0.378173828125, 0.8492774963378906, 6.319419860839844, 2.6895751953125, -4.163341522216797, 0.8526821136474609, 0.8170356750488281, -2.7136383056640625, 5.318275451660156, 1.3082046508789062, -0.6444244384765625, 0.63482666015625, 2.6158294677734375, -0.281646728515625, 2.2904090881347656, 5.7501983642578125, 0.20134735107421875, 5.483123779296875, 1.501373291015625, -2.325958251953125, 4.573638916015625, 1.2550888061523438, 7.121673583984375, 1.2615814208984375, -1.134359359741211, 2.4667701721191406, 2.9142837524414062, 5.8945465087890625, 0.3117847442626953, -1.6776046752929688, 3.0043106079101562, 2.2573623657226562, -2.538623809814453, -0.703765869140625, 3.8954238891601562, 1.5015811920166016, 1.44927978515625, 6.2860870361328125, 4.1200408935546875, 4.399776458740234, 0.21271514892578125, -1.14208984375, -3.1515464782714844, -0.10772705078125, -1.33892822265625, 2.3947067260742188, 0.574737548828125, -0.4796142578125, 1.2045249938964844, -5.575101852416992, -1.828948974609375, -0.263702392578125, 1.3598098754882812, -3.3790740966796875, -0.3128318786621094, 1.2586822509765625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000112.npy"}
{"epoch": 0.1693121693121693, "step": 113, "batch_size": 64, "mean": 0.917809784412384, "std": 2.615405321121216, "min": -7.3254852294921875, "p10": -1.2757476806640624, "median": 0.8273401260375977, "p90": 3.7368812561035156, "max": 9.00640869140625, "pos_frac": 0.625, "sample": [-0.8460006713867188, -3.532806396484375, -0.9055938720703125, 0.9854545593261719, -0.40377044677734375, 2.661449432373047, 0.9169464111328125, -1.2498779296875, -3.0634613037109375, 2.6448097229003906, -1.00250244140625, -1.1474647521972656, -1.5215644836425781, -0.22119140625, -2.9399642944335938, -0.45020294189453125, 0.6307048797607422, 2.7759437561035156, 1.8544235229492188, -5.350940704345703, -1.286834716796875, -0.2120819091796875, 5.107414245605469, 0.6354217529296875, 1.4112892150878906, 1.8415107727050781, 2.5464935302734375, 1.4475784301757812, 4.462028503417969, -0.3250999450683594, 1.8564834594726562, 9.00640869140625, 0.8427028656005859, 0.8119773864746094, 0.4100799560546875, 2.0294876098632812, -7.3254852294921875, 0.11897659301757812, 2.256927490234375, 2.1932830810546875, 0.09505081176757812, 1.6351852416992188, -0.16876220703125, 1.52435302734375, 2.3523521423339844, 6.6555938720703125, 0.43447113037109375, 1.8154163360595703, 0.6658535003662109, 3.6286239624023438, -1.1459503173828125, -0.6040496826171875, -0.14559173583984375, 5.834197998046875, 3.743438720703125, -0.8623275756835938, 1.6867904663085938, 1.461141586303711, 3.7215805053710938, 3.4227828979492188, 5.014457702636719, -0.5836296081542969, -0.5804405212402344, 1.4763374328613281], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000113.npy"}
{"epoch": 0.1708238851095994, "step": 114, "batch_size": 64, "mean": 1.3459113836288452, "std": 3.1439263820648193, "min": -4.2093658447265625, "p10": -2.3454772949218747, "median": 1.0628223419189453, "p90": 5.194883728027344, "max": 10.39031982421875, "pos_frac": 0.640625, "sample": [-1.6239776611328125, -1.7529296875, 2.3123016357421875, 2.4864425659179688, 8.353034973144531, 2.0828094482421875, 4.892982482910156, 5.5004730224609375, 1.4271926879882812, -3.8335952758789062, 4.967292785644531, -1.7443695068359375, 1.9617500305175781, -2.428680419921875, 0.7197036743164062, 1.2508926391601562, 1.4199371337890625, 0.41933441162109375, -0.4893836975097656, -0.6460647583007812, 0.625335693359375, -0.48931121826171875, 10.39031982421875, -1.7529411315917969, 1.0357780456542969, 3.8816070556640625, 4.364646911621094, 2.6198596954345703, -0.45243072509765625, 3.201934814453125, -2.7414817810058594, 0.7733840942382812, 4.7948760986328125, -3.741455078125, 2.2609481811523438, -0.3476448059082031, 5.1956939697265625, 0.9210662841796875, 2.4850425720214844, -2.851367950439453, 8.515045166015625, 1.9185676574707031, 5.1929931640625, 0.09370994567871094, 1.762298583984375, -1.0457649230957031, 8.206161499023438, -0.08013916015625, -1.3985595703125, -2.151336669921875, 1.0898666381835938, 0.751739501953125, 1.3305015563964844, 0.04559326171875, 3.1340408325195312, 4.931396484375, 1.8050537109375, -1.380441665649414, -2.6525802612304688, -4.2093658447265625, 5.8335113525390625, -0.54339599609375, 1.1639213562011719, -1.6234970092773438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000114.npy"}
{"epoch": 0.17233560090702948, "step": 115, "batch_size": 64, "mean": 1.2582770586013794, "std": 2.6060709953308105, "min": -3.57574462890625, "p10": -2.02966194152832, "median": 1.008697509765625, "p90": 4.937807083129886, "max": 7.689544677734375, "pos_frac": 0.65625, "sample": [-1.9641304016113281, -0.5098304748535156, 3.884521484375, 0.0985870361328125, 7.5987091064453125, -0.5365791320800781, 1.2755203247070312, 4.122127532958984, -0.4109649658203125, 6.391632080078125, 0.6110267639160156, -2.6547927856445312, 7.388084411621094, -1.3476486206054688, -2.1613082885742188, -3.4794464111328125, 0.7633209228515625, 2.116729736328125, 0.4634246826171875, 3.614776611328125, 1.4989395141601562, 3.4636077880859375, 1.4055557250976562, 0.9444198608398438, -2.0577468872070312, 0.9873504638671875, 1.6978530883789062, 1.646209716796875, -3.57574462890625, 5.759895324707031, -0.397918701171875, 3.9534378051757812, 2.2038116455078125, 5.8459930419921875, 0.00247955322265625, 3.83489990234375, 0.8473739624023438, 1.11737060546875, -0.979034423828125, -0.2989349365234375, 1.2492141723632812, -2.2477951049804688, 2.0908355712890625, 2.0382080078125, 3.6369171142578125, 1.2439002990722656, 3.8173065185546875, 0.041294097900390625, -2.3554420471191406, 7.689544677734375, 1.0629730224609375, -0.6231937408447266, 2.5018062591552734, -0.09044265747070312, -0.03119659423828125, -0.23895263671875, -1.4659042358398438, -0.13888931274414062, 1.0300445556640625, -0.45025634765625, 1.8725318908691406, 0.09183120727539062, 1.3544349670410156, 5.287384033203125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000115.npy"}
{"epoch": 0.17384731670445955, "step": 116, "batch_size": 64, "mean": 1.6396100521087646, "std": 2.685100793838501, "min": -4.768516540527344, "p10": -1.1589344024658204, "median": 1.209869384765625, "p90": 5.789421844482423, "max": 8.132965087890625, "pos_frac": 0.703125, "sample": [7.107933044433594, -1.1700820922851562, 0.50115966796875, 5.8265533447265625, 5.702781677246094, 1.4634933471679688, 4.72833251953125, 0.047298431396484375, 3.4341697692871094, 4.171577453613281, 2.5273590087890625, -0.9784164428710938, -1.1240997314453125, 4.82696533203125, 1.3858108520507812, 0.5978488922119141, -0.4344940185546875, 0.06592559814453125, 2.2373199462890625, 3.4585342407226562, -1.8739585876464844, 0.3606109619140625, 2.4148788452148438, -0.0682220458984375, -1.1329231262207031, 6.220497131347656, 0.9791793823242188, 2.9455108642578125, 7.576385498046875, 1.2311630249023438, 1.6529922485351562, -0.9765434265136719, -2.318521499633789, 0.42038536071777344, 4.6393890380859375, 0.246185302734375, 3.0934524536132812, 2.957061767578125, -1.0249176025390625, -1.49615478515625, 2.083038330078125, 1.7308120727539062, 1.2067489624023438, -1.4284439086914062, 6.4933624267578125, 2.1562843322753906, 5.8523712158203125, -1.633056640625, -0.39043426513671875, 0.7797622680664062, -0.94696044921875, -0.1113128662109375, 0.5598869323730469, -4.768516540527344, -0.24879074096679688, 2.2080535888671875, 1.2129898071289062, 0.6887588500976562, 0.6100711822509766, 8.132965087890625, 3.7112960815429688, 2.6157569885253906, 5.038032531738281, -0.840057373046875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000116.npy"}
{"epoch": 0.17535903250188964, "step": 117, "batch_size": 64, "mean": 1.58555006980896, "std": 2.8660714626312256, "min": -4.825714111328125, "p10": -2.356103134155273, "median": 1.4262542724609375, "p90": 5.752631378173828, "max": 6.906414031982422, "pos_frac": 0.734375, "sample": [-2.4241485595703125, 5.9868011474609375, 3.11907958984375, 0.6918869018554688, 6.2183685302734375, 0.7890167236328125, 3.6885986328125, 0.9397411346435547, 3.1857986450195312, 2.0877685546875, 0.7969932556152344, -0.04315948486328125, 5.755828857421875, -4.825714111328125, 1.2081298828125, 2.4600296020507812, -2.5511322021484375, -0.8271484375, 3.8653640747070312, 3.1914901733398438, 0.27156829833984375, 4.024059295654297, 1.4170455932617188, 0.53387451171875, 4.294878005981445, -2.0806121826171875, 0.84808349609375, 2.951568603515625, 2.856121063232422, 6.049915313720703, 1.585296630859375, 5.745170593261719, 6.654735565185547, 0.36862754821777344, 3.575408935546875, -0.2641639709472656, 1.4209671020507812, 5.507209777832031, -0.2635498046875, 3.8366012573242188, 1.1021881103515625, 1.5372314453125, 3.1632232666015625, 6.906414031982422, -2.1973304748535156, -3.8044204711914062, 1.6129570007324219, 0.6684226989746094, 4.306938171386719, 2.140472412109375, 0.7448921203613281, 1.4315414428710938, 4.126655578613281, -2.489887237548828, 6.845245361328125, -3.5394020080566406, -1.9157695770263672, 2.9082412719726562, -4.44525146484375, -1.7293167114257812, 2.1243515014648438, 1.033935546875, -1.1400146484375, -0.5625114440917969], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000117.npy"}
{"epoch": 0.17687074829931973, "step": 118, "batch_size": 64, "mean": 1.2571250200271606, "std": 2.8933498859405518, "min": -6.461418151855469, "p10": -2.744528198242187, "median": 1.0069828033447266, "p90": 4.929497718811035, "max": 7.6602630615234375, "pos_frac": 0.6875, "sample": [-0.26195526123046875, 0.932098388671875, -0.009033203125, 1.3757781982421875, -6.461418151855469, -2.8940200805664062, 0.6191329956054688, 3.107250213623047, 5.138797760009766, -1.5477619171142578, 4.012720108032227, 2.9121475219726562, -0.7284698486328125, 3.4670372009277344, 0.6156044006347656, -2.3544769287109375, 2.628814697265625, 4.1045379638671875, 0.00644683837890625, 0.10647964477539062, -1.3427085876464844, 1.0805816650390625, 3.1880836486816406, 3.14434814453125, 2.9574737548828125, -3.1022796630859375, -2.3957138061523438, -1.5392684936523438, 2.9624786376953125, -0.038818359375, 2.5064620971679688, 2.1005420684814453, 6.1013031005859375, -2.1323375701904297, 0.550262451171875, 2.9186553955078125, 0.9333839416503906, 1.9077911376953125, 4.2787628173828125, -3.2415084838867188, 0.3964805603027344, 4.446830749511719, 1.5671768188476562, 7.6602630615234375, 0.2326812744140625, -2.9871444702148438, -3.8956375122070312, 4.969383239746094, 5.539440155029297, 5.48834228515625, 4.795143127441406, -4.090213775634766, 3.4888153076171875, 2.564311981201172, 0.3992919921875, -0.3749103546142578, -1.8456573486328125, 2.5376338958740234, 3.00054931640625, 4.836431503295898, 5.7599334716796875, -0.3706932067871094, 0.5743350982666016, 0.15601348876953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000118.npy"}
{"epoch": 0.17838246409674982, "step": 119, "batch_size": 64, "mean": 1.3585854768753052, "std": 3.595616579055786, "min": -7.8508148193359375, "p10": -2.3931060791015626, "median": 1.035400390625, "p90": 6.830061340332032, "max": 11.485382080078125, "pos_frac": 0.578125, "sample": [-1.9416465759277344, 7.855354309082031, -1.4693450927734375, -2.7133522033691406, -5.4279327392578125, -0.6942253112792969, 2.546722412109375, 1.1347885131835938, -1.0883026123046875, -0.8300399780273438, -2.9480018615722656, -0.91656494140625, -0.6423416137695312, 7.08184814453125, 4.68718147277832, -7.8508148193359375, -1.6843147277832031, -2.8837966918945312, 1.7854194641113281, 4.1619415283203125, -1.7878990173339844, 2.254117965698242, 0.37615966796875, 1.084564208984375, 7.285285949707031, 2.1289215087890625, -3.4058074951171875, 4.622344970703125, 6.84075927734375, 6.8050994873046875, 4.7218780517578125, 2.699565887451172, 2.169384002685547, 3.3328704833984375, 5.026947021484375, 11.485382080078125, 1.3166694641113281, -1.86090087890625, 0.8584403991699219, 5.90301513671875, 7.595920562744141, 1.596923828125, -1.9045677185058594, 1.7257232666015625, -2.3697357177734375, -2.4031219482421875, 5.9678192138671875, 3.5450820922851562, 2.898590087890625, -1.6223602294921875, 0.6519775390625, -2.097381591796875, -0.01300048828125, -1.6325263977050781, -0.0605010986328125, 0.986236572265625, 3.307659149169922, 7.3131866455078125, 0.9578018188476562, 2.2190628051757812, -0.752471923828125, 1.4459304809570312, -0.10728836059570312, -0.3188629150390625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000119.npy"}
{"epoch": 0.17989417989417988, "step": 120, "batch_size": 64, "mean": 1.971259355545044, "std": 2.9352805614471436, "min": -4.38763427734375, "p10": -0.8684968948364256, "median": 1.628732681274414, "p90": 6.088597106933595, "max": 10.04794692993164, "pos_frac": 0.75, "sample": [2.61676025390625, 0.03186798095703125, -0.4257087707519531, 10.04794692993164, 2.119659423828125, -0.298675537109375, 4.236415863037109, -0.5287818908691406, 1.351593017578125, -4.38763427734375, 7.375457763671875, 0.6561164855957031, 3.3857345581054688, 0.39910888671875, 2.72235107421875, 6.2185516357421875, 4.540153503417969, -0.4364509582519531, 3.0268478393554688, 1.0015029907226562, 1.9601974487304688, 0.2744903564453125, 2.5482635498046875, 3.4250869750976562, -1.4828643798828125, 1.4038734436035156, 6.612236022949219, -1.875213623046875, 2.857391357421875, 4.1995391845703125, 5.785369873046875, -4.171661376953125, 8.924652099609375, -0.9602184295654297, 4.230739593505859, -0.5365180969238281, 3.6266708374023438, -0.38871192932128906, 5.0552215576171875, 6.6783447265625, 0.6054859161376953, 1.4888725280761719, 0.57647705078125, 1.9908905029296875, 1.5123825073242188, -2.9816360473632812, -0.5591621398925781, 2.1124267578125, 2.9506454467773438, 0.928253173828125, 0.19566726684570312, 0.9544448852539062, 4.99298095703125, -0.138946533203125, 2.6217880249023438, -0.65447998046875, 7.0048675537109375, 4.629425048828125, 0.10249710083007812, -2.968505859375, 1.826416015625, 0.91650390625, 4.488521575927734, 1.7450828552246094], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000120.npy"}
{"epoch": 0.18140589569160998, "step": 121, "batch_size": 64, "mean": 1.4688822031021118, "std": 3.2715003490448, "min": -4.1588134765625, "p10": -2.3399314880371094, "median": 1.1727046966552734, "p90": 6.338874053955084, "max": 10.617691040039062, "pos_frac": 0.640625, "sample": [3.3331451416015625, -4.1588134765625, -0.444488525390625, -2.724651336669922, 3.414205551147461, 7.897705078125, 2.9120864868164062, -3.705656051635742, 1.8878040313720703, 4.980674743652344, -0.7975540161132812, -1.6775360107421875, 10.617691040039062, 3.0382633209228516, -1.9892120361328125, 0.5327186584472656, 3.57843017578125, 1.2310562133789062, 4.170280456542969, -2.416046142578125, 7.8732452392578125, 0.0955352783203125, 4.084663391113281, 4.039100646972656, 3.5618515014648438, -1.9681243896484375, 2.1813926696777344, -2.6499900817871094, 1.578582763671875, -0.36861419677734375, -0.3325538635253906, 7.90985107421875, -0.8917007446289062, 1.0978851318359375, 2.446533203125, 0.8280105590820312, -2.1623306274414062, 6.92095947265625, 2.1835098266601562, -0.6782989501953125, -0.17180442810058594, 0.5596389770507812, -0.7616424560546875, 3.6507644653320312, -2.0128097534179688, 2.2563247680664062, -4.0394287109375, 9.019546508789062, 0.873809814453125, 0.7868137359619141, 1.8113670349121094, 1.1143531799316406, 3.2621116638183594, 1.8329696655273438, -1.8226394653320312, 7.238525390625, -1.8524894714355469, 1.3856544494628906, 1.0956382751464844, 1.284576416015625, -0.27895355224609375, 4.7380218505859375, 1.7243194580078125, -3.115823745727539], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000121.npy"}
{"epoch": 0.18291761148904007, "step": 122, "batch_size": 64, "mean": 1.4933542013168335, "std": 3.6356201171875, "min": -5.1476593017578125, "p10": -2.8016117095947264, "median": 0.9376964569091797, "p90": 5.897290420532228, "max": 10.9666748046875, "pos_frac": 0.65625, "sample": [10.378982543945312, -4.566005706787109, -3.8470458984375, 2.15692138671875, 0.86322021484375, -1.3631629943847656, 10.35546875, 4.2403411865234375, -3.559051513671875, 4.9033355712890625, 2.558521270751953, -1.3164825439453125, -2.798137664794922, 1.0121726989746094, -0.11417388916015625, 10.7501220703125, -1.3168487548828125, 0.3930206298828125, -3.64923095703125, 3.9370384216308594, 4.50996208190918, 5.378507614135742, 0.6058921813964844, 1.928964614868164, 5.969123840332031, 7.7021026611328125, -2.3788490295410156, -1.2119789123535156, 2.7141799926757812, 5.729679107666016, -3.140380859375, 4.971343994140625, 1.095266342163086, 1.083953857421875, -5.1476593017578125, 1.5712432861328125, 1.0612926483154297, 0.28966522216796875, 1.910654067993164, -0.8599967956542969, 0.38745880126953125, 0.275726318359375, -1.8384857177734375, 6.8976898193359375, -2.8031005859375, 4.315910339355469, 0.84344482421875, 0.46942901611328125, -1.1300506591796875, 1.6120147705078125, -0.3205375671386719, -1.3110275268554688, 2.5189208984375, 1.0648574829101562, -0.13840103149414062, 0.11631965637207031, 1.175506591796875, -0.49430084228515625, -0.20359039306640625, 3.9197158813476562, 2.9723587036132812, 2.619792938232422, 0.8563690185546875, 10.9666748046875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000122.npy"}
{"epoch": 0.18442932728647016, "step": 123, "batch_size": 64, "mean": 1.6300681829452515, "std": 3.1398682594299316, "min": -5.22528076171875, "p10": -2.296855926513671, "median": 1.4682111740112305, "p90": 5.574101257324221, "max": 12.495559692382812, "pos_frac": 0.71875, "sample": [3.588275909423828, 6.584781646728516, 0.9992218017578125, 12.495559692382812, -1.4511222839355469, -2.9131813049316406, 3.374755859375, 2.0176620483398438, 4.381500244140625, 1.8265609741210938, 0.09647369384765625, 4.12078857421875, 2.161663055419922, -1.39501953125, 2.929931640625, 2.6812820434570312, 2.7555999755859375, 4.567483901977539, 0.134246826171875, 5.046783447265625, 1.269683837890625, 2.4002609252929688, 2.2929153442382812, 3.1268539428710938, 3.9085006713867188, 5.8000946044921875, 1.3740882873535156, -1.3555984497070312, 2.5754547119140625, 4.049274444580078, 1.5623340606689453, 3.4995040893554688, -5.22528076171875, 5.95402717590332, -0.9706878662109375, -4.75054931640625, -2.98394775390625, -2.659313201904297, -0.2716407775878906, 0.9697761535644531, 3.3467483520507812, -0.12241363525390625, 3.2421512603759766, -2.8088455200195312, 2.7228851318359375, 7.324317932128906, -1.1113452911376953, -2.8386459350585938, 0.27902984619140625, 0.19965744018554688, -1.416778564453125, 1.2152328491210938, 1.03167724609375, 0.8404769897460938, -1.159027099609375, 7.434906005859375, -1.440948486328125, 0.6391372680664062, -1.2188186645507812, 1.0630111694335938, 1.854548454284668, 2.3343048095703125, 7.2498779296875, 1.0942230224609375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000123.npy"}
{"epoch": 0.18594104308390022, "step": 124, "batch_size": 64, "mean": 1.9508872032165527, "std": 3.770692825317383, "min": -6.5952911376953125, "p10": -2.6334531784057615, "median": 1.8933486938476562, "p90": 6.783976745605469, "max": 12.922683715820312, "pos_frac": 0.71875, "sample": [-0.25109100341796875, 9.307708740234375, 1.36920166015625, -4.617580413818359, 1.8947296142578125, 2.722564697265625, 6.785491943359375, 2.7699737548828125, 9.963272094726562, 12.922683715820312, 2.193359375, 3.1839828491210938, -0.24572372436523438, 8.2781982421875, 1.2162675857543945, 3.0796051025390625, -1.2175979614257812, 2.0922622680664062, -0.03682708740234375, -1.960439682006836, -2.791839599609375, 2.972869873046875, -1.59503173828125, 3.7683258056640625, 0.6432952880859375, 3.1753387451171875, 0.24046897888183594, 3.4919357299804688, -1.047332763671875, -3.869415283203125, 5.846111297607422, 5.252044677734375, 2.7886962890625, 1.286468505859375, 6.1558685302734375, 8.193389892578125, 6.7804412841796875, -6.5952911376953125, 1.521453857421875, 4.490386962890625, -2.7030715942382812, 3.3608779907226562, 0.67266845703125, 2.34454345703125, 1.9410686492919922, 3.0440597534179688, 1.2755889892578125, 1.7545700073242188, -2.471010208129883, 1.8919677734375, 9.390495300292969, 0.6501007080078125, -0.2519512176513672, 0.2623443603515625, -3.3774871826171875, 1.1048965454101562, 3.0673294067382812, -1.1420974731445312, -5.820640563964844, 5.4605560302734375, -1.8315887451171875, 3.574390411376953, 0.5128021240234375, 1.9881439208984375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000124.npy"}
{"epoch": 0.1874527588813303, "step": 125, "batch_size": 64, "mean": 2.074455738067627, "std": 3.6205034255981445, "min": -7.869659423828125, "p10": -2.7113361358642574, "median": 1.8382186889648438, "p90": 5.945433044433594, "max": 11.6954345703125, "pos_frac": 0.78125, "sample": [-1.24005126953125, 3.4215621948242188, 5.607215881347656, 1.2240753173828125, 2.65911865234375, -0.9513130187988281, 4.407562255859375, 9.811809539794922, 6.343658447265625, 1.3895893096923828, 9.965179443359375, -2.3152389526367188, 0.3315010070800781, 1.37640380859375, 0.15320587158203125, 5.775112152099609, 4.85809326171875, 1.251556396484375, 0.3262214660644531, -1.4978446960449219, 5.657512664794922, 5.948234558105469, 1.833038330078125, 2.8221893310546875, -0.77362060546875, 2.571929931640625, 0.18332672119140625, -1.8705940246582031, -3.6293258666992188, 0.4841461181640625, 3.6837844848632812, 1.5227813720703125, 0.9668502807617188, 5.537017822265625, 5.498325347900391, -3.94110107421875, 2.088777542114258, 1.611968994140625, 5.938896179199219, 11.6954345703125, 4.1800537109375, 1.8433990478515625, 7.413185119628906, 4.2628173828125, 5.036552429199219, -2.9038467407226562, 1.9161453247070312, 3.896259307861328, 0.3436546325683594, 1.0181808471679688, 3.9763641357421875, 1.377279281616211, 3.5701446533203125, 1.33551025390625, -1.946502685546875, 6.16143798828125, 2.8611602783203125, -3.2460098266601562, 2.6400299072265625, 0.5409088134765625, 3.7555694580078125, -2.881092071533203, -5.243377685546875, -7.869659423828125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000125.npy"}
{"epoch": 0.1889644746787604, "step": 126, "batch_size": 64, "mean": 1.667033314704895, "std": 4.308374404907227, "min": -6.6609344482421875, "p10": -3.214411163330078, "median": 0.6038579940795898, "p90": 8.196736907958986, "max": 11.62451171875, "pos_frac": 0.609375, "sample": [9.605937957763672, -0.8315715789794922, 4.203100204467773, 1.2863922119140625, 1.1244125366210938, 7.7201385498046875, 6.2731781005859375, 0.6074352264404297, 1.4434967041015625, -1.6358413696289062, 8.710979461669922, 6.9045257568359375, -6.6609344482421875, 0.60028076171875, 11.165428161621094, 3.0099105834960938, -3.1121978759765625, -1.3613052368164062, 0.08036994934082031, -2.962352752685547, 6.160308837890625, 6.1077117919921875, -1.0182666778564453, 2.467344284057617, -3.2582168579101562, 7.006645202636719, 3.5486297607421875, -3.5655288696289062, 10.558609008789062, -0.415679931640625, -3.5168228149414062, -4.66815185546875, -2.5628204345703125, 0.25504302978515625, -2.223560333251953, 1.3204421997070312, -0.06457138061523438, 2.1846466064453125, 0.5501832962036133, -6.0106201171875, 5.2007293701171875, -1.6693267822265625, 3.428272247314453, 0.004180908203125, 10.023597717285156, -0.55609130859375, 1.9796676635742188, 11.62451171875, 1.6145706176757812, 2.6196365356445312, -1.3139114379882812, 2.3360671997070312, 2.8786888122558594, 0.579315185546875, -0.0564117431640625, -0.8057327270507812, 0.5971527099609375, 7.0156402587890625, -1.9184722900390625, -2.5994338989257812, -0.37432861328125, 8.400993347167969, -3.767223358154297, 2.42132568359375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000126.npy"}
{"epoch": 0.19047619047619047, "step": 127, "batch_size": 64, "mean": 2.2443761825561523, "std": 3.2799723148345947, "min": -5.6021881103515625, "p10": -1.4646903991699218, "median": 2.087646484375, "p90": 6.469207000732422, "max": 11.396827697753906, "pos_frac": 0.78125, "sample": [5.7148895263671875, 9.051422119140625, 2.7960586547851562, 5.173900604248047, -1.3025360107421875, 7.2354278564453125, -1.2376556396484375, 3.6499900817871094, 1.1493377685546875, 4.702705383300781, 7.251129150390625, 1.4560394287109375, 0.8933563232421875, 6.00537109375, 0.306610107421875, 0.4231452941894531, 4.166877746582031, 1.8402748107910156, 0.3675422668457031, 1.9907913208007812, 6.493095397949219, -5.6021881103515625, 2.4299850463867188, 2.956104278564453, 2.441864013671875, -4.102783203125, 0.3680572509765625, 0.4847602844238281, 2.0297393798828125, 0.338470458984375, -1.5601234436035156, 4.242847442626953, 4.589208602905273, 2.1455535888671875, -0.322052001953125, 4.696659088134766, 1.0115432739257812, 6.4134674072265625, 3.9616661071777344, 11.396827697753906, -4.554210662841797, 4.3202667236328125, 3.6267662048339844, -1.4386978149414062, -2.8971099853515625, 6.5023956298828125, 1.3496017456054688, -0.8118438720703125, 6.160491943359375, -0.98406982421875, -1.475830078125, 4.09619140625, 2.7862625122070312, 0.21308517456054688, 2.9361953735351562, -1.6724815368652344, 3.7769241333007812, 0.4939727783203125, 2.651763916015625, 1.8719902038574219, 2.210195541381836, 0.8491687774658203, 8.08285140991211, -0.5011844635009766], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000127.npy"}
{"epoch": 0.19198790627362056, "step": 128, "batch_size": 64, "mean": 1.4706671237945557, "std": 4.742069244384766, "min": -10.261245727539062, "p10": -5.125218200683594, "median": 1.5137176513671875, "p90": 7.464718627929687, "max": 12.352508544921875, "pos_frac": 0.640625, "sample": [3.40283203125, -2.8609619140625, 1.8182754516601562, 1.3102188110351562, 0.9637603759765625, 1.7689247131347656, -0.7353668212890625, 8.59320068359375, -0.08049774169921875, 8.204391479492188, 0.5130844116210938, -10.261245727539062, 2.7190017700195312, 0.2606353759765625, 7.151399612426758, -7.141578674316406, 1.4490509033203125, 7.441497802734375, -5.814552307128906, 9.688117980957031, 12.352508544921875, -1.2825641632080078, 2.490131378173828, 7.47467041015625, 11.458038330078125, 0.752960205078125, -1.638641357421875, 7.350303649902344, -5.3664093017578125, 3.4369659423828125, 3.283050537109375, 0.1566009521484375, -0.8776054382324219, 1.857696533203125, 3.8806381225585938, 1.5783843994140625, -1.9445610046386719, 4.504055023193359, -8.355396270751953, 4.470508575439453, -2.061237335205078, 1.974355697631836, 1.7189178466796875, 0.19735336303710938, -4.56243896484375, 3.5949249267578125, 1.7117443084716797, -0.029781341552734375, 2.8607330322265625, -0.22146320343017578, -7.558380126953125, 5.72027587890625, 4.05023193359375, -0.0444793701171875, 2.2387657165527344, -1.046438217163086, 1.1369171142578125, 3.505207061767578, 11.810096740722656, -1.24652099609375, -7.929420471191406, -0.749786376953125, 5.321540832519531, -0.2399444580078125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000128.npy"}
{"epoch": 0.19349962207105065, "step": 129, "batch_size": 64, "mean": 1.6215569972991943, "std": 4.666401386260986, "min": -9.0047607421875, "p10": -3.824726867675781, "median": 1.6556625366210938, "p90": 7.52489471435547, "max": 13.6529541015625, "pos_frac": 0.625, "sample": [-1.7901077270507812, -2.3668365478515625, 5.443902969360352, 0.985931396484375, -1.053171157836914, 8.201374053955078, 4.2070159912109375, -4.416419982910156, 4.961250305175781, 8.044010162353516, 6.749446868896484, 4.952484130859375, 1.67138671875, -0.16742706298828125, 0.11946868896484375, 1.2641143798828125, 7.598114013671875, -2.2304630279541016, -0.4949493408203125, 5.156848907470703, 6.0746612548828125, -9.0047607421875, 1.4843978881835938, 4.832878112792969, -1.8053665161132812, 13.6529541015625, 0.12949371337890625, -2.5648651123046875, 2.3815250396728516, 8.8323974609375, 1.7711334228515625, -0.876495361328125, -8.560073852539062, 4.618749618530273, 3.3645858764648438, 5.385799407958984, -6.062919616699219, -3.7047348022460938, -2.4105911254882812, 7.298641204833984, 2.2977981567382812, -1.8939628601074219, -2.1216888427734375, 0.9383945465087891, 7.3540496826171875, 3.3459110260009766, -3.8761520385742188, 7.0865936279296875, 2.0408477783203125, 11.248954772949219, 0.9296875, 4.868431091308594, 7.9640350341796875, -0.15894317626953125, 2.1200180053710938, -5.304353713989258, -2.5879287719726562, -7.419136047363281, 2.12628173828125, -1.395517349243164, 1.9470100402832031, 1.6399383544921875, -2.7134647369384766, 3.669464111328125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000129.npy"}
{"epoch": 0.19501133786848074, "step": 130, "batch_size": 64, "mean": 2.0028371810913086, "std": 4.052952766418457, "min": -8.529281616210938, "p10": -2.7823623657226557, "median": 1.466780662536621, "p90": 7.1230300903320325, "max": 10.845001220703125, "pos_frac": 0.6875, "sample": [2.549591064453125, 2.8541946411132812, 5.8087921142578125, 10.599601745605469, 1.13250732421875, -2.071807861328125, 1.48455810546875, 3.6536808013916016, -1.8273906707763672, 0.9903221130371094, -1.4196624755859375, 6.153564453125, 6.157432556152344, 1.4490032196044922, -3.1745452880859375, 8.058990478515625, -1.0685577392578125, -3.5303497314453125, -2.2885513305664062, 10.71734619140625, 2.145915985107422, -1.1121292114257812, -0.56768798828125, 6.6088409423828125, -0.2486724853515625, -4.129005432128906, 0.8631362915039062, 0.2255706787109375, 1.5081748962402344, -2.0228347778320312, 4.7649993896484375, 2.9799728393554688, 9.828800201416016, 5.948890686035156, 2.3445205688476562, 7.294158935546875, 7.2722930908203125, -1.2797775268554688, 10.845001220703125, 2.5548095703125, -1.1696701049804688, 2.8646697998046875, 1.2477035522460938, -2.9939956665039062, -0.9964523315429688, 2.381277084350586, 3.016948699951172, 1.316650390625, 0.24212646484375, -4.317829132080078, -5.359687805175781, 0.0937957763671875, 3.532196044921875, 5.86192512512207, -8.529281616210938, 5.9051971435546875, 5.929779052734375, -0.15849876403808594, 4.55908203125, 0.536285400390625, 3.803417205810547, 0.17302703857421875, 1.414468765258789, 6.774749755859375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000130.npy"}
{"epoch": 0.1965230536659108, "step": 131, "batch_size": 64, "mean": 2.5501041412353516, "std": 4.615158557891846, "min": -8.464622497558594, "p10": -2.7781002044677736, "median": 1.6845779418945312, "p90": 9.267514038085942, "max": 14.060760498046875, "pos_frac": 0.703125, "sample": [0.9132156372070312, 4.284971237182617, -2.7824134826660156, 0.5530548095703125, -3.3196563720703125, 14.060760498046875, 2.239002227783203, 3.192218780517578, -0.42943572998046875, 11.862167358398438, 3.5148773193359375, 0.8129768371582031, -0.9930534362792969, 0.426513671875, 5.80615234375, 3.6973800659179688, 6.7219390869140625, -1.5682373046875, 2.4825057983398438, -8.464622497558594, -2.4275970458984375, 1.715545654296875, 13.10150146484375, 0.6757888793945312, 1.6536102294921875, -3.41888427734375, -1.3812904357910156, -1.4180221557617188, -4.777248382568359, -2.768035888671875, 2.4439430236816406, 0.825897216796875, 5.965545654296875, 1.4696941375732422, 8.139396667480469, 7.8408660888671875, 1.5124130249023438, 8.372283935546875, 1.624216079711914, 5.4157562255859375, 3.0512237548828125, 9.65118408203125, -2.138193130493164, 7.979736328125, 5.087039947509766, -3.1826705932617188, 1.4413642883300781, -4.462547302246094, -0.29084014892578125, 9.868743896484375, 1.258687973022461, -0.4800872802734375, 7.347620010375977, -2.1535072326660156, -2.20025634765625, 9.89398193359375, 2.3503007888793945, 1.4205360412597656, 4.373565673828125, 5.026721954345703, 2.665283203125, 3.697784423828125, 10.26198959350586, 5.163318634033203], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000131.npy"}
{"epoch": 0.1980347694633409, "step": 132, "batch_size": 64, "mean": 1.7558989524841309, "std": 3.92497181892395, "min": -7.576530456542969, "p10": -2.218587493896484, "median": 1.8247013092041016, "p90": 5.989785385131837, "max": 12.79840087890625, "pos_frac": 0.75, "sample": [3.716327667236328, 0.6645421981811523, 3.4678726196289062, 3.5017166137695312, -1.1893024444580078, -1.9270858764648438, 6.37420654296875, 1.5958127975463867, -6.868095397949219, 7.140327453613281, 2.9685020446777344, 2.1308212280273438, 2.363872528076172, 2.6201820373535156, -1.606781005859375, 2.2892227172851562, -0.6319541931152344, 8.021575927734375, 3.8654327392578125, 3.170543670654297, -7.576530456542969, -7.416107177734375, 6.120693206787109, -1.5158462524414062, -0.8161354064941406, -2.0409011840820312, 4.587371826171875, 1.3362274169921875, 5.684333801269531, -1.6078147888183594, 10.420295715332031, 0.6099424362182617, -7.496925354003906, 4.406074523925781, 3.49774169921875, 3.065185546875, 4.721824645996094, 3.5606822967529297, 5.14268684387207, 1.88720703125, 0.9678573608398438, 0.625091552734375, 0.43927001953125, 11.3221435546875, 1.8320350646972656, 0.19260406494140625, 3.6749839782714844, 4.214668273925781, 0.7494449615478516, 1.622650146484375, 3.568531036376953, -0.0900726318359375, 0.401214599609375, 1.9850082397460938, 2.0459413528442383, 1.8173675537109375, 12.79840087890625, -2.6338157653808594, 1.2059783935546875, 0.7351837158203125, 1.1316909790039062, 1.6944427490234375, -3.8661041259765625, -2.29473876953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000132.npy"}
{"epoch": 0.19954648526077098, "step": 133, "batch_size": 64, "mean": 2.0136892795562744, "std": 4.3267645835876465, "min": -6.781517028808594, "p10": -3.841481781005859, "median": 1.7706289291381836, "p90": 7.333692169189457, "max": 13.603805541992188, "pos_frac": 0.640625, "sample": [8.024063110351562, -6.781517028808594, 1.1408233642578125, 4.628440856933594, -0.14914894104003906, -6.045799255371094, 1.9971504211425781, 11.236251831054688, -0.1455364227294922, 3.6305503845214844, -1.0356826782226562, 4.969757080078125, 5.355806350708008, 2.8546218872070312, -2.867778778076172, 2.7971343994140625, 7.86669921875, -6.11688232421875, 3.48919677734375, 4.7189178466796875, -4.16815185546875, 5.344306945800781, -1.3667831420898438, 5.166986465454102, 1.8340740203857422, -3.9160995483398438, 2.7172088623046875, -2.8165435791015625, 6.165351867675781, -0.18898487091064453, 5.105228424072266, 0.5127449035644531, 3.0579986572265625, 7.935310363769531, 0.3701324462890625, 0.5019378662109375, -4.761749267578125, 6.421623229980469, 5.8011016845703125, 7.724578857421875, -3.6673736572265625, -0.2476806640625, 0.4198760986328125, -1.3558616638183594, 1.1787605285644531, 9.996482849121094, 4.916664123535156, 4.934898376464844, -2.4746856689453125, 13.603805541992188, -0.6590805053710938, 1.2002449035644531, -4.025808334350586, 1.4852447509765625, 5.806571960449219, -0.17702102661132812, 3.5099639892578125, 6.348865509033203, 5.509185791015625, -0.7904052734375, 4.4635162353515625, -0.32685089111328125, -3.4877243041992188, 1.707183837890625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000133.npy"}
{"epoch": 0.20105820105820105, "step": 134, "batch_size": 64, "mean": 2.901494026184082, "std": 4.749834060668945, "min": -5.3199462890625, "p10": -3.73126220703125, "median": 2.757261276245117, "p90": 7.762834167480469, "max": 17.5504150390625, "pos_frac": 0.71875, "sample": [-4.352447509765625, 1.913787841796875, -1.2036094665527344, 4.828639984130859, 1.3860511779785156, 1.4935321807861328, 2.75433349609375, -1.0354690551757812, 4.191978454589844, 5.32452392578125, 3.3817672729492188, 4.356395721435547, 3.8235015869140625, 1.4830398559570312, -0.43491363525390625, 2.7601890563964844, 7.8524322509765625, 4.7506866455078125, -4.317615509033203, 6.359409332275391, 8.16424560546875, 10.355422973632812, 5.926731109619141, -4.529655456542969, 1.456939697265625, 6.3366241455078125, -1.0480575561523438, 3.3983078002929688, -0.12277889251708984, 2.8306427001953125, -3.795257568359375, -0.26306915283203125, 6.156982421875, -4.2601776123046875, 1.5057544708251953, -2.0405502319335938, 7.1397857666015625, 2.5925350189208984, 2.970001220703125, 4.856271743774414, 17.5504150390625, 5.246061325073242, -3.581939697265625, 4.5235595703125, -0.3536262512207031, 3.90447998046875, 7.27581787109375, -5.3199462890625, 11.938819885253906, 1.4998092651367188, 2.982746124267578, 1.7729110717773438, -0.6343231201171875, 2.091766357421875, 14.43182373046875, -4.943931579589844, 3.357696533203125, 2.5129470825195312, 1.6580162048339844, 3.9009628295898438, 7.55377197265625, 16.292579650878906, 0.0161590576171875, -0.9278688430786133], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000134.npy"}
{"epoch": 0.20256991685563114, "step": 135, "batch_size": 64, "mean": 2.007038116455078, "std": 4.245513916015625, "min": -11.668121337890625, "p10": -1.6020975112915037, "median": 2.04837703704834, "p90": 7.624228668212892, "max": 14.772270202636719, "pos_frac": 0.671875, "sample": [2.351184844970703, -0.8386344909667969, 0.8287582397460938, 4.212135314941406, 2.536632537841797, -0.11798858642578125, -4.7391357421875, -0.507598876953125, 2.617015838623047, 8.873138427734375, 0.060638427734375, -5.21484375, 3.7007369995117188, 0.2767353057861328, 2.6943740844726562, -5.968894958496094, 8.418083190917969, -1.0311431884765625, -1.2919502258300781, 3.6678829193115234, 2.28057861328125, -1.01678466796875, -11.668121337890625, 7.2068328857421875, 1.73773193359375, 0.7918777465820312, 4.146007537841797, 2.9364395141601562, 2.073211669921875, 4.845310211181641, 1.5868721008300781, -2.58013916015625, 6.559959411621094, 2.0648555755615234, 6.860992431640625, -0.5258598327636719, 0.619659423828125, -0.515350341796875, 4.730968475341797, 0.2723350524902344, -0.47840118408203125, 7.8326263427734375, -0.7880802154541016, 3.841949462890625, 3.9139270782470703, 7.454627990722656, 1.4566802978515625, -0.15351104736328125, 2.645832061767578, 1.9529266357421875, 2.0318984985351562, -6.177886962890625, -0.48242950439453125, 3.264444351196289, 7.6969146728515625, 10.060401916503906, -1.256490707397461, 2.208589553833008, -1.7350177764892578, 7.9809112548828125, 3.5972900390625, 6.58056640625, 14.772270202636719, -0.7041091918945312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000135.npy"}
{"epoch": 0.20408163265306123, "step": 136, "batch_size": 64, "mean": 2.4650299549102783, "std": 6.00989294052124, "min": -9.244285583496094, "p10": -5.1899169921875, "median": 0.5709304809570312, "p90": 9.629695892333986, "max": 17.4251708984375, "pos_frac": 0.609375, "sample": [5.98675537109375, 7.05573844909668, -2.574371337890625, -0.33818817138671875, -5.20635986328125, 5.896522521972656, -9.226211547851562, -2.4345550537109375, -8.799880981445312, 7.757551193237305, -0.5383377075195312, -5.482383728027344, -0.7776870727539062, 0.08531761169433594, -2.1141738891601562, -0.6419601440429688, 8.351245880126953, -0.5056648254394531, 0.18637847900390625, -1.1082115173339844, 9.814651489257812, 2.2025833129882812, 9.736343383789062, 9.380851745605469, -2.43017578125, 0.569549560546875, 11.743354797363281, -1.2096633911132812, 0.384124755859375, 14.75897216796875, 8.28643798828125, -0.15219879150390625, 6.03253173828125, 7.6422576904296875, 8.223556518554688, -4.02496337890625, 0.16827392578125, 6.017120361328125, 8.891151428222656, -5.15155029296875, -6.6071367263793945, 0.5723114013671875, 3.1872406005859375, 8.255813598632812, 3.100006103515625, 0.9244155883789062, -5.919853210449219, 4.460365295410156, 2.010589599609375, 17.4251708984375, 0.5451927185058594, -0.6577301025390625, 4.6781005859375, 9.14874267578125, 7.75982666015625, 16.989105224609375, -0.8180694580078125, 1.0833740234375, 6.068870544433594, 9.842269897460938, 0.0146942138671875, -1.5007705688476562, -0.0110626220703125, -9.244285583496094], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000136.npy"}
{"epoch": 0.20559334845049132, "step": 137, "batch_size": 64, "mean": 2.5589451789855957, "std": 4.649223804473877, "min": -6.2491912841796875, "p10": -3.2012615203857417, "median": 2.7170896530151367, "p90": 8.023923873901369, "max": 17.50506591796875, "pos_frac": 0.671875, "sample": [-4.133487701416016, -6.2491912841796875, 0.6954879760742188, 9.390663146972656, 3.8506698608398438, 5.739616394042969, 8.21976089477539, 13.783355712890625, 2.4554824829101562, 17.50506591796875, 4.465982437133789, 3.7671661376953125, 7.5669708251953125, -1.1717300415039062, 8.721221923828125, 1.8693313598632812, -3.7704849243164062, 4.769561767578125, -2.7111053466796875, -2.6572494506835938, 1.7307968139648438, 2.218059539794922, 5.781402587890625, -3.372791290283203, 2.978696823120117, 0.23928070068359375, 10.242530822753906, 10.399093627929688, 4.845298767089844, 5.101936340332031, 0.7314910888671875, 4.066734313964844, -5.148105621337891, 3.2821197509765625, 7.4023590087890625, 3.3869667053222656, -1.5542621612548828, -4.181354522705078, -0.12995147705078125, 0.6006736755371094, 1.5972785949707031, 1.6218681335449219, 0.6671829223632812, 4.867279052734375, 6.328071594238281, -0.7188148498535156, -1.8505630493164062, 5.374053955078125, 4.8447265625, 5.1251678466796875, -0.018384933471679688, -0.76446533203125, -0.5265970230102539, -4.9815521240234375, 3.408893585205078, -1.8875465393066406, -1.01531982421875, -2.801025390625, -2.5385513305664062, 3.57928466796875, 5.926605224609375, 5.760944366455078, 4.869232177734375, 6.1766510009765625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000137.npy"}
{"epoch": 0.20710506424792138, "step": 138, "batch_size": 64, "mean": 1.8820289373397827, "std": 5.135687351226807, "min": -8.242149353027344, "p10": -3.5949535369873042, "median": 1.1806440353393555, "p90": 8.047792053222656, "max": 19.1795654296875, "pos_frac": 0.640625, "sample": [0.4712791442871094, 3.1830978393554688, 0.7828330993652344, 2.839405059814453, 1.2856407165527344, 0.7652854919433594, 2.0377349853515625, -1.138275146484375, -6.1342010498046875, -3.6923370361328125, 1.3911056518554688, -6.3446807861328125, 7.932403564453125, 0.00392913818359375, 13.621864318847656, -7.3034820556640625, -1.4723129272460938, 4.4413299560546875, 9.573715209960938, 1.3057060241699219, 6.014549255371094, -2.3438873291015625, 19.1795654296875, 3.4102401733398438, 0.7569999694824219, 5.671236038208008, 5.46278190612793, -1.7374114990234375, 2.568798065185547, -0.4883842468261719, 8.097244262695312, -3.367725372314453, 5.0313720703125, -0.412628173828125, -0.4624462127685547, -1.854278564453125, 1.0156059265136719, -8.242149353027344, -0.3674182891845703, 7.083015441894531, 3.9673919677734375, 10.572685241699219, -2.0270233154296875, 4.486610412597656, -0.2798576354980469, -8.08651351928711, 4.472448348999023, 0.7168159484863281, -1.5774993896484375, 1.4245147705078125, -1.5748291015625, 4.92877197265625, 5.9885101318359375, 2.5646743774414062, 1.1935272216796875, -5.331083297729492, 0.006481170654296875, 1.1677608489990234, -0.5771560668945312, 3.0069198608398438, 10.207099914550781, 3.296337127685547, 13.82806396484375, -0.48992156982421875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000138.npy"}
{"epoch": 0.20861678004535147, "step": 139, "batch_size": 64, "mean": 1.9987211227416992, "std": 4.974520206451416, "min": -15.499847412109375, "p10": -4.315220642089843, "median": 2.891476631164551, "p90": 7.970032691955568, "max": 10.878044128417969, "pos_frac": 0.71875, "sample": [-4.7248687744140625, 8.169294357299805, 7.505088806152344, 1.3470382690429688, -5.573314666748047, 3.568817138671875, 3.405059814453125, 4.321949005126953, 9.612922668457031, -2.1253318786621094, -1.6134872436523438, 6.373451232910156, 3.1695404052734375, -5.155366897583008, -0.668548583984375, 1.170501708984375, 9.675979614257812, 2.8885879516601562, 2.9002304077148438, 2.5791072845458984, 0.4097137451171875, -15.499847412109375, 0.33301544189453125, 2.8943653106689453, 5.159721374511719, 5.122886657714844, 8.436904907226562, 1.7205333709716797, 4.548408508300781, -3.359375, 3.215057373046875, 2.9022598266601562, 3.4207916259765625, -8.30419921875, 4.045799255371094, 0.60272216796875, 0.8442955017089844, 1.6379127502441406, 10.604400634765625, -0.24290847778320312, 5.840019226074219, -3.2316856384277344, 5.098663330078125, 0.1954784393310547, 1.6515693664550781, -6.428619384765625, 5.461029052734375, 4.932868957519531, 7.2453765869140625, 7.4385528564453125, -2.194976806640625, 4.860996246337891, 0.9209671020507812, -8.899284362792969, -1.288726806640625, 4.244758605957031, -2.4875030517578125, 6.458740234375, -2.63253116607666, 8.592926025390625, 5.074188232421875, 10.878044128417969, 1.958043098449707, -1.0898590087890625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000139.npy"}
{"epoch": 0.21012849584278157, "step": 140, "batch_size": 64, "mean": 3.2170231342315674, "std": 5.601644039154053, "min": -10.918167114257812, "p10": -3.734488677978515, "median": 2.6887855529785156, "p90": 10.402757263183595, "max": 15.672935485839844, "pos_frac": 0.65625, "sample": [11.268325805664062, 10.458892822265625, -0.5400009155273438, -2.641632080078125, -2.7161331176757812, 8.066394805908203, -4.441902160644531, -4.836761474609375, 3.12933349609375, 1.6774520874023438, -10.918167114257812, 7.244773864746094, 4.1172027587890625, 3.3941650390625, 7.662933349609375, -2.910614013671875, -1.1857185363769531, 5.46002197265625, 5.286285400390625, 8.089637756347656, -0.4930229187011719, 8.64727783203125, 7.792499542236328, 7.576438903808594, 1.356475830078125, 10.026809692382812, -0.6585235595703125, 11.007244110107422, 0.9961318969726562, -2.5351181030273438, -4.418731689453125, 0.3689117431640625, 5.639434814453125, -4.823570251464844, -1.8210868835449219, 4.7567138671875, 2.4327049255371094, 9.974563598632812, 10.869224548339844, 11.272354125976562, 8.902595520019531, 8.737422943115234, 10.271774291992188, 9.080482482910156, -0.261260986328125, -1.0624752044677734, 14.360847473144531, -2.0451812744140625, 15.672935485839844, -2.3573875427246094, -4.138031005859375, 6.34051513671875, 0.5926895141601562, 2.4164600372314453, 8.662879943847656, 0.9853620529174805, 4.4933319091796875, -2.6679000854492188, 4.069305419921875, 1.6297149658203125, 2.654327392578125, 2.7232437133789062, -2.7158126831054688, -4.087577819824219], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000140.npy"}
{"epoch": 0.21164021164021163, "step": 141, "batch_size": 64, "mean": 3.0974936485290527, "std": 5.657039165496826, "min": -10.441390991210938, "p10": -4.064863586425781, "median": 3.377063751220703, "p90": 11.314088249206543, "max": 14.020111083984375, "pos_frac": 0.65625, "sample": [4.2279052734375, -5.599891662597656, 9.043991088867188, -0.7010726928710938, 7.7653961181640625, -0.2365875244140625, 4.106986999511719, 11.364448547363281, 3.6256866455078125, -1.7838211059570312, 0.322021484375, -3.5090866088867188, 4.0497283935546875, -6.058338165283203, 12.13067626953125, -0.18307876586914062, 3.116596221923828, 0.07891082763671875, -3.469799041748047, -10.441390991210938, 0.34447479248046875, -4.606395721435547, -2.2393646240234375, 4.815998077392578, -0.3574371337890625, 1.9310989379882812, 5.914680480957031, 9.584808349609375, 3.6917572021484375, -4.53131103515625, -0.32413482666015625, -4.191535949707031, 5.498584747314453, 11.655616760253906, 7.420036315917969, -3.7692947387695312, 8.962791442871094, -1.9105300903320312, 11.19658088684082, 14.020111083984375, 5.6375885009765625, 5.010824203491211, 3.855926513671875, 3.1284408569335938, 12.974498748779297, 4.020416259765625, -1.5094718933105469, 6.19781494140625, 2.1974563598632812, 9.931137084960938, -2.9914779663085938, 2.6818161010742188, 6.817207336425781, 13.153350830078125, 4.5805206298828125, 12.84869384765625, -3.1699771881103516, 8.853885650634766, 2.30108642578125, 8.659210205078125, 6.510734558105469, -0.7171421051025391, -6.3610382080078125, 2.672271728515625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000141.npy"}
{"epoch": 0.21315192743764172, "step": 142, "batch_size": 64, "mean": 2.2193617820739746, "std": 5.051835060119629, "min": -11.565814971923828, "p10": -2.488247776031494, "median": 1.8910818099975586, "p90": 6.535311508178711, "max": 18.065399169921875, "pos_frac": 0.703125, "sample": [5.1274566650390625, -8.646507263183594, 0.3932380676269531, 6.273235321044922, 3.5399551391601562, 12.802200317382812, 5.9604339599609375, -6.992015838623047, 2.6621780395507812, 1.0082473754882812, 3.170013427734375, 1.185882568359375, 6.480113983154297, -5.579261779785156, 0.6705818176269531, 2.1950759887695312, -1.43536376953125, -11.565814971923828, -2.569854736328125, 1.8904361724853516, 1.6457366943359375, 4.2751007080078125, 1.8917274475097656, -1.7218666076660156, 4.585941314697266, 10.319580078125, -2.06640625, 2.0110015869140625, 6.558967590332031, 1.6739692687988281, 1.2572784423828125, 2.75653076171875, 18.065399169921875, 3.885406494140625, -0.32988739013671875, -3.8918609619140625, -0.5144157409667969, -2.2978315353393555, 1.8809738159179688, 3.8341407775878906, 13.081802368164062, -0.24994659423828125, 0.3826141357421875, -1.8164749145507812, -1.3267173767089844, 0.943023681640625, 0.1368694305419922, 6.038482666015625, -1.65338134765625, 6.426002502441406, 2.7878494262695312, -2.260862350463867, -0.776031494140625, 14.128494262695312, 5.882286071777344, 6.1631622314453125, 4.904087066650391, 5.3840484619140625, 7.949432373046875, 3.6884403228759766, 3.0640029907226562, -4.0011444091796875, 0.43869781494140625, 2.334705352783203], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000142.npy"}
{"epoch": 0.2146636432350718, "step": 143, "batch_size": 64, "mean": 2.055307149887085, "std": 6.037508487701416, "min": -13.154647827148438, "p10": -4.10455322265625, "median": 1.211202621459961, "p90": 9.759379768371582, "max": 15.78448486328125, "pos_frac": 0.5625, "sample": [0.913848876953125, 6.6396484375, 8.768146514892578, 8.059562683105469, 2.2562408447265625, 4.223186492919922, -0.6099853515625, -0.49493980407714844, 12.03292465209961, -2.719799041748047, 2.4220809936523438, -4.025299072265625, 14.981491088867188, -1.5722846984863281, 7.576850891113281, 2.2275733947753906, -9.827621459960938, -3.5582275390625, 9.07666015625, 1.0052947998046875, -4.138519287109375, -5.0828094482421875, 10.503303527832031, 9.842887878417969, 2.889678955078125, 2.4682159423828125, -0.40816497802734375, -7.555694580078125, 0.22366905212402344, -0.7254714965820312, -2.397430419921875, 3.4417953491210938, -4.288536071777344, -0.22747802734375, 10.556026458740234, -0.14574432373046875, -3.9156951904296875, -0.5717601776123047, 0.5736465454101562, 3.3991241455078125, 2.4412803649902344, -11.439666748046875, 7.115703582763672, 2.7607574462890625, 1.4171104431152344, 8.265045166015625, 12.935470581054688, 8.130706787109375, -3.7527923583984375, -1.3338241577148438, -0.7575225830078125, -2.6166248321533203, 5.256664276123047, 15.78448486328125, -13.154647827148438, 7.897617340087891, 6.242218017578125, -0.16776275634765625, 9.56452751159668, 7.588174819946289, 2.288484573364258, -0.6591949462890625, -3.249462127685547, -0.8334770202636719], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000143.npy"}
{"epoch": 0.2161753590325019, "step": 144, "batch_size": 64, "mean": 1.5377315282821655, "std": 6.939415454864502, "min": -12.092117309570312, "p10": -6.789602279663085, "median": 0.0988311767578125, "p90": 10.635671234130863, "max": 20.897750854492188, "pos_frac": 0.515625, "sample": [8.96487045288086, 5.1621246337890625, -11.112876892089844, -0.7752723693847656, 0.0798492431640625, 4.524202346801758, 0.1178131103515625, 8.780830383300781, 5.289215087890625, 1.5778579711914062, -4.250953674316406, 13.84451675415039, -7.075695037841797, -0.6556415557861328, -2.3915863037109375, -12.092117309570312, 10.887260437011719, -0.013996124267578125, -7.847576141357422, 8.392013549804688, 1.040313720703125, -2.6158065795898438, -3.016693115234375, 4.055999755859375, -5.909694671630859, -3.73687744140625, -5.888092041015625, 7.890129089355469, -2.382099151611328, -0.7596626281738281, 11.862060546875, -4.2649993896484375, -5.923103332519531, -0.9999847412109375, 6.487091064453125, -0.6605987548828125, 1.792755126953125, 12.739971160888672, 17.184112548828125, 8.257736206054688, 3.15020751953125, 1.9492721557617188, -7.028297424316406, 6.635955810546875, -6.232646942138672, -1.3879241943359375, 6.873920440673828, -5.279659271240234, 7.699745178222656, -3.1193389892578125, -0.2718830108642578, -4.21636962890625, 2.981182098388672, 9.276556015014648, -8.670852661132812, 12.329353332519531, 20.897750854492188, 10.048629760742188, -0.744537353515625, -7.921974182128906, 4.223197937011719, -2.0951976776123047, 1.736215591430664, 1.0241127014160156], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000144.npy"}
{"epoch": 0.21768707482993196, "step": 145, "batch_size": 64, "mean": 2.7969002723693848, "std": 6.737921714782715, "min": -16.716751098632812, "p10": -6.2876945495605465, "median": 3.0442142486572266, "p90": 10.492483901977545, "max": 18.20166015625, "pos_frac": 0.765625, "sample": [4.617641448974609, 4.550849914550781, 4.597272872924805, 6.73272705078125, 0.39156341552734375, -16.716751098632812, -0.4813079833984375, 0.01451873779296875, 14.502540588378906, -7.931854248046875, 3.8607521057128906, 3.727336883544922, 2.1739587783813477, 7.8262939453125, 1.8996849060058594, 0.93084716796875, 2.0649948120117188, 14.250419616699219, -9.553207397460938, 6.483001708984375, 3.217071533203125, 2.2191162109375, 0.8500175476074219, 7.124786376953125, 9.067615509033203, 1.650238037109375, 1.1603622436523438, 2.93780517578125, -2.3863525390625, -7.201435089111328, 11.103141784667969, 2.8714599609375, 4.999908447265625, 16.471160888671875, -12.11724853515625, 4.713022232055664, 7.014789581298828, 1.1956672668457031, 6.801685333251953, 3.150623321533203, -5.9769744873046875, 7.070304870605469, 15.102767944335938, 6.231269836425781, -6.420860290527344, 2.75848388671875, 2.4139060974121094, 6.510101318359375, -1.290374755859375, 14.363357543945312, 6.811107635498047, 1.2189521789550781, 6.7583770751953125, 0.19738006591796875, -0.22554779052734375, 6.558563232421875, 18.20166015625, -2.4523239135742188, -11.336257934570312, 3.2074203491210938, 6.82269287109375, -5.0453033447265625, -5.0253448486328125, 3.7635345458984375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000145.npy"}
{"epoch": 0.21919879062736206, "step": 146, "batch_size": 64, "mean": 2.072317600250244, "std": 5.719551086425781, "min": -12.541839599609375, "p10": -4.9481302261352536, "median": 2.187885284423828, "p90": 8.481858825683593, "max": 21.4583740234375, "pos_frac": 0.671875, "sample": [3.2983245849609375, 0.1961822509765625, 1.0406646728515625, 6.9669036865234375, 3.9212646484375, -6.508594512939453, 0.5802841186523438, -3.4677467346191406, 6.287258148193359, -3.859466552734375, -9.864479064941406, -12.541839599609375, 5.701080322265625, 6.264705657958984, -5.344642639160156, -4.82281494140625, 9.267967224121094, -2.941802978515625, 12.027275085449219, -5.79205322265625, 4.620449066162109, 4.33880615234375, -0.2792816162109375, -4.334661483764648, -5.001836776733398, 6.364406585693359, -0.45293235778808594, 1.7161483764648438, -2.5241241455078125, 1.3312606811523438, 6.488582611083984, -3.805908203125, 11.340301513671875, 2.6596221923828125, 5.9937591552734375, -0.8769454956054688, 3.443866729736328, -0.8847274780273438, -4.410003662109375, 0.9686965942382812, 10.307769775390625, 0.6653785705566406, -1.3439064025878906, 3.7737503051757812, 4.32269287109375, 8.444816589355469, 5.918067932128906, 3.011688232421875, 8.277034759521484, 3.170055389404297, 4.545976638793945, -5.40863037109375, 8.94906234741211, -3.5048446655273438, 1.5619220733642578, 0.004486083984375, 21.4583740234375, 7.296661376953125, 1.0683574676513672, 3.9446144104003906, 6.105701446533203, 0.6951751708984375, 3.7624282836914062, 8.497734069824219], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000146.npy"}
{"epoch": 0.22071050642479215, "step": 147, "batch_size": 64, "mean": 1.9967000484466553, "std": 6.012650966644287, "min": -10.506317138671875, "p10": -4.726283264160156, "median": 1.0494632720947266, "p90": 10.862616348266602, "max": 15.209281921386719, "pos_frac": 0.625, "sample": [0.9004096984863281, 2.9288711547851562, 11.543769836425781, 0.5336151123046875, -0.9287643432617188, 2.7650527954101562, -0.3851776123046875, -3.07861328125, 0.32562255859375, -1.7259941101074219, 0.8373336791992188, 5.727210998535156, 10.560791015625, 4.32672119140625, -1.2981948852539062, 6.388298034667969, 10.99197006225586, -10.506317138671875, 8.404563903808594, -0.2985830307006836, 7.022138595581055, 3.5647964477539062, -2.0676727294921875, -4.624908447265625, -2.3508453369140625, 4.310455322265625, -0.8744277954101562, 3.811105728149414, 1.1781196594238281, 5.314952850341797, -2.6792068481445312, 0.920806884765625, -0.34478759765625, -9.172332763671875, 6.589759826660156, 6.523170471191406, 6.075675964355469, 14.512886047363281, 14.333152770996094, -3.9502601623535156, 10.340408325195312, 5.038017272949219, -4.7697296142578125, 3.0740966796875, 1.7877826690673828, -4.856235504150391, 13.912765502929688, -2.666248321533203, 15.209281921386719, -3.3287429809570312, -4.306388854980469, 3.839550018310547, 0.16522216796875, 2.0206336975097656, -8.059295654296875, 2.1667709350585938, 0.48825836181640625, -4.289712905883789, 3.2553024291992188, -7.819671630859375, -6.966224670410156, 3.1761550903320312, 14.233444213867188, 0.038204193115234375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000147.npy"}
{"epoch": 0.2222222222222222, "step": 148, "batch_size": 64, "mean": 3.392505407333374, "std": 6.88906717300415, "min": -12.774070739746094, "p10": -5.09604091644287, "median": 2.6921043395996094, "p90": 11.79086990356446, "max": 24.620590209960938, "pos_frac": 0.75, "sample": [6.7637481689453125, -4.228593826293945, 0.166534423828125, -8.400325775146484, 2.574188232421875, 10.332588195800781, 9.2799072265625, -5.467803955078125, 7.862884521484375, 12.415847778320312, -3.491769790649414, 2.0923328399658203, 7.668632507324219, 6.688102722167969, 6.754058837890625, -8.314804077148438, -2.3747901916503906, 0.9384765625, 4.627067565917969, -5.779655456542969, 1.6178627014160156, -8.772666931152344, 24.620590209960938, 4.275557518005371, 1.2508277893066406, 9.72216796875, 8.242729187011719, 0.7482147216796875, 19.3955078125, 5.998584747314453, 4.490947723388672, 6.952880859375, 15.831531524658203, -3.3585357666015625, 4.270195007324219, 14.597694396972656, 1.2679576873779297, 2.68902587890625, -12.774070739746094, 0.16773223876953125, 5.779258728027344, -2.0345916748046875, 3.1618270874023438, 14.326576232910156, -8.940685272216797, 2.2096729278564453, 6.043376922607422, 2.5203514099121094, 13.692161560058594, -2.9723968505859375, 5.13252067565918, 1.2065200805664062, 1.6604251861572266, -0.90936279296875, 1.672292709350586, 1.662017822265625, 9.864532470703125, 5.5835723876953125, 2.6951828002929688, 4.280815124511719, 3.6610488891601562, 9.614906311035156, -2.9016647338867188, -1.2293701171875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000148.npy"}
{"epoch": 0.2237339380196523, "step": 149, "batch_size": 64, "mean": 2.0515592098236084, "std": 5.491276741027832, "min": -13.151130676269531, "p10": -4.4337223052978505, "median": 1.9888038635253906, "p90": 8.946919250488286, "max": 13.511756896972656, "pos_frac": 0.640625, "sample": [6.686180114746094, 6.249443054199219, 1.4981613159179688, -0.5356063842773438, -2.22735595703125, 1.4461936950683594, -4.824287414550781, 1.2778854370117188, -5.153495788574219, 11.558731079101562, -13.151130676269531, 2.908781051635742, 4.838115692138672, 1.6061134338378906, -2.2718353271484375, -3.3875083923339844, 0.43996429443359375, 1.98565673828125, 7.282379150390625, 7.8559112548828125, 3.3823776245117188, 4.841339111328125, 9.338829040527344, -2.0002670288085938, 13.511756896972656, -1.8125534057617188, 2.950571060180664, -0.33018970489501953, 4.430446624755859, -7.224231719970703, 3.114694595336914, 4.044929504394531, 6.5809173583984375, 13.360570907592773, 4.568265914916992, 13.037544250488281, 7.726142883300781, -6.524532318115234, 1.785919189453125, -3.5224037170410156, 1.83856201171875, 11.854461669921875, -1.9863910675048828, 9.773178100585938, 4.6201324462890625, 4.520465850830078, 1.9919509887695312, -6.664516448974609, -0.7014875411987305, 3.9965362548828125, -2.0345802307128906, 3.1094894409179688, -2.5305633544921875, -10.6787109375, -2.4244585037231445, 0.8320159912109375, 2.6045684814453125, 8.032463073730469, 2.4983863830566406, 4.238559722900391, -1.6739349365234375, 6.177816390991211, -1.170684814453125, -0.26589202880859375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000149.npy"}
{"epoch": 0.2252456538170824, "step": 150, "batch_size": 64, "mean": 3.746462106704712, "std": 7.216642379760742, "min": -10.735149383544922, "p10": -6.128597259521484, "median": 4.057223320007324, "p90": 12.91476097106934, "max": 20.00762939453125, "pos_frac": 0.6875, "sample": [6.567516326904297, 1.873316764831543, 17.86871337890625, 2.848052978515625, -8.241256713867188, 1.4227561950683594, 3.7402420043945312, 0.7021560668945312, -2.0186767578125, 5.484355926513672, -0.48773193359375, 9.625686645507812, -7.42828369140625, -6.163032531738281, -1.3604011535644531, 6.484222412109375, 1.4569320678710938, 16.411041259765625, -9.99090576171875, -6.048248291015625, 10.442230224609375, -0.4611320495605469, 16.553848266601562, -10.031906127929688, -0.3511810302734375, 7.538339614868164, 4.843891143798828, 4.831291198730469, 7.741661071777344, 13.36214828491211, 3.0275230407714844, 6.886932373046875, 11.0062255859375, 7.236183166503906, 6.889427185058594, -3.984130859375, 2.8775711059570312, 5.050676345825195, 10.3558349609375, 14.3785400390625, 5.842437744140625, 10.125717163085938, -6.663795471191406, 2.3550796508789062, 17.741336822509766, -4.187652587890625, 4.374204635620117, 7.106670379638672, -10.735149383544922, 20.00762939453125, 6.31280517578125, -1.8369064331054688, 8.734344482421875, -3.744213104248047, 1.1606483459472656, 6.544960021972656, -2.948272705078125, -1.86492919921875, 11.870857238769531, -3.8640518188476562, 2.6315231323242188, 8.772018432617188, 1.169952392578125, 9.92791748046875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000150.npy"}
{"epoch": 0.22675736961451248, "step": 151, "batch_size": 64, "mean": 3.980025291442871, "std": 6.261089324951172, "min": -10.311302185058594, "p10": -2.1533870697021484, "median": 4.001585960388184, "p90": 10.39380226135254, "max": 23.292877197265625, "pos_frac": 0.671875, "sample": [7.011039733886719, 11.465328216552734, 7.184831619262695, -0.4861602783203125, 4.949316024780273, 0.29744720458984375, 21.258880615234375, -1.7040748596191406, -1.4937744140625, -0.6349582672119141, 4.140445709228516, -0.21479415893554688, 4.6086273193359375, -10.311302185058594, 6.608392715454102, 5.848833084106445, 9.802520751953125, -2.1028785705566406, 2.7882003784179688, -0.2541236877441406, -1.4647674560546875, -2.6296615600585938, 9.318466186523438, 0.21369171142578125, -0.18963623046875, -6.6862335205078125, 4.650056838989258, 7.165916442871094, -1.7061595916748047, -2.335693359375, 5.049893379211426, -1.6532363891601562, -0.60870361328125, 7.4202728271484375, 3.114715576171875, -0.27588653564453125, 2.837066650390625, -3.9252395629882812, 1.6695556640625, 23.292877197265625, 1.7158012390136719, 6.742393493652344, 3.2146682739257812, 8.813161849975586, 17.981983184814453, 15.739494323730469, 9.18111801147461, 4.09503173828125, 10.569923400878906, 5.601829528808594, 2.791027069091797, 7.594940185546875, -2.1750335693359375, 4.9511260986328125, -7.500698089599609, 6.641244888305664, 5.908262252807617, 9.839962005615234, -2.072010040283203, 4.712432861328125, 1.385345458984375, 9.982852935791016, 3.908140182495117, 13.079544067382812], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000151.npy"}
{"epoch": 0.22826908541194255, "step": 152, "batch_size": 64, "mean": 2.8387527465820312, "std": 6.621760368347168, "min": -11.68768310546875, "p10": -5.080463027954101, "median": 2.5074024200439453, "p90": 10.41343574523926, "max": 21.7152099609375, "pos_frac": 0.6875, "sample": [1.521759033203125, 0.69677734375, 6.259269714355469, 1.7890625, -11.07330322265625, 6.7108154296875, 7.170906066894531, 5.536796569824219, 12.962776184082031, -2.0015106201171875, 8.871644973754883, -2.7294464111328125, 1.3780441284179688, -11.282196044921875, 5.488128662109375, 6.4886474609375, 0.4560222625732422, 5.3105926513671875, -0.9571037292480469, -3.3023223876953125, -7.702476501464844, 11.978134155273438, 6.624713897705078, -6.1519775390625, -4.331615447998047, 0.37487030029296875, 3.6195144653320312, 4.367179870605469, 3.661989212036133, 10.710182189941406, 1.4837169647216797, 9.721027374267578, 2.3755226135253906, -0.14263916015625, 1.7868194580078125, -11.68768310546875, 0.2551155090332031, 1.3130264282226562, 8.37860107421875, -2.77783203125, -3.1155929565429688, 7.530693054199219, 7.352783203125, -3.1941680908203125, 11.558784484863281, 9.474624633789062, 3.2394752502441406, 5.900630950927734, -0.1313934326171875, 3.781402587890625, -7.834075927734375, 2.6392822265625, 0.24323654174804688, 20.979408264160156, 7.097921371459961, 9.212295532226562, 11.190017700195312, -1.0881423950195312, 3.431671142578125, 21.7152099609375, -1.6090011596679688, -5.401397705078125, -0.9149246215820312, 6.469871520996094], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000152.npy"}
{"epoch": 0.22978080120937264, "step": 153, "batch_size": 64, "mean": 3.1567811965942383, "std": 5.953783988952637, "min": -11.257251739501953, "p10": -3.9953002929687496, "median": 2.700010299682617, "p90": 11.0498046875, "max": 15.849205017089844, "pos_frac": 0.703125, "sample": [9.232513427734375, 2.3111190795898438, -3.5731277465820312, 6.8199615478515625, 12.35232162475586, 3.2594451904296875, -2.6129379272460938, -1.171173095703125, 4.582786560058594, -0.08868408203125, 1.7376365661621094, 5.781871795654297, 1.32867431640625, 15.25927734375, 3.9355697631835938, 9.846694946289062, 0.16042327880859375, -2.684661865234375, 4.169395446777344, 11.130668640136719, 3.6186065673828125, 4.38555908203125, -11.257251739501953, -4.6949615478515625, 15.849205017089844, -3.2128219604492188, 3.4727210998535156, 1.489858627319336, -2.6341552734375, 3.7075767517089844, 0.52301025390625, 6.265926361083984, -7.125938415527344, -3.285022735595703, 6.1834716796875, 5.233543395996094, -7.4263458251953125, 9.890426635742188, 11.65765380859375, 2.0511932373046875, -5.69049072265625, 1.0128822326660156, 10.794692993164062, -5.4397735595703125, 14.315933227539062, -0.5239582061767578, 10.861122131347656, 9.22784423828125, 8.01849365234375, -0.7071247100830078, -1.9149932861328125, 1.610198974609375, 1.2139091491699219, -4.176231384277344, 3.0889015197753906, 0.9409103393554688, -2.5227279663085938, 6.142662048339844, 3.92620849609375, 10.220382690429688, 10.301128387451172, 1.8251209259033203, 1.617156982421875, 11.421722412109375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000153.npy"}
{"epoch": 0.23129251700680273, "step": 154, "batch_size": 64, "mean": 2.216388702392578, "std": 6.587747097015381, "min": -13.582893371582031, "p10": -6.993770599365234, "median": 2.5121593475341797, "p90": 11.327079391479497, "max": 16.086647033691406, "pos_frac": 0.671875, "sample": [-13.582893371582031, -7.30316162109375, -7.8299102783203125, -4.7663421630859375, 11.878646850585938, 3.7229156494140625, -5.629119873046875, -1.6900062561035156, -8.375503540039062, -11.249267578125, 3.543516159057617, 0.23741722106933594, 2.773273468017578, 8.881866455078125, -0.1492919921875, -7.040519714355469, 2.70416259765625, 3.945812225341797, 1.6821060180664062, -4.933082580566406, -0.7482528686523438, 16.086647033691406, 13.433027267456055, 4.97747802734375, -6.8846893310546875, 15.075252532958984, 7.36126708984375, -5.8222503662109375, 3.772552490234375, -9.268890380859375, 8.449090957641602, -0.16672515869140625, 3.3287506103515625, 6.211761474609375, 11.867950439453125, -3.843841552734375, 7.141124725341797, 11.785823822021484, 1.6378583908081055, -2.332660675048828, 0.5792350769042969, 0.0900726318359375, 6.5725250244140625, -4.197998046875, 2.0579757690429688, 5.956348419189453, 13.629444122314453, 0.9528617858886719, 4.036216735839844, 0.350067138671875, 4.926727294921875, 2.0137252807617188, -1.5853958129882812, 7.9133758544921875, 7.373291015625, 10.256675720214844, 2.4431304931640625, -0.8492527008056641, 7.597129821777344, 9.7646484375, 0.5234527587890625, 2.831709861755371, 2.581188201904297, 7.149833679199219], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000154.npy"}
{"epoch": 0.2328042328042328, "step": 155, "batch_size": 64, "mean": 3.2796502113342285, "std": 6.881336212158203, "min": -14.856525421142578, "p10": -5.641850090026855, "median": 2.791341781616211, "p90": 13.030307769775392, "max": 16.461132049560547, "pos_frac": 0.703125, "sample": [1.0169219970703125, 0.5231170654296875, 13.97076416015625, 2.7172889709472656, -2.05596923828125, 6.1689300537109375, 2.348613739013672, 8.89175033569336, -5.45378303527832, 10.85150146484375, 4.860267639160156, 4.421727180480957, -1.248931884765625, 8.243537902832031, -2.224395751953125, 1.7051620483398438, 0.7804412841796875, 15.2354736328125, 1.5407333374023438, 13.265872955322266, 12.717643737792969, -0.6341323852539062, 3.2880859375, 0.2550811767578125, 6.437103271484375, 16.461132049560547, -5.722450256347656, 3.4696807861328125, -1.0101737976074219, 7.950714111328125, -6.27178955078125, 6.14122200012207, 15.95394515991211, 6.567924499511719, 13.164306640625, 5.320953369140625, -2.7198638916015625, 10.873458862304688, -11.313705444335938, 4.705657958984375, -2.6125335693359375, 7.494533538818359, 0.42267608642578125, -0.6141319274902344, 14.539764404296875, 4.969390869140625, 3.4728622436523438, 5.14544677734375, -13.615203857421875, 2.502971649169922, 11.363548278808594, -0.690704345703125, 1.3799476623535156, -1.0548553466796875, 1.9242897033691406, 2.2898712158203125, 2.8653945922851562, -6.5500335693359375, -14.856525421142578, -0.6661529541015625, 12.479293823242188, 7.2683868408203125, 8.023981094360352, -6.778419494628906], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000155.npy"}
{"epoch": 0.23431594860166288, "step": 156, "batch_size": 64, "mean": 3.818819522857666, "std": 6.0648956298828125, "min": -10.808975219726562, "p10": -2.442132568359375, "median": 3.394817352294922, "p90": 12.305561828613282, "max": 17.096111297607422, "pos_frac": 0.734375, "sample": [-0.828521728515625, -2.2873497009277344, -0.19055557250976562, 0.09986114501953125, 12.434051513671875, 11.162567138671875, 14.284273147583008, -6.19000244140625, 2.7276077270507812, 10.087638854980469, 2.9000244140625, 17.096111297607422, -8.047348022460938, 3.672088623046875, -2.442535400390625, 9.314693450927734, 9.930557250976562, 13.479660034179688, 0.8459739685058594, 5.8957672119140625, 2.3164186477661133, 1.3940563201904297, 3.0378990173339844, 6.315887451171875, 3.1341552734375, 4.7021026611328125, -0.27087974548339844, 10.928779602050781, 6.1031951904296875, 3.3967208862304688, -1.2388839721679688, 6.22088623046875, -0.41124725341796875, -2.0364532470703125, 13.617355346679688, 13.990470886230469, 1.910247802734375, 0.605926513671875, -2.4009323120117188, 7.0064239501953125, 12.005752563476562, 3.392913818359375, -1.2941741943359375, 0.5146903991699219, 5.3335723876953125, 3.4094161987304688, -3.6612472534179688, -2.441192626953125, 4.0816192626953125, 5.170196533203125, 0.6467323303222656, 5.7314453125, 5.807828903198242, -10.808975219726562, 7.8838653564453125, 0.7929611206054688, -7.8675994873046875, 6.274879455566406, 11.484512329101562, 2.2101669311523438, 8.13138198852539, 14.210845947265625, 7.9203033447265625, -6.7921295166015625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000156.npy"}
{"epoch": 0.23582766439909297, "step": 157, "batch_size": 64, "mean": 2.6955344676971436, "std": 6.390186786651611, "min": -14.248001098632812, "p10": -3.98048858642578, "median": 2.6409988403320312, "p90": 9.774538040161136, "max": 21.62519073486328, "pos_frac": 0.65625, "sample": [16.94278335571289, 12.616268157958984, 0.9498672485351562, 6.858089447021484, 3.9197235107421875, -14.248001098632812, 3.7901763916015625, 8.024749755859375, -0.4179649353027344, -3.1400184631347656, 2.314586639404297, -0.45189666748046875, 8.2373046875, -0.865203857421875, -6.88861083984375, 0.23679542541503906, -2.3902511596679688, 0.7051315307617188, 8.908149719238281, 6.063783645629883, 2.7446136474609375, 10.14584732055664, 21.62519073486328, 2.8262901306152344, 2.537384033203125, 2.7863845825195312, -7.621124267578125, -3.1425628662109375, 4.040809631347656, -0.4917182922363281, 6.054841995239258, -3.1036758422851562, 2.8678359985351562, 5.061565399169922, 5.447338104248047, 14.990936279296875, 1.279205322265625, 6.533514022827148, 12.780487060546875, -1.4014205932617188, 4.785388946533203, 6.53936767578125, -1.5907554626464844, 14.693328857421875, -4.339599609375, -7.085611343383789, -0.4517669677734375, 5.840250015258789, -1.6088180541992188, -0.628692626953125, 8.066131591796875, -4.913021087646484, -2.594898223876953, 1.0694999694824219, 3.4466323852539062, 7.926811218261719, 1.2841033935546875, 7.538135528564453, -11.19451904296875, 1.8191070556640625, 5.4460906982421875, 0.01926422119140625, -2.821014404296875, 4.1415863037109375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000157.npy"}
{"epoch": 0.23733938019652306, "step": 158, "batch_size": 64, "mean": 3.13816237449646, "std": 6.929567337036133, "min": -12.594558715820312, "p10": -5.448365783691406, "median": 3.158781051635742, "p90": 12.143331909179691, "max": 23.65228271484375, "pos_frac": 0.6875, "sample": [-0.16943359375, -3.3694190979003906, 0.6097869873046875, 3.72613525390625, -0.480377197265625, 4.330543518066406, 14.658950805664062, 14.591812133789062, 0.6960601806640625, 1.2071990966796875, 4.994377136230469, 8.571823120117188, 8.860916137695312, 8.007736206054688, -7.842449188232422, 5.891166687011719, 23.65228271484375, 1.1830482482910156, -0.9783315658569336, 9.421989440917969, -5.290733337402344, 0.789459228515625, 7.338279724121094, 6.876552581787109, 10.146156311035156, 12.869224548339844, 1.4043121337890625, 0.882476806640625, 14.551193237304688, -5.16497802734375, -12.594558715820312, 0.38471221923828125, 6.257335662841797, -4.6832427978515625, 8.591724395751953, 8.598609924316406, 13.649337768554688, -0.496795654296875, -3.6391983032226562, -8.625953674316406, 10.993377685546875, 12.63616943359375, 3.721893310546875, 10.049652099609375, -4.90130615234375, 6.1214447021484375, -1.089691162109375, 6.4456939697265625, 10.65869140625, 6.083580017089844, 3.5441856384277344, 3.9333724975585938, -7.073041915893555, -5.515922546386719, 0.31601524353027344, -2.3427505493164062, -1.2138595581054688, 2.4229888916015625, 4.885898590087891, 1.2316703796386719, -5.652034759521484, 2.77337646484375, 3.8293533325195312, -10.424102783203125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000158.npy"}
{"epoch": 0.23885109599395313, "step": 159, "batch_size": 64, "mean": 3.6435976028442383, "std": 6.4621262550354, "min": -19.493331909179688, "p10": -2.9438739776611325, "median": 3.326251983642578, "p90": 12.00088882446289, "max": 19.588058471679688, "pos_frac": 0.765625, "sample": [8.55499267578125, 4.594062805175781, 1.7313003540039062, 15.050033569335938, 7.6085968017578125, -8.170196533203125, 17.56927490234375, 9.407432556152344, 1.8540611267089844, 10.2620849609375, 3.6894378662109375, 0.6190032958984375, 11.828441619873047, -4.1966094970703125, 0.9186592102050781, 8.830322265625, 5.4269866943359375, 19.588058471679688, 10.830463409423828, 2.1384410858154297, 5.033653259277344, -2.1621246337890625, 1.3887338638305664, 2.0784378051757812, 5.682609558105469, 0.5670166015625, 0.37874794006347656, -2.0945053100585938, 7.033416748046875, 12.07479476928711, -6.842571258544922, 13.004348754882812, 7.298196792602539, -0.7230758666992188, 4.2357940673828125, 3.0549793243408203, 0.534820556640625, 3.35595703125, -4.1560821533203125, -0.06598663330078125, 3.488483428955078, 0.5526275634765625, 15.810165405273438, -0.4251976013183594, 0.8775825500488281, 4.784246444702148, -2.620830535888672, 5.891193389892578, -3.0823211669921875, 6.7135162353515625, 0.6914825439453125, 14.13261604309082, 4.4050750732421875, 4.006309509277344, -1.4881458282470703, -19.493331909179688, 5.214811325073242, -3.6545257568359375, 1.9352035522460938, 0.6717910766601562, 4.736660003662109, -1.4391021728515625, 3.2965469360351562, 10.373374938964844], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000159.npy"}
{"epoch": 0.24036281179138322, "step": 160, "batch_size": 64, "mean": 2.933891773223877, "std": 6.035837173461914, "min": -8.504730224609375, "p10": -4.962599563598633, "median": 3.3675947189331055, "p90": 10.89316787719727, "max": 20.032058715820312, "pos_frac": 0.6875, "sample": [4.031810760498047, -2.4238510131835938, 8.144454956054688, 8.570121765136719, -1.935892105102539, 11.358299255371094, 2.0346832275390625, -2.4663658142089844, 12.454524993896484, -1.9595870971679688, 17.36151123046875, 0.6923370361328125, 8.788393020629883, 3.5630874633789062, 3.3584041595458984, 2.4929962158203125, 3.4282054901123047, -7.405925750732422, 3.748159408569336, -5.874107360839844, 0.8462600708007812, -4.930660247802734, 4.5382080078125, 0.05181121826171875, 8.601139068603516, 5.335426330566406, 8.095043182373047, 0.3927440643310547, 4.960845947265625, 5.257511138916016, 11.431549072265625, 3.3767852783203125, 20.032058715820312, 13.921703338623047, 0.14870643615722656, 1.145904541015625, 3.9262771606445312, -1.0909576416015625, 4.568483352661133, 4.593620300292969, -0.233489990234375, 7.008087158203125, -6.1319427490234375, -8.504730224609375, -1.5351886749267578, 4.70574951171875, -4.976287841796875, -3.441516876220703, 3.5718154907226562, 4.777286529541016, 6.5247955322265625, -5.358238220214844, 2.6515426635742188, 9.807861328125, 0.392333984375, -1.837890625, -0.8060302734375, 9.290840148925781, -4.149467468261719, 2.9366531372070312, 7.7695159912109375, 12.804210662841797, -4.0467529296875, -6.613800048828125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000160.npy"}
{"epoch": 0.2418745275888133, "step": 161, "batch_size": 64, "mean": 3.161600112915039, "std": 6.1573381423950195, "min": -19.682212829589844, "p10": -1.5079566955566404, "median": 3.006074905395508, "p90": 11.632267379760744, "max": 17.12154769897461, "pos_frac": 0.75, "sample": [8.785980224609375, -13.413459777832031, -1.3272476196289062, 7.4398651123046875, -1.5854034423828125, -0.8101654052734375, 4.65606689453125, 12.642837524414062, 1.6001434326171875, 2.372537612915039, -4.958282470703125, 0.2003021240234375, 7.985136032104492, 3.7493667602539062, 3.2469406127929688, -0.9128799438476562, 3.3368568420410156, -19.682212829589844, 5.272483825683594, -8.433731079101562, -0.2730560302734375, 4.8123321533203125, 4.608436584472656, -0.21596527099609375, 10.168121337890625, -0.6371803283691406, 4.2691497802734375, 1.7995376586914062, -4.821632385253906, 12.911064147949219, 4.102146148681641, 3.6083526611328125, 3.143077850341797, 0.0413818359375, 10.8907470703125, 3.4267005920410156, 0.1802959442138672, 3.3918304443359375, 2.8690719604492188, 12.619461059570312, 0.9996490478515625, 8.415748596191406, -0.12177276611328125, 1.1851654052734375, 7.364040374755859, 1.5706863403320312, 2.255847930908203, 4.121068954467773, 7.125114440917969, 8.650688171386719, 0.17517852783203125, 10.566699981689453, 13.835166931152344, 2.263427734375, 0.8416290283203125, -0.10450172424316406, 0.3203125, -0.7627639770507812, 11.950061798095703, 12.779777526855469, 17.12154769897461, 2.4106979370117188, -3.5268421173095703, 5.846778869628906], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000161.npy"}
{"epoch": 0.24338624338624337, "step": 162, "batch_size": 64, "mean": 4.608911514282227, "std": 5.775290012359619, "min": -16.430450439453125, "p10": -2.1179763793945305, "median": 5.109415054321289, "p90": 11.494060516357424, "max": 20.790512084960938, "pos_frac": 0.78125, "sample": [7.6812896728515625, 2.0338668823242188, -16.430450439453125, 1.403045654296875, 5.262359619140625, 5.062431335449219, 2.6055450439453125, 5.9367828369140625, 7.466339111328125, 12.226848602294922, 5.6376953125, 2.9954185485839844, -0.6261253356933594, 20.790512084960938, 3.50341796875, 9.129268646240234, 3.938934326171875, 9.633697509765625, 1.845794677734375, 10.94814682006836, 11.809928894042969, 7.184207916259766, 4.479278564453125, -4.0504150390625, 8.005870819091797, 6.7015380859375, -4.610992431640625, 4.352317810058594, -3.82958984375, 9.502777099609375, 2.023378372192383, 0.251983642578125, -1.5323715209960938, 9.56985855102539, 6.2482757568359375, 12.604839324951172, 11.728023529052734, 4.62713623046875, 0.37452125549316406, -0.9315681457519531, -1.0346221923828125, 5.380241394042969, 18.821487426757812, 12.313995361328125, 5.156398773193359, -2.3689498901367188, -1.130096435546875, -1.2765426635742188, 6.526420593261719, -0.7449951171875, -2.6431312561035156, 6.1046905517578125, 7.0207672119140625, 7.390846252441406, 8.333488464355469, 5.051979064941406, 7.597766876220703, -2.5323028564453125, 2.9363861083984375, 2.3863697052001953, 3.7381458282470703, 8.807430267333984, 7.362911224365234, 6.217811584472656], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000162.npy"}
{"epoch": 0.24489795918367346, "step": 163, "batch_size": 64, "mean": 3.6132428646087646, "std": 6.373258590698242, "min": -8.459403991699219, "p10": -5.714022064208985, "median": 3.670795440673828, "p90": 11.920911216735846, "max": 18.388336181640625, "pos_frac": 0.6875, "sample": [-2.37908935546875, 3.5424423217773438, 1.5301055908203125, 3.9057693481445312, 3.7214813232421875, 4.225067138671875, -7.531970977783203, -0.3708953857421875, 18.388336181640625, -0.70819091796875, 6.491371154785156, 14.444183349609375, 10.017654418945312, 3.6201095581054688, 3.1832275390625, -0.5036392211914062, 18.100208282470703, 0.38323211669921875, 6.914794921875, 15.389732360839844, -0.9893035888671875, 4.495826721191406, 9.079166412353516, -1.8848190307617188, -5.71923828125, 12.909046173095703, 2.9980125427246094, 6.519737243652344, 0.7586593627929688, -0.1331634521484375, 8.572538375854492, 0.617401123046875, 6.333549499511719, 4.891483306884766, 7.7107391357421875, 4.2640533447265625, -6.079399108886719, 7.814247131347656, -4.397830963134766, 3.5954017639160156, 5.888866424560547, 12.487205505371094, -5.701850891113281, 9.330764770507812, -7.025970458984375, -1.7585716247558594, 1.7871894836425781, 4.059688568115234, -0.30890655517578125, -6.541259765625, -7.649940490722656, 9.590286254882812, 5.233711242675781, 0.3754711151123047, -0.1917266845703125, -8.459403991699219, 15.370018005371094, 4.620033264160156, 9.777229309082031, 10.599557876586914, 7.608970642089844, -2.71173095703125, 3.034944534301758, 8.112930297851562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000163.npy"}
{"epoch": 0.24640967498110355, "step": 164, "batch_size": 64, "mean": 4.495945930480957, "std": 7.003328800201416, "min": -11.209259033203125, "p10": -3.601396942138672, "median": 4.259998321533203, "p90": 14.540938186645509, "max": 16.74982452392578, "pos_frac": 0.734375, "sample": [-2.0004806518554688, -0.700927734375, 0.9226226806640625, -1.9789390563964844, 7.697319030761719, 8.049186706542969, 16.74982452392578, 11.332855224609375, 7.7680816650390625, -1.8412933349609375, 0.7355804443359375, -3.5248184204101562, 6.627195358276367, -0.8007946014404297, 7.775424957275391, 2.11151123046875, 6.660499572753906, 3.1678428649902344, 13.082275390625, -0.7096290588378906, 0.7471923828125, -10.573783874511719, 4.294090270996094, 4.6556549072265625, 0.1426982879638672, 0.1655864715576172, 7.822139739990234, 3.3343734741210938, 3.6531982421875, 4.999080657958984, 8.575439453125, 14.72658920288086, 2.1346588134765625, 5.951572418212891, 4.7054595947265625, -11.209259033203125, 12.841621398925781, 15.054290771484375, 14.649349212646484, -11.003829956054688, -3.7855758666992188, -0.3627166748046875, 4.6485443115234375, 13.985977172851562, 3.9783401489257812, 4.2259063720703125, 8.007514953613281, 0.09021759033203125, 15.260650634765625, 11.9803466796875, 10.714279174804688, -3.63421630859375, 14.287979125976562, -8.331932067871094, -0.30034637451171875, -6.363651275634766, -1.2021026611328125, 16.334266662597656, 15.396858215332031, 1.9073677062988281, 13.989269256591797, 10.345474243164062, 5.891294479370117, 3.8873291015625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000164.npy"}
{"epoch": 0.24792139077853365, "step": 165, "batch_size": 64, "mean": 4.228605270385742, "std": 6.558839321136475, "min": -12.842857360839844, "p10": -2.988765335083008, "median": 4.049417495727539, "p90": 11.534843444824219, "max": 21.0535888671875, "pos_frac": 0.765625, "sample": [19.34522247314453, 6.359813690185547, 9.42216968536377, 9.083015441894531, 7.3251800537109375, 1.1573486328125, 1.5171394348144531, -2.9945068359375, 4.695228576660156, 2.7035140991210938, 6.021293640136719, 3.9569435119628906, 13.190818786621094, 0.09363937377929688, 0.8069839477539062, 4.1418914794921875, 1.7954177856445312, 8.58810043334961, -1.9486160278320312, -1.5145797729492188, -12.842857360839844, -3.093036651611328, 8.515281677246094, -0.9533615112304688, 14.263214111328125, 8.254592895507812, 1.2000579833984375, 5.860691070556641, 10.960186004638672, 11.621063232421875, 3.2201919555664062, -0.9336128234863281, 21.0535888671875, 4.865116119384766, 8.723758697509766, -10.832717895507812, -0.42548370361328125, 4.938022613525391, 11.333663940429688, 8.443939208984375, -0.18139266967773438, -2.9753684997558594, 2.7252655029296875, 2.699260711669922, 10.572357177734375, -3.072021484375, 1.5829315185546875, 4.430694580078125, 4.825296401977539, -2.7434005737304688, 0.07585906982421875, 4.454559326171875, -9.0028076171875, 2.9305877685546875, 3.348886489868164, -6.827415466308594, 8.637863159179688, 5.676277160644531, 9.487625122070312, 3.6579742431640625, 17.185684204101562, 3.2921981811523438, 16.47955322265625, 5.451972961425781], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000165.npy"}
{"epoch": 0.2494331065759637, "step": 166, "batch_size": 64, "mean": 3.097219467163086, "std": 6.417323112487793, "min": -11.0562744140625, "p10": -3.840604782104492, "median": 1.9457206726074219, "p90": 12.884130096435548, "max": 17.207611083984375, "pos_frac": 0.625, "sample": [14.331390380859375, -0.6110458374023438, 1.113739013671875, 7.5546875, 9.815727233886719, -1.9302825927734375, -3.1402740478515625, 16.04668426513672, -0.6043548583984375, 1.408721923828125, 2.4689102172851562, 2.6556167602539062, 10.022407531738281, -11.0562744140625, 8.937957763671875, 4.208717346191406, -0.29489898681640625, 4.255130767822266, 0.6610488891601562, 5.670326232910156, -1.9757137298583984, 10.548053741455078, -6.662391662597656, 1.2771987915039062, 12.611061096191406, -0.15457916259765625, -0.22161865234375, 7.92901611328125, 7.296348571777344, 5.102783203125, -7.3851776123046875, -5.838035583496094, -2.129213333129883, -1.18292236328125, 14.504142761230469, 1.7923469543457031, 12.262924194335938, -3.2041702270507812, 8.46677017211914, 7.140697479248047, 3.41021728515625, 17.207611083984375, -2.2791290283203125, 0.11651611328125, 5.790740966796875, -5.881065368652344, -1.3947219848632812, -3.9367294311523438, 3.598155975341797, 16.757736206054688, 6.029670715332031, 4.750694274902344, -2.9538726806640625, 1.1651477813720703, -3.616313934326172, -0.15857696533203125, 13.035140991210938, -2.097627639770508, 13.00115966796875, -5.388530731201172, 4.321952819824219, 2.7801666259765625, 2.0990943908691406, 0.17315673828125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000166.npy"}
{"epoch": 0.2509448223733938, "step": 167, "batch_size": 64, "mean": 3.750251293182373, "std": 6.08634090423584, "min": -17.519912719726562, "p10": -1.9343416213989257, "median": 4.373200416564941, "p90": 10.739791870117188, "max": 18.33140754699707, "pos_frac": 0.703125, "sample": [4.507438659667969, 4.747974395751953, 3.770153045654297, 4.9244842529296875, -0.28018951416015625, 9.62860107421875, 8.098564147949219, -0.46384429931640625, 1.6013145446777344, 14.125709533691406, 3.171415328979492, -0.32315826416015625, 4.33079719543457, -4.422508239746094, -1.230560302734375, -1.8661251068115234, 4.054962158203125, 15.82708740234375, 9.103500366210938, -8.981178283691406, -5.944019317626953, -0.4901580810546875, 1.3515701293945312, 5.627769470214844, 5.406198501586914, -17.519912719726562, 5.071746826171875, -0.6339130401611328, -0.2415924072265625, 5.0132598876953125, 4.6561737060546875, 9.178007125854492, -2.387526512145996, 8.908527374267578, 5.208744049072266, 4.787841796875, 6.352931976318359, 10.041091918945312, 13.896575927734375, 2.667205810546875, -0.4543495178222656, 11.039253234863281, -1.9635772705078125, 3.2425575256347656, 10.575248718261719, -0.8496856689453125, -6.288719177246094, 10.810310363769531, 0.46319580078125, 4.4156036376953125, 9.991416931152344, 0.4671478271484375, 4.897792816162109, 18.33140754699707, 4.433723449707031, 6.428985595703125, 1.1481552124023438, 6.9610443115234375, 4.072154998779297, 15.962837219238281, 6.0998382568359375, -0.011526107788085938, 0.133392333984375, -1.16510009765625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000167.npy"}
{"epoch": 0.25245653817082386, "step": 168, "batch_size": 64, "mean": 2.981163740158081, "std": 5.286889553070068, "min": -7.188194274902344, "p10": -4.063514709472656, "median": 2.344005584716797, "p90": 9.70596618652344, "max": 18.816574096679688, "pos_frac": 0.703125, "sample": [3.4620513916015625, -0.6257095336914062, 8.161361694335938, -3.6246414184570312, 1.5661258697509766, 7.856361389160156, 2.0684661865234375, -3.9227371215820312, -6.9534149169921875, -5.124664306640625, 2.8097686767578125, 0.06498146057128906, -2.2021942138671875, 10.03353500366211, 6.188934326171875, 6.234272003173828, 8.941638946533203, 6.937213897705078, 2.8255615234375, 11.361587524414062, 7.348602294921875, 3.097532272338867, 0.3967628479003906, 2.9590320587158203, -0.8393058776855469, 1.2951469421386719, 10.729732513427734, 1.516815185546875, -4.6643218994140625, 3.13037109375, -0.415740966796875, 2.2918701171875, 7.3253326416015625, 3.7719039916992188, 5.2916259765625, -4.1709442138671875, 14.413818359375, 0.18148040771484375, -1.0947532653808594, 0.8366470336914062, 5.8784637451171875, 18.816574096679688, 0.8570632934570312, 8.463050842285156, 6.237556457519531, -7.188194274902344, 7.148681640625, 1.015655517578125, 2.3961410522460938, -0.07961273193359375, 13.601058959960938, 1.2006645202636719, 3.05438232421875, -4.6673126220703125, -0.7315292358398438, 11.401885986328125, -4.123847961425781, -0.967529296875, -0.1614837646484375, 0.9136924743652344, 5.597260475158691, 4.7197265625, 8.3079833984375, -0.35596466064453125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000168.npy"}
{"epoch": 0.25396825396825395, "step": 169, "batch_size": 64, "mean": 4.0221123695373535, "std": 6.711499214172363, "min": -14.7320556640625, "p10": -3.9906871795654295, "median": 3.693115234375, "p90": 13.35278282165528, "max": 20.145050048828125, "pos_frac": 0.6875, "sample": [-4.563819885253906, 5.488273620605469, 10.534271240234375, -3.1590003967285156, 2.122211456298828, 0.6299495697021484, 7.024658203125, -1.3644943237304688, 1.7134780883789062, 1.5862236022949219, -5.498847961425781, -1.4787368774414062, 17.3756103515625, 6.0416259765625, 7.5984649658203125, -4.435611724853516, 5.162605285644531, 2.7576141357421875, 8.238319396972656, 3.0199127197265625, -0.609405517578125, 10.654953002929688, 11.87369155883789, 6.224601745605469, -4.082195281982422, 20.145050048828125, 9.012611389160156, 4.3663177490234375, 0.07595062255859375, 5.417694091796875, 0.15860366821289062, 5.280426025390625, 1.254608154296875, -2.5267648696899414, 2.533353805541992, -1.7811622619628906, 13.986679077148438, -5.1231689453125, 9.82412338256836, 4.4113616943359375, -0.5168571472167969, 2.6170501708984375, 6.765775680541992, 5.775119781494141, -0.4427642822265625, 9.538566589355469, 6.917392730712891, -2.375244140625, 8.697906494140625, 11.608001708984375, -2.9966812133789062, 6.353086471557617, 9.5899658203125, -14.7320556640625, -5.18475341796875, 15.51812744140625, 14.869575500488281, 17.79969024658203, -0.7141494750976562, 14.058120727539062, -3.7771682739257812, 8.750173568725586, -1.224700927734375, 0.6309623718261719], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000169.npy"}
{"epoch": 0.25547996976568405, "step": 170, "batch_size": 64, "mean": 2.7229256629943848, "std": 8.282783508300781, "min": -16.884498596191406, "p10": -8.761390018463134, "median": 3.5337095260620117, "p90": 13.923411560058595, "max": 19.732528686523438, "pos_frac": 0.65625, "sample": [14.061203002929688, -9.10171890258789, 3.2687911987304688, -0.5011138916015625, 6.544397354125977, 16.45708465576172, 1.80242919921875, 1.398686408996582, 6.437370300292969, 9.61198616027832, -9.297164916992188, 3.525327682495117, 8.5543212890625, 16.926376342773438, 12.105987548828125, -6.252429962158203, 4.009252548217773, 7.841495513916016, -2.6269073486328125, -8.309677124023438, 2.6794891357421875, 17.359397888183594, -2.3494110107421875, -1.8603477478027344, -7.6302490234375, -3.4529876708984375, -8.497886657714844, 4.2134552001953125, -16.884498596191406, 5.6807403564453125, 9.892841339111328, 5.9365081787109375, 16.322479248046875, 5.416645050048828, -8.874320030212402, 0.48490142822265625, -7.12890625, 3.5420913696289062, 9.609428405761719, -4.4803924560546875, -0.08994293212890625, -4.875925064086914, 1.6484336853027344, 5.64056396484375, 6.159721374511719, 13.601898193359375, 3.7812957763671875, 16.835647583007812, 1.006988525390625, 3.0947418212890625, -8.918701171875, -7.667793273925781, 4.496498107910156, 11.109542846679688, -12.662094116210938, 4.676235198974609, 3.9818878173828125, 19.732528686523438, -3.507526397705078, 4.6197357177734375, 10.943428039550781, 11.53463363647461, 2.3668899536132812, -9.676116943359375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000170.npy"}
{"epoch": 0.25699168556311414, "step": 171, "batch_size": 64, "mean": 4.3172712326049805, "std": 7.5332231521606445, "min": -15.304069519042969, "p10": -4.403187370300293, "median": 4.256259918212891, "p90": 13.54803466796875, "max": 19.961196899414062, "pos_frac": 0.6875, "sample": [4.442512512207031, 1.1230087280273438, 12.416641235351562, 13.5526123046875, 8.284339904785156, 16.216453552246094, 12.558479309082031, 14.442939758300781, 0.8647031784057617, 8.015600204467773, -1.4303512573242188, 4.07000732421875, 7.2977447509765625, 11.714759826660156, 8.451255798339844, 3.3537368774414062, 5.415180206298828, 3.67889404296875, 7.234982490539551, -1.2996368408203125, -0.045589447021484375, 9.479639053344727, 6.203849792480469, 15.55084228515625, 9.095857620239258, -0.13762664794921875, -4.901702880859375, -4.296897888183594, 18.6170654296875, -4.448740005493164, 11.302677154541016, 1.970428466796875, 6.4238739013671875, -3.804248809814453, 7.541603088378906, 6.798336029052734, -4.019294738769531, -15.304069519042969, 14.575786590576172, 13.337265014648438, -15.143646240234375, -0.6308021545410156, 7.464511871337891, 3.3938140869140625, 1.92681884765625, 19.961196899414062, 7.563934326171875, 0.498199462890625, 3.4020767211914062, 8.024803161621094, 3.496112823486328, 13.537353515625, 11.099451065063477, -1.2460556030273438, -11.438385009765625, -5.7296142578125, 6.0420074462890625, 2.0266571044921875, -1.55902099609375, -7.43048095703125, 10.04248046875, -0.025295257568359375, -1.21759033203125, -2.0960922241210938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000171.npy"}
{"epoch": 0.2585034013605442, "step": 172, "batch_size": 64, "mean": 3.8189544677734375, "std": 6.94365119934082, "min": -14.802200317382812, "p10": -2.8043815612792966, "median": 2.6162986755371094, "p90": 12.663570022583007, "max": 20.69732666015625, "pos_frac": 0.6875, "sample": [-2.9456863403320312, 2.4285736083984375, 1.1812896728515625, 8.364242553710938, -2.0657806396484375, 2.8463401794433594, 0.23584556579589844, 11.254196166992188, 13.60162353515625, 12.579891204833984, 20.69732666015625, -2.147064208984375, -0.31354522705078125, -0.026981353759765625, 2.0879287719726562, 18.082080841064453, 8.738410949707031, -0.9492874145507812, 0.7885589599609375, 4.228092193603516, 12.07177734375, -0.07293319702148438, -6.743890762329102, 2.8040237426757812, 4.953178405761719, 0.48637962341308594, -0.2410430908203125, 5.217193603515625, 8.50518798828125, -4.137115478515625, 1.8814964294433594, 1.94879150390625, 7.22784423828125, 12.699432373046875, 1.2617874145507812, -4.359643936157227, 13.342912673950195, -2.4135971069335938, 12.004405975341797, -0.1230010986328125, -2.4465885162353516, 5.438934326171875, 3.7532615661621094, 11.716289520263672, 5.29387092590332, -12.179901123046875, -0.47967529296875, 9.663543701171875, -0.403472900390625, 4.6853790283203125, -9.121173858642578, 12.435150146484375, 15.216049194335938, -14.802200317382812, 3.3469505310058594, 10.341583251953125, 3.6662139892578125, 0.3985137939453125, 0.879302978515625, 14.309890747070312, -2.47467041015625, 2.059375762939453, 9.787269592285156, 8.349945068359375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000172.npy"}
{"epoch": 0.2600151171579743, "step": 173, "batch_size": 64, "mean": 3.7884161472320557, "std": 8.2000150680542, "min": -13.698738098144531, "p10": -6.371321105957031, "median": 2.9685535430908203, "p90": 15.181273269653321, "max": 25.745712280273438, "pos_frac": 0.671875, "sample": [-11.507535934448242, 2.382598876953125, -7.794704437255859, 4.309471130371094, 2.1265640258789062, 16.5411376953125, 9.947463989257812, -4.8004302978515625, -0.26508140563964844, 6.613437652587891, 0.0341796875, 6.5585784912109375, 12.845802307128906, -6.513145446777344, -3.816923141479492, 13.0269775390625, 14.076644897460938, 1.8698005676269531, 2.020109176635742, 1.4896049499511719, 10.9442138671875, 4.2743988037109375, -4.595310211181641, 3.5565185546875, -13.698738098144531, -9.719802856445312, 8.457969665527344, -0.01470947265625, 0.5369529724121094, 3.3112945556640625, 0.069580078125, 17.123886108398438, -3.078763961791992, 10.353897094726562, 7.719444274902344, -0.9305534362792969, 8.95151138305664, -9.287971496582031, 11.060646057128906, 16.310745239257812, 3.7322845458984375, 3.360992431640625, 3.9561920166015625, 13.62744140625, -0.665069580078125, -7.251434326171875, 3.22998046875, -6.040397644042969, 25.745712280273438, 14.786148071289062, 15.35061264038086, 1.8714408874511719, 2.9306869506835938, 14.096054077148438, 16.077346801757812, 0.7712554931640625, -3.6641006469726562, -3.2803993225097656, 3.006420135498047, 17.158660888671875, -1.0173416137695312, 9.878555297851562, -2.95318603515625, -2.738983154296875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000173.npy"}
{"epoch": 0.2615268329554044, "step": 174, "batch_size": 64, "mean": 4.738399982452393, "std": 7.657214164733887, "min": -17.29550552368164, "p10": -4.462965393066406, "median": 4.466153144836426, "p90": 15.102838516235352, "max": 21.028411865234375, "pos_frac": 0.703125, "sample": [5.551551818847656, 11.427597045898438, 8.8662109375, 2.0213050842285156, 6.605121612548828, 4.3247528076171875, 10.119583129882812, 15.090194702148438, -4.845367431640625, 5.50274658203125, 19.295578002929688, 20.99518585205078, 2.9970169067382812, 0.7762489318847656, 6.7830352783203125, -0.4152851104736328, 15.108257293701172, 9.167827606201172, 7.278495788574219, -13.070003509521484, 18.920021057128906, 3.0628738403320312, -0.24433517456054688, 6.43841552734375, 21.028411865234375, -0.7813224792480469, 7.338775634765625, 0.9023895263671875, 13.921188354492188, 8.422874450683594, 0.5815353393554688, 12.993019104003906, 15.54189682006836, -0.340057373046875, 5.691242218017578, -1.9255828857421875, 5.9232025146484375, -6.682830810546875, 5.6961669921875, 13.389190673828125, -0.4497528076171875, 9.235580444335938, -1.3830795288085938, 1.6610488891601562, -5.275054931640625, 14.341102600097656, 3.910888671875, -0.21215057373046875, 15.749374389648438, 4.20921516418457, -5.685813903808594, 2.4415359497070312, 7.6592254638671875, 4.607553482055664, -3.5706939697265625, 2.3916702270507812, 7.52032470703125, 2.5074615478515625, -0.8730697631835938, -17.29550552368164, 6.738471984863281, -2.4314804077148438, -5.3685760498046875, -0.6277923583984375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000174.npy"}
{"epoch": 0.26303854875283444, "step": 175, "batch_size": 64, "mean": 4.225754737854004, "std": 8.187215805053711, "min": -19.593246459960938, "p10": -4.9439952850341795, "median": 3.130678176879883, "p90": 15.46101417541504, "max": 21.138843536376953, "pos_frac": 0.75, "sample": [-9.639278411865234, 10.003644943237305, -4.6699676513671875, 2.9571571350097656, 2.1563034057617188, 13.565170288085938, -4.2213945388793945, 5.22528076171875, 1.280059814453125, 3.210895538330078, 6.0056915283203125, 15.957954406738281, 2.028564453125, 8.145980834960938, -0.7685737609863281, 14.43582534790039, 4.1035003662109375, 18.444740295410156, 4.353771209716797, -4.078052520751953, 2.7563095092773438, 4.760759353637695, 7.1284332275390625, 9.443618774414062, 1.6028289794921875, 1.4595489501953125, 0.24831771850585938, -10.252479553222656, 17.11475372314453, -19.593246459960938, 7.826423645019531, 1.325225830078125, 9.275711059570312, -12.063812255859375, -0.8497467041015625, 1.9106330871582031, 9.943222045898438, 21.138843536376953, 2.3466835021972656, -2.4295654296875, 3.0504608154296875, -10.556564331054688, 1.3357696533203125, 2.115142822265625, -4.0034027099609375, 5.087493896484375, 8.175163269042969, -6.708282470703125, 15.464302062988281, 15.266029357910156, 5.050392150878906, 13.898822784423828, -1.2326583862304688, -0.61431884765625, 6.20257568359375, 8.633872985839844, -5.061435699462891, 15.628616333007812, 1.4017791748046875, 18.862213134765625, 4.823081970214844, 14.371826171875, 2.2143402099609375, 15.45334243774414], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000175.npy"}
{"epoch": 0.26455026455026454, "step": 176, "batch_size": 64, "mean": 2.73274564743042, "std": 8.10048770904541, "min": -14.596939086914062, "p10": -6.185530471801758, "median": 2.2277584075927734, "p90": 13.054252815246583, "max": 26.06732177734375, "pos_frac": 0.625, "sample": [22.992752075195312, -14.596939086914062, 3.4010772705078125, 12.649312973022461, -2.779582977294922, 11.887771606445312, -8.897148132324219, 4.854949951171875, 0.738983154296875, 4.881126403808594, -5.6309051513671875, -5.002967834472656, -3.9779815673828125, 15.159591674804688, -11.000473022460938, 4.246219635009766, 1.4915657043457031, 5.693508148193359, 1.677764892578125, -0.805419921875, -1.749267578125, -6.291835784912109, 3.320920944213867, 18.01470184326172, 14.602378845214844, -0.18609619140625, -4.686363220214844, 6.704322814941406, -1.6266403198242188, 9.663551330566406, -2.6414794921875, -3.2040576934814453, 0.8850936889648438, 3.4838409423828125, 4.168832778930664, 13.227798461914062, 1.3065948486328125, 8.813262939453125, 2.432025909423828, -3.4143600463867188, -3.0032501220703125, 1.9382476806640625, 7.999650955200195, 13.421127319335938, 5.963508605957031, 9.349777221679688, -3.1048736572265625, -10.498374938964844, 9.518753051757812, 3.4965362548828125, -0.432220458984375, 10.807708740234375, 2.7295188903808594, -2.421966552734375, 5.366058349609375, 3.7528858184814453, 12.08160400390625, 1.7805557250976562, -5.9374847412109375, 26.06732177734375, 2.0234909057617188, -8.404190063476562, 6.116851806640625, -13.521957397460938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000176.npy"}
{"epoch": 0.2660619803476946, "step": 177, "batch_size": 64, "mean": 7.25890588760376, "std": 8.064958572387695, "min": -12.845260620117188, "p10": -2.4705984115600583, "median": 6.9391632080078125, "p90": 17.096764373779298, "max": 27.293975830078125, "pos_frac": 0.828125, "sample": [4.185569763183594, -7.013206481933594, 7.1472015380859375, 27.293975830078125, 7.0384674072265625, 4.883903503417969, 1.5442352294921875, 5.656951904296875, 13.772140502929688, 6.8398590087890625, 20.976367950439453, 13.543418884277344, 11.195869445800781, -4.03839111328125, 0.6153144836425781, -2.6546268463134766, 13.579566955566406, -0.6815834045410156, 0.7710914611816406, 10.517313003540039, 7.99359130859375, 0.4661407470703125, -3.041889190673828, 12.554847717285156, 13.650920867919922, 7.290641784667969, 6.055980682373047, 5.782985687255859, 0.49706268310546875, 3.5651473999023438, -0.1466217041015625, 5.519443511962891, 19.000869750976562, -0.532440185546875, 1.6681900024414062, 17.14124298095703, 19.054046630859375, 0.12209320068359375, -2.04119873046875, 8.511383056640625, 4.192684173583984, 4.717254638671875, 12.944461822509766, 12.183563232421875, 16.99298095703125, -8.604179382324219, 1.2950286865234375, -3.5635986328125, 12.117731094360352, 3.5153656005859375, 22.919998168945312, 1.0539169311523438, -12.845260620117188, 16.580684661865234, 9.08294677734375, 14.59765625, 13.77658462524414, 21.02178955078125, 12.6456298828125, 7.978084564208984, 13.246307373046875, 15.735563278198242, 3.0303497314453125, 11.668540954589844], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000177.npy"}
{"epoch": 0.2675736961451247, "step": 178, "batch_size": 64, "mean": 4.633016109466553, "std": 6.791265487670898, "min": -12.851715087890625, "p10": -2.94424819946289, "median": 3.779956817626953, "p90": 13.673771667480471, "max": 23.403762817382812, "pos_frac": 0.78125, "sample": [5.440074920654297, 3.8629989624023438, 7.391349792480469, -0.031768798828125, 5.6521148681640625, 0.9255294799804688, 2.641876220703125, 11.889896392822266, 6.317512512207031, 9.835769653320312, -3.0550155639648438, 12.427833557128906, 10.6917724609375, -6.740436553955078, 5.1430816650390625, 3.9826393127441406, 14.122615814208984, 0.1586761474609375, 4.701507568359375, 1.6847305297851562, 18.911739349365234, 2.172992706298828, 2.81378173828125, 0.373382568359375, 3.5838088989257812, -3.865142822265625, 2.0939407348632812, 23.403762817382812, -2.685791015625, -0.10649871826171875, 3.6969146728515625, 0.3675117492675781, 1.4683837890625, 15.18124008178711, -0.1009674072265625, 8.823944091796875, 8.484466552734375, 0.9041290283203125, 13.912490844726562, 1.3880233764648438, -0.6083526611328125, 7.1914215087890625, 13.11676025390625, 5.9308929443359375, -12.851715087890625, 14.527114868164062, -3.333606719970703, 1.7807350158691406, 16.883712768554688, 0.6395626068115234, 10.530158996582031, 10.542251586914062, 3.561676025390625, 8.768844604492188, 0.30149269104003906, 8.134166717529297, 4.5076751708984375, -1.1221389770507812, -10.634689331054688, -2.4498252868652344, 9.917961120605469, 5.060401916503906, 12.967361450195312, -4.713705062866211], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000178.npy"}
{"epoch": 0.2690854119425548, "step": 179, "batch_size": 64, "mean": 3.95438289642334, "std": 8.341879844665527, "min": -14.738388061523438, "p10": -7.644245910644529, "median": 3.0215377807617188, "p90": 12.858528137207031, "max": 30.448501586914062, "pos_frac": 0.71875, "sample": [2.3035507202148438, -5.96026611328125, 7.76702880859375, 0.413299560546875, 1.8762893676757812, 8.742687225341797, 3.2498302459716797, 12.863357543945312, 16.304092407226562, 7.297763824462891, 10.448659896850586, 1.906545639038086, 14.931167602539062, -13.42641830444336, 9.717327117919922, -9.274463653564453, -14.738388061523438, -9.860088348388672, -2.5543670654296875, 4.549781799316406, -1.794698715209961, 12.365528106689453, 12.840042114257812, 5.0996246337890625, 13.07302474975586, -8.365951538085938, -3.4262771606445312, 12.847259521484375, 3.041440963745117, -0.4520072937011719, 6.324634552001953, 6.045967102050781, 1.8839988708496094, 1.69244384765625, 1.8099365234375, -11.272235870361328, -0.7815399169921875, 19.59478759765625, 11.625350952148438, -5.284294128417969, 8.902912139892578, 3.3533248901367188, 1.9518108367919922, 9.654556274414062, -2.0295562744140625, 12.144424438476562, 18.71239471435547, 7.2051544189453125, 8.274375915527344, 30.448501586914062, 10.026191711425781, 3.0016345977783203, 2.2203826904296875, 0.3595085144042969, -2.3279876708984375, 6.8794403076171875, -0.7819328308105469, 1.786529541015625, 9.235832214355469, 0.704376220703125, -9.997970581054688, 2.3129520416259766, 10.001785278320312, -2.382547378540039], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000179.npy"}
{"epoch": 0.2705971277399849, "step": 180, "batch_size": 64, "mean": 3.8914151191711426, "std": 7.504227161407471, "min": -12.824928283691406, "p10": -3.3849769592285153, "median": 1.6826820373535156, "p90": 14.899643325805666, "max": 25.83453369140625, "pos_frac": 0.640625, "sample": [25.83453369140625, -1.8229522705078125, 5.626142501831055, -1.1261978149414062, -1.322052001953125, 5.642353057861328, -5.970741271972656, 8.555770874023438, 0.27394866943359375, -3.126220703125, 14.206829071044922, 1.0041122436523438, 0.3001556396484375, -2.738414764404297, 9.228988647460938, -1.4065608978271484, -2.134033203125, 5.34454345703125, -2.1784820556640625, -7.50396728515625, 6.421989440917969, 8.738687515258789, 4.585607528686523, 10.946517944335938, 10.918502807617188, 8.935142517089844, -0.9256668090820312, 9.322868347167969, 6.079120635986328, 4.193778991699219, 4.7013397216796875, 10.12472915649414, -2.1337203979492188, 1.6310577392578125, -3.4958724975585938, 18.125213623046875, -1.174386978149414, 12.702301025390625, -2.9117431640625, -6.069450378417969, -1.430032730102539, 9.169788360595703, 16.717864990234375, -0.9326343536376953, -0.8979568481445312, 1.6894302368164062, 18.262359619140625, 0.3741035461425781, 12.826152801513672, 0.27361297607421875, 15.196563720703125, -12.824928283691406, -1.8328475952148438, 2.0299606323242188, 1.1813507080078125, 1.675933837890625, 18.322296142578125, -4.708953857421875, 6.743324279785156, -4.986869812011719, 0.677398681640625, 3.8353538513183594, 17.902938842773438, 2.3825740814208984], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000180.npy"}
{"epoch": 0.272108843537415, "step": 181, "batch_size": 64, "mean": 4.935826301574707, "std": 7.206709861755371, "min": -8.804412841796875, "p10": -3.1779144287109373, "median": 3.253936767578125, "p90": 15.051104736328124, "max": 23.963088989257812, "pos_frac": 0.765625, "sample": [8.127986907958984, 5.072185516357422, -2.6417007446289062, 0.24462318420410156, 3.3236160278320312, 0.13720321655273438, 2.2609481811523438, -8.804412841796875, 3.1842575073242188, -0.6443367004394531, 8.377891540527344, 3.06597900390625, 0.5937423706054688, -3.2664031982421875, 3.7227096557617188, 5.128257751464844, 10.061393737792969, 7.49591064453125, 13.4586181640625, -5.148162841796875, 2.7933197021484375, 8.148193359375, 16.854022979736328, 4.762413024902344, 21.921653747558594, 5.7187042236328125, 2.717559814453125, 23.963088989257812, -4.446746826171875, -5.136680603027344, 2.435333251953125, 1.9125518798828125, 19.859527587890625, -6.93370246887207, 15.041748046875, 2.787181854248047, 5.092916488647461, 22.308578491210938, -0.5752143859863281, -1.856231689453125, 10.705718994140625, 3.928375244140625, 17.985931396484375, 3.6142730712890625, 9.605300903320312, -0.25384521484375, 0.49549293518066406, 4.329231262207031, 3.056243896484375, 0.5127487182617188, -3.1617431640625, 8.957622528076172, 15.05511474609375, 0.11145782470703125, 9.834808349609375, 11.541258811950684, 2.3918304443359375, -0.306671142578125, 11.460813522338867, -3.184844970703125, 3.106739044189453, -1.024810791015625, 11.762725830078125, 4.2505950927734375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000181.npy"}
{"epoch": 0.273620559334845, "step": 182, "batch_size": 64, "mean": 6.296449661254883, "std": 8.558837890625, "min": -11.533248901367188, "p10": -4.593047332763672, "median": 5.722110748291016, "p90": 17.15248336791992, "max": 31.857192993164062, "pos_frac": 0.765625, "sample": [7.949920654296875, -4.686561584472656, -0.4832344055175781, 3.4653167724609375, 8.690826416015625, 17.164222717285156, 5.536663055419922, -4.839801788330078, 0.22888946533203125, 18.68829345703125, 12.262908935546875, 8.337909698486328, 18.582489013671875, 2.4743881225585938, 5.907558441162109, 9.688629150390625, -2.1261062622070312, -4.9709320068359375, 13.583549499511719, 4.408294677734375, 7.789546966552734, 1.026597023010254, 6.777191162109375, 6.35089111328125, 7.696193695068359, -1.0384979248046875, 16.652124404907227, 6.25450325012207, 13.14971923828125, 17.125091552734375, 15.637130737304688, 4.013740539550781, 8.52685546875, 2.3214569091796875, 22.6810302734375, -10.711807250976562, 2.63604736328125, 9.159431457519531, 2.8527069091796875, -0.5307140350341797, 6.9007415771484375, -10.83597183227539, 15.064064025878906, 5.2251434326171875, 14.859397888183594, 31.857192993164062, -0.635345458984375, -2.2393035888671875, 18.232513427734375, 5.4106292724609375, 12.893524169921875, 21.527755737304688, 4.336845397949219, 14.706840515136719, 9.219497680664062, 2.1499996185302734, 2.7482681274414062, 0.4718208312988281, 13.95989990234375, 1.3582744598388672, -1.2084884643554688, -4.374847412109375, -5.354892730712891, -11.533248901367188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000182.npy"}
{"epoch": 0.2751322751322751, "step": 183, "batch_size": 64, "mean": 2.27756667137146, "std": 7.266908168792725, "min": -16.15521240234375, "p10": -6.76564598083496, "median": 1.3488225936889648, "p90": 11.68285827636719, "max": 18.681594848632812, "pos_frac": 0.65625, "sample": [4.811885833740234, 9.581230163574219, 0.4451904296875, -1.3638381958007812, 1.0268230438232422, -3.13897705078125, -7.0719757080078125, 18.681594848632812, -11.325630187988281, -6.0291900634765625, -7.048389434814453, -3.0921630859375, 2.634197235107422, 1.82177734375, 5.903766632080078, 8.041879653930664, -0.46842384338378906, 4.620109558105469, 9.665229797363281, 15.759943008422852, 11.090835571289062, -6.1059112548828125, 10.48736572265625, -0.5024490356445312, 5.7849884033203125, 3.064006805419922, 0.0506439208984375, -1.315093994140625, -14.518356323242188, 0.5171966552734375, 3.951812744140625, -4.2841339111328125, -0.7462806701660156, 4.701576232910156, 0.6365737915039062, 3.3812179565429688, 14.421524047851562, 11.201362609863281, -16.15521240234375, 11.889213562011719, -8.900382995605469, -1.8572616577148438, 16.969505310058594, 0.3934593200683594, 0.7408103942871094, 0.4974803924560547, 4.1861572265625, 0.7859725952148438, -7.702003479003906, -1.0376663208007812, -3.117279052734375, 2.7392349243164062, 1.5705394744873047, 2.7744522094726562, 12.446197509765625, -3.188629150390625, 4.235908508300781, 1.127105712890625, -1.81048583984375, 2.8954925537109375, 4.8551483154296875, 15.954818725585938, 9.16981315612793, 11.02996826171875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000183.npy"}
{"epoch": 0.2766439909297052, "step": 184, "batch_size": 64, "mean": 4.506134986877441, "std": 6.61939811706543, "min": -11.527545928955078, "p10": -2.619543457031249, "median": 3.0513362884521484, "p90": 14.4859432220459, "max": 18.60361099243164, "pos_frac": 0.78125, "sample": [-0.9040374755859375, 7.924736022949219, 13.858818054199219, -6.045352935791016, 10.620201110839844, 0.5068359375, 1.4203071594238281, -3.6824951171875, -0.5819511413574219, -0.6831512451171875, -0.7451248168945312, 7.098136901855469, 1.8830604553222656, 0.5835638046264648, 4.804126739501953, -3.6060523986816406, 5.4618988037109375, 12.796257019042969, 5.739597320556641, 8.742172241210938, -4.163963317871094, 16.373416900634766, 3.5247535705566406, 14.66952133178711, 1.050933837890625, 3.630046844482422, 1.6122627258300781, 6.0614471435546875, 12.500873565673828, 1.9768943786621094, 14.78173828125, 5.670112609863281, 3.63470458984375, -7.7656402587890625, 0.5246448516845703, 0.5621070861816406, 18.60361099243164, 4.799171447753906, -1.5989799499511719, 14.057594299316406, 0.8373298645019531, 8.143379211425781, 9.787437438964844, 1.9033985137939453, 13.946395874023438, 4.954778671264648, 3.2322235107421875, 7.414100646972656, -1.713653564453125, 18.419662475585938, -0.8575305938720703, -3.007781982421875, 2.6757431030273438, 2.8704490661621094, 17.51513671875, 17.179122924804688, 0.5692062377929688, 6.336311340332031, 2.8562393188476562, 7.624916076660156, -11.527545928955078, 2.5726985931396484, 0.0203857421875, 0.943450927734375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000184.npy"}
{"epoch": 0.2781557067271353, "step": 185, "batch_size": 64, "mean": 4.972773551940918, "std": 9.667619705200195, "min": -16.72594451904297, "p10": -7.8503063201904295, "median": 4.343951225280762, "p90": 18.41788482666016, "max": 23.1712646484375, "pos_frac": 0.6875, "sample": [4.526273727416992, 23.1712646484375, 4.592430114746094, 13.848243713378906, 20.214019775390625, -13.800910949707031, 6.715274810791016, 4.019020080566406, 1.4632720947265625, -3.9432525634765625, 17.628040313720703, 19.377029418945312, -6.890861511230469, 0.116302490234375, 16.852577209472656, 1.3907012939453125, 19.42392349243164, 1.4198493957519531, 14.641889572143555, 11.288858413696289, -1.7316780090332031, 11.608200073242188, 3.8754425048828125, 3.6490554809570312, 1.7849597930908203, 15.120025634765625, 5.699430465698242, 12.778243064880371, 4.0343170166015625, 19.871978759765625, -6.403099060058594, -11.153671264648438, 16.949493408203125, 12.541915893554688, -11.096784591674805, 9.653976440429688, 22.286041259765625, 6.5716094970703125, 3.476043701171875, -0.31424522399902344, -7.530670166015625, 8.586418151855469, -1.2319068908691406, 11.947303771972656, 8.500686645507812, -10.765838623046875, 11.234466552734375, -0.6295051574707031, 8.025001525878906, -6.139951705932617, 9.090118408203125, -16.72594451904297, 8.543136596679688, -2.042705535888672, 7.261383056640625, -4.837760925292969, -7.987293243408203, -9.546977996826172, 3.9434585571289062, -2.7480316162109375, -2.1277084350585938, 4.161628723144531, 15.266582489013672, 18.756389617919922], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000185.npy"}
{"epoch": 0.2796674225245654, "step": 186, "batch_size": 64, "mean": 5.201438903808594, "std": 9.866981506347656, "min": -11.248603820800781, "p10": -7.211063003540039, "median": 3.2522945404052734, "p90": 18.244377517700197, "max": 27.882781982421875, "pos_frac": 0.65625, "sample": [16.99005126953125, 10.176605224609375, -3.6789627075195312, 1.7315597534179688, -0.733154296875, -0.3411865234375, 14.937484741210938, -9.5430908203125, -2.2206268310546875, 10.156539916992188, 15.858009338378906, -2.8472633361816406, 4.5330352783203125, 14.058782577514648, 27.882781982421875, 19.85494613647461, -3.9346694946289062, 3.415569305419922, 11.748764038085938, -6.4290771484375, -7.2417755126953125, 14.241081237792969, 5.783298492431641, 23.622085571289062, -11.103469848632812, -5.190528869628906, -7.139400482177734, -2.80096435546875, -5.449554443359375, 2.4726638793945312, 5.324371337890625, 15.491493225097656, 9.9061279296875, -10.601837158203125, -0.5063056945800781, -2.215869903564453, 15.793403625488281, 18.635665893554688, 14.452552795410156, 2.5765838623046875, -1.9718742370605469, -8.163681030273438, 4.7747039794921875, 0.124481201171875, 15.950393676757812, 2.5592041015625, 17.331371307373047, 12.996185302734375, -11.248603820800781, 0.105560302734375, 18.636077880859375, 12.705739974975586, 10.838058471679688, 20.176010131835938, 2.4011917114257812, 4.891700744628906, 25.75390625, 5.024036407470703, -0.6063346862792969, 2.8710861206054688, 0.54669189453125, 3.089019775390625, -10.737876892089844, 7.179317474365234], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000186.npy"}
{"epoch": 0.2811791383219955, "step": 187, "batch_size": 64, "mean": 4.727651119232178, "std": 8.189643859863281, "min": -16.883575439453125, "p10": -5.304922485351561, "median": 4.812360763549805, "p90": 15.85353775024414, "max": 24.166900634765625, "pos_frac": 0.75, "sample": [7.2021484375, -10.369842529296875, 16.015838623046875, 18.726394653320312, 24.166900634765625, 8.44720458984375, 14.212028503417969, 2.9196434020996094, 5.6952056884765625, 7.713802337646484, -5.9893341064453125, 0.9657440185546875, 0.49796295166015625, 5.140838623046875, 5.7985382080078125, 4.737525939941406, 1.5453338623046875, 11.033195495605469, 6.9953460693359375, 13.489009857177734, -5.786224365234375, -16.883575439453125, 8.806411743164062, 0.8245677947998047, 1.3462600708007812, 6.106636047363281, 0.5659217834472656, -0.9483642578125, -4.181884765625, 15.855453491210938, 7.675991058349609, 16.42052459716797, -7.607746124267578, -3.769775390625, -0.832916259765625, 7.9664764404296875, 4.997562408447266, 3.5178146362304688, -2.4656829833984375, 11.271379470825195, -1.2371406555175781, -2.054546356201172, 19.382217407226562, 2.796916961669922, 5.628944396972656, 1.2634506225585938, 8.130363464355469, 3.8614349365234375, 2.3518524169921875, 2.82891845703125, 13.275215148925781, -7.7765350341796875, 12.02374267578125, 20.40350341796875, 2.6557559967041016, 5.553764343261719, 10.2950439453125, 15.849067687988281, 4.887195587158203, 15.734970092773438, -0.8489761352539062, -3.5205631256103516, -13.453826904296875, 2.716583251953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000187.npy"}
{"epoch": 0.28269085411942557, "step": 188, "batch_size": 64, "mean": 4.612437725067139, "std": 7.951073169708252, "min": -12.401412963867188, "p10": -3.58906135559082, "median": 3.7337307929992676, "p90": 17.503802299499515, "max": 22.763595581054688, "pos_frac": 0.671875, "sample": [3.2430648803710938, -2.0937118530273438, 18.703956604003906, -0.3388957977294922, -7.698272705078125, 5.4224700927734375, 8.708793640136719, 9.181438446044922, 6.0560302734375, 4.2552337646484375, -1.9573860168457031, 22.763595581054688, 20.81808090209961, 4.1331634521484375, 7.083610534667969, -3.3497467041015625, 2.3514938354492188, 10.269081115722656, 1.1028594970703125, -3.4081172943115234, 8.965774536132812, 0.0345916748046875, -0.78253173828125, -7.381683349609375, 0.459930419921875, 0.63482666015625, -2.8715057373046875, 0.8052597045898438, 4.8331298828125, 16.33523178100586, 12.236625671386719, -0.5963973999023438, 5.1683349609375, 19.161903381347656, 8.491966247558594, -0.5338935852050781, 13.797775268554688, 6.2040863037109375, -5.687950134277344, -0.45121002197265625, 18.73271942138672, -2.8107452392578125, 5.7843828201293945, 0.6508560180664062, 8.5689697265625, 0.1740264892578125, 7.312313079833984, 20.07659149169922, 5.933937072753906, -0.054302215576171875, 13.6246337890625, -12.401412963867188, -4.364173889160156, 0.3501777648925781, -2.5964508056640625, 8.701560974121094, 3.3342981338500977, -3.3326873779296875, 18.002487182617188, 16.3402042388916, 6.911533355712891, -6.50390625, -3.6666088104248047, 12.356613159179688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000188.npy"}
{"epoch": 0.2842025699168556, "step": 189, "batch_size": 64, "mean": 3.8941988945007324, "std": 9.21821117401123, "min": -23.08417510986328, "p10": -7.246711730957031, "median": 4.106578826904297, "p90": 16.419058418273927, "max": 18.835281372070312, "pos_frac": 0.703125, "sample": [-2.3306884765625, 8.347412109375, 4.25372314453125, -6.113334655761719, -5.1833648681640625, 6.498016357421875, 18.414440155029297, 12.023504257202148, 0.722442626953125, -6.196258544921875, -13.036422729492188, 14.023303985595703, -23.08417510986328, 3.5602035522460938, -6.94427490234375, 1.2194366455078125, 18.200984954833984, 12.411338806152344, -8.76523208618164, 4.5018463134765625, 5.628326416015625, 16.86655044555664, 8.165748596191406, 0.84307861328125, 5.3184356689453125, -4.064516067504883, 4.050743103027344, 0.9437942504882812, 3.354511260986328, 1.2501678466796875, 5.1278228759765625, -8.36151123046875, 1.5891914367675781, 18.246803283691406, 17.310211181640625, 9.002307891845703, 0.39766693115234375, 3.2189254760742188, 12.080142974853516, 11.088050842285156, 11.633110046386719, 4.16241455078125, 8.479286193847656, -1.2248344421386719, 6.742687225341797, 11.11834716796875, -2.2541236877441406, -7.3763275146484375, -1.740692138671875, 2.5665512084960938, -1.892181396484375, 16.460554122924805, 13.921379089355469, 14.789230346679688, 13.983062744140625, 6.1908721923828125, -2.082305908203125, -16.57776641845703, 6.071643829345703, -15.871612548828125, 3.9434051513671875, -1.5508346557617188, 18.835281372070312, 16.322235107421875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000189.npy"}
{"epoch": 0.2857142857142857, "step": 190, "batch_size": 64, "mean": 6.525960922241211, "std": 8.43247127532959, "min": -18.325353622436523, "p10": -5.458893966674804, "median": 5.925701141357422, "p90": 16.36370735168457, "max": 24.123428344726562, "pos_frac": 0.828125, "sample": [12.30223274230957, 12.266357421875, 21.772281646728516, -7.5002899169921875, 4.734615325927734, 7.769111633300781, 4.867271423339844, 3.71380615234375, -18.325353622436523, 24.123428344726562, -5.903312683105469, 5.983818054199219, 9.32038688659668, 10.767570495605469, 4.403043746948242, 2.4221153259277344, 1.250030517578125, 9.3035888671875, 5.6943511962890625, 6.197505950927734, 8.341972351074219, 11.28204345703125, 1.7183837890625, 3.4252853393554688, -4.421916961669922, 12.216156005859375, -7.376472473144531, 1.6769218444824219, 5.827907562255859, 8.631816864013672, 20.067073822021484, -8.232418060302734, 12.328964233398438, 22.5291748046875, 4.8721771240234375, 3.3526573181152344, 17.8529052734375, -0.25830841064453125, 4.788928985595703, 12.264236450195312, 2.5723876953125, 5.867584228515625, 12.5330810546875, 20.870941162109375, 5.4874420166015625, -0.4476318359375, -9.428024291992188, 9.713134765625, -11.074127197265625, 1.1606426239013672, 0.5651969909667969, 10.758340835571289, -1.2803192138671875, 12.521102905273438, 15.437652587890625, 14.255294799804688, 16.398658752441406, 3.3349838256835938, 16.282154083251953, 11.336143493652344, 1.7887802124023438, 14.450347900390625, 8.449066162109375, 10.05862808227539], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000190.npy"}
{"epoch": 0.2872260015117158, "step": 191, "batch_size": 64, "mean": 3.9295034408569336, "std": 8.450912475585938, "min": -16.30194091796875, "p10": -6.158495330810546, "median": 3.8335609436035156, "p90": 14.297635650634765, "max": 23.09337615966797, "pos_frac": 0.703125, "sample": [13.840126037597656, -2.6757068634033203, 2.633758544921875, 22.837905883789062, 23.09337615966797, -1.7313308715820312, 4.277313232421875, 12.897193908691406, 1.4990501403808594, 12.039634704589844, -5.535308837890625, 20.906631469726562, -2.5634918212890625, 4.058902740478516, 16.844558715820312, 5.710323333740234, -16.30194091796875, 1.23980712890625, 11.27923583984375, 8.645278930664062, 7.1640167236328125, 3.43927001953125, 16.350200653076172, 3.966339111328125, -14.090873718261719, 3.7007827758789062, -2.411468505859375, 5.613086700439453, -6.425575256347656, -0.3804512023925781, 1.7383689880371094, 1.0793228149414062, 2.5430984497070312, 6.2218017578125, 5.44378662109375, -10.4720458984375, -1.304229736328125, -7.8490753173828125, 10.734546661376953, 7.8126068115234375, 11.31439208984375, 14.219017028808594, -5.043636322021484, 4.18707275390625, 6.262565612792969, 2.053924560546875, 2.3596038818359375, 10.449310302734375, 0.29517364501953125, 1.0825958251953125, -3.746601104736328, 14.634258270263672, 8.198883056640625, -8.008934020996094, -1.19818115234375, 12.65087890625, 8.057487487792969, -15.05517578125, 6.106800079345703, 0.73602294921875, 5.052833557128906, 14.331329345703125, -0.5915603637695312, -2.7286758422851562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000191.npy"}
{"epoch": 0.2887377173091459, "step": 192, "batch_size": 64, "mean": 3.0240795612335205, "std": 7.678213596343994, "min": -26.302833557128906, "p10": -4.147983169555664, "median": 1.7469596862792969, "p90": 11.596362400054936, "max": 22.987030029296875, "pos_frac": 0.625, "sample": [-26.302833557128906, 4.264293670654297, 0.6607589721679688, 6.504749298095703, 5.545524597167969, 3.3461437225341797, -4.462068557739258, 9.21343994140625, 1.0009841918945312, 9.254653930664062, -1.0101165771484375, -7.027122497558594, -9.461170196533203, 0.9946460723876953, -0.1809539794921875, -2.2634429931640625, 17.794532775878906, -1.906890869140625, 0.2961883544921875, 1.7180252075195312, 8.297340393066406, 22.987030029296875, 2.088777542114258, -1.8859100341796875, 7.397125244140625, 6.318511962890625, -1.7945098876953125, 7.402252197265625, -0.5818023681640625, 6.175697326660156, 12.048606872558594, 6.903526306152344, -3.2047195434570312, 12.983428955078125, -0.9213619232177734, -1.906930923461914, 1.2604904174804688, 1.5230751037597656, -4.987194061279297, 7.8607940673828125, 6.504299163818359, -9.13299560546875, -4.0768280029296875, 22.978500366210938, 5.2855987548828125, 8.31869888305664, -4.178478240966797, 10.541125297546387, 5.573539733886719, -1.2156982421875, 1.7758941650390625, 13.419010162353516, -1.2323036193847656, 5.1758880615234375, -1.9462146759033203, 9.234695434570312, 3.8231964111328125, -0.44097900390625, 10.049530029296875, 16.428665161132812, 3.4481468200683594, -3.4406280517578125, -0.332763671875, 1.0376205444335938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000192.npy"}
{"epoch": 0.29024943310657597, "step": 193, "batch_size": 64, "mean": 6.229554176330566, "std": 9.068280220031738, "min": -10.867141723632812, "p10": -3.714996337890625, "median": 4.4963579177856445, "p90": 19.11042652130127, "max": 25.47167205810547, "pos_frac": 0.6875, "sample": [4.517524719238281, 4.352180480957031, 10.797950744628906, -2.175323486328125, 2.5284805297851562, -3.72552490234375, 4.467353820800781, 1.5083465576171875, -3.6904296875, -0.4399604797363281, 7.6967926025390625, -2.931406021118164, 11.377082824707031, -1.4264678955078125, 21.736614227294922, -4.943572998046875, 12.369754791259766, 3.6406021118164062, 4.833551406860352, 1.334259033203125, -3.8143768310546875, 15.5230712890625, 17.055580139160156, 0.7173042297363281, 6.43603515625, 13.581596374511719, 2.128812789916992, 9.545799255371094, 1.2997970581054688, 14.265575408935547, 20.927539825439453, 6.032051086425781, -4.794620513916016, 22.975814819335938, 8.73223876953125, 6.5152130126953125, -1.6739006042480469, -7.397853851318359, 13.387519836425781, 5.224662780761719, 7.18951416015625, -3.1386566162109375, 18.358688354492188, 25.047409057617188, -1.6612358093261719, -0.783233642578125, 2.748676300048828, -2.0473670959472656, -1.4907455444335938, 14.796760559082031, 1.4249954223632812, -2.1039886474609375, 25.47167205810547, 19.432600021362305, 7.25811767578125, 16.792312622070312, 23.29779052734375, -10.867141723632812, 15.727363586425781, 18.326751708984375, -2.188079833984375, 4.475191116333008, -9.186012268066406, 13.314422607421875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000193.npy"}
{"epoch": 0.29176114890400606, "step": 194, "batch_size": 64, "mean": 2.970244884490967, "std": 9.624678611755371, "min": -20.875625610351562, "p10": -9.597011566162108, "median": 2.341733932495117, "p90": 15.551016998291017, "max": 23.793426513671875, "pos_frac": 0.640625, "sample": [2.447052001953125, -1.1759185791015625, 6.1479644775390625, -4.516756057739258, 3.7290573120117188, -4.096687316894531, -11.217367172241211, 8.321846008300781, -2.1664962768554688, 4.945404052734375, -2.176250457763672, 3.48223876953125, 1.6569137573242188, 16.360809326171875, -17.290924072265625, 0.7471847534179688, 12.874534606933594, -4.392181396484375, 2.2576637268066406, 1.6054611206054688, 9.255905151367188, -3.0600452423095703, 2.054443359375, 8.860084533691406, -8.481071472167969, 2.4258041381835938, -3.568817138671875, 10.681442260742188, 1.183624267578125, 16.384780883789062, -6.834926605224609, 2.5848846435546875, -10.63153076171875, -13.138694763183594, -16.392440795898438, -6.70989990234375, 15.338371276855469, 17.171768188476562, 13.264324188232422, 13.147808074951172, -10.075271606445312, 1.0146484375, 4.798236846923828, -2.827190399169922, 14.695808410644531, 20.625164031982422, 8.89422607421875, -20.875625610351562, 23.793426513671875, 4.869232177734375, 9.847070693969727, -7.025108337402344, -0.97540283203125, 17.84262466430664, 10.8697509765625, 13.468017578125, 12.737241744995117, 10.771942138671875, 0.19271469116210938, 15.64215087890625, 0.7434730529785156, -0.8070640563964844, -4.0675201416015625, 4.863750457763672], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000194.npy"}
{"epoch": 0.29327286470143615, "step": 195, "batch_size": 64, "mean": 4.652583122253418, "std": 8.033452987670898, "min": -14.226333618164062, "p10": -5.008870697021484, "median": 4.457464218139648, "p90": 14.328727340698242, "max": 25.433937072753906, "pos_frac": 0.734375, "sample": [3.1164093017578125, 7.627525329589844, -9.935600280761719, 5.570014953613281, 21.955482482910156, 6.899463653564453, 3.20343017578125, 2.10443115234375, 10.33734130859375, 21.832122802734375, 7.4193878173828125, 4.231376647949219, -2.29583740234375, -11.75323486328125, -4.688323974609375, -14.226333618164062, -1.5917282104492188, 9.307273864746094, 1.2370147705078125, 7.061689376831055, -3.2525177001953125, 0.30789947509765625, 14.30350112915039, 8.130966186523438, 4.31951904296875, 3.7441329956054688, 3.7143478393554688, 6.939674377441406, 5.8444061279296875, 6.726593017578125, 4.595409393310547, 3.4169082641601562, 21.042346954345703, 7.046867370605469, 12.864753723144531, -1.8324165344238281, 4.744901657104492, 6.374969482421875, 3.551025390625, 25.433937072753906, -0.5191879272460938, -3.5530853271484375, 8.573890686035156, -2.026123046875, 4.829658508300781, 7.872562408447266, -5.9662933349609375, 14.33953857421875, -4.3642578125, -5.465747833251953, -5.146247863769531, -3.2281951904296875, 9.250595092773438, 17.73339080810547, 2.6458892822265625, 8.00347900390625, 2.331573486328125, 8.432106018066406, -5.896942138671875, 22.55877685546875, 2.742950439453125, 4.296146392822266, 5.748626708984375, 9.143096923828125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000195.npy"}
{"epoch": 0.2947845804988662, "step": 196, "batch_size": 64, "mean": 4.098524570465088, "std": 8.668481826782227, "min": -21.068702697753906, "p10": -4.756643676757812, "median": 3.848785400390625, "p90": 14.232024383544921, "max": 35.534423828125, "pos_frac": 0.734375, "sample": [-0.23406982421875, -21.068702697753906, 0.06846237182617188, -2.4317169189453125, 5.236948013305664, -0.9264774322509766, 4.371063232421875, 3.393768310546875, 6.6239776611328125, -7.686065673828125, -3.42498779296875, 20.015228271484375, 5.972206115722656, 3.1122894287109375, 5.19244384765625, 1.5015411376953125, -9.79085922241211, 22.776676177978516, -9.551849365234375, 16.360977172851562, -4.9908905029296875, 8.913749694824219, 17.86663055419922, 1.3515167236328125, -5.164085388183594, 8.937179565429688, 9.665267944335938, 6.572395324707031, 4.303802490234375, 35.534423828125, 1.0375251770019531, 8.056854248046875, 0.314910888671875, -2.8354568481445312, 1.8244209289550781, 14.237052917480469, -2.735921859741211, 9.155315399169922, 1.0866012573242188, 2.173583984375, 10.573883056640625, 5.2314453125, 8.504467010498047, -2.061676025390625, 15.069587707519531, -14.650482177734375, 0.09409332275390625, 2.0820770263671875, 8.171615600585938, 8.347702026367188, 4.4759063720703125, 11.344829559326172, -0.58538818359375, -2.0237808227539062, 2.251680374145508, 6.788970947265625, 7.578403472900391, 11.0047607421875, 1.7140007019042969, -4.2100677490234375, 6.537395477294922, 0.9291458129882812, 6.1009674072265625, 14.220291137695312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000196.npy"}
{"epoch": 0.2962962962962963, "step": 197, "batch_size": 64, "mean": 5.416136741638184, "std": 10.946637153625488, "min": -26.186843872070312, "p10": -7.985976409912107, "median": 6.525389671325684, "p90": 18.016596603393555, "max": 27.835853576660156, "pos_frac": 0.71875, "sample": [-3.7846908569335938, 25.88219451904297, 4.827205657958984, 2.1741485595703125, 11.271636962890625, 4.422950744628906, -13.64544677734375, -16.273727416992188, 7.763641357421875, 9.666244506835938, 0.7201690673828125, 20.783344268798828, -15.098190307617188, -4.960468292236328, 7.002410888671875, 10.901992797851562, 10.081680297851562, 4.052009582519531, 17.399871826171875, 20.991455078125, -2.1600799560546875, 5.070762634277344, 10.27020263671875, 9.016355514526367, 18.118507385253906, -0.37276458740234375, 15.022354125976562, -5.30821418762207, 5.550502777099609, 8.541427612304688, 16.710494995117188, 22.647308349609375, -26.186843872070312, 15.75668716430664, -4.345592498779297, -15.396438598632812, 4.068672180175781, 9.116622924804688, 17.607437133789062, -1.6178817749023438, -2.1015548706054688, 8.890941619873047, 17.778804779052734, 1.904754638671875, 1.5433120727539062, 8.884384155273438, -4.909454345703125, 0.3773460388183594, 7.1279449462890625, 16.881446838378906, -9.133588790893555, 0.5791778564453125, 21.897789001464844, 10.330423355102539, -4.60546875, -14.531570434570312, 27.835853576660156, 10.807395935058594, 9.629230499267578, 10.282485961914062, 2.5932083129882812, -1.4919281005859375, 6.048368453979492, 13.725479125976562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000197.npy"}
{"epoch": 0.29780801209372637, "step": 198, "batch_size": 64, "mean": 5.6805620193481445, "std": 8.66405963897705, "min": -16.69189453125, "p10": -5.923395347595214, "median": 5.908120155334473, "p90": 16.881035614013673, "max": 24.29905891418457, "pos_frac": 0.78125, "sample": [-6.705268859863281, 14.316070556640625, 1.8916702270507812, 1.7770004272460938, 3.7773056030273438, 17.29589080810547, 1.7184333801269531, 16.720962524414062, 4.003440856933594, 14.623497009277344, 9.296279907226562, 8.782546997070312, 2.546449661254883, -5.162666320800781, 17.160865783691406, 16.94963836669922, 1.45184326171875, 13.632568359375, 1.5745391845703125, 1.3908615112304688, 7.639823913574219, 8.731456756591797, 11.692150115966797, -10.483726501464844, 14.082954406738281, -16.69189453125, -8.8604736328125, -1.2329597473144531, 18.94959259033203, 11.874191284179688, 9.932960510253906, 11.42608642578125, 1.1120147705078125, -14.109832763671875, 7.7048797607421875, 1.98876953125, 0.12729644775390625, -2.832324981689453, 5.6343231201171875, 5.919681549072266, 12.621559143066406, 19.37042999267578, 15.967016220092773, 6.411079406738281, 7.2001190185546875, 4.712646484375, -6.405422210693359, 21.349178314208984, 24.29905891418457, -2.0128860473632812, -3.1859283447265625, 3.6347122192382812, 13.795326232910156, 13.711099624633789, -1.6079177856445312, 1.4933395385742188, 8.904571533203125, -4.723930358886719, -6.249422073364258, 5.89655876159668, 8.961456298828125, 8.6739501953125, 8.304573059082031, 2.7878990173339844], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000198.npy"}
{"epoch": 0.29931972789115646, "step": 199, "batch_size": 64, "mean": 4.949808120727539, "std": 10.646992683410645, "min": -21.276702880859375, "p10": -6.071768188476561, "median": 5.257386207580566, "p90": 17.088740158081055, "max": 25.9998779296875, "pos_frac": 0.640625, "sample": [15.904422760009766, 12.793052673339844, 8.266265869140625, 9.583457946777344, 0.6402816772460938, -12.441455841064453, 25.9998779296875, 17.34195327758789, 7.45477294921875, 12.324539184570312, 8.108261108398438, -4.3900604248046875, 12.870361328125, 14.742446899414062, 9.220832824707031, -1.191009521484375, 18.95263671875, 11.897415161132812, 7.878662109375, 9.282005310058594, -0.39971923828125, -14.710397720336914, 16.019359588623047, 16.497909545898438, 0.033885955810546875, 2.1967926025390625, 1.0224380493164062, -4.763214111328125, 4.382389068603516, 13.540206909179688, 4.11181640625, 6.132383346557617, 13.36428451538086, -3.5502471923828125, -0.3171520233154297, -2.4355316162109375, 24.1572265625, -0.5311279296875, -6.599693298339844, 9.32501220703125, -20.75020980834961, 9.568634033203125, -0.6047592163085938, 15.929534912109375, -13.168601989746094, 6.4498443603515625, 13.280975341796875, 1.06488037109375, 0.35079193115234375, -0.37209320068359375, -2.8788089752197266, -15.85626220703125, -4.839942932128906, 20.143165588378906, -21.276702880859375, -2.6514968872070312, 13.396781921386719, -3.650787353515625, 13.566802978515625, -2.099384307861328, 21.67864990234375, -0.12844085693359375, 24.173233032226562, 2.7465591430664062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000199.npy"}
{"epoch": 0.30083144368858655, "step": 200, "batch_size": 64, "mean": 5.892240524291992, "std": 11.343127250671387, "min": -24.444595336914062, "p10": -6.798461532592772, "median": 4.892614364624023, "p90": 20.641191864013674, "max": 32.447662353515625, "pos_frac": 0.734375, "sample": [4.537298202514648, 6.429485321044922, 20.7294921875, 13.723617553710938, 14.368057250976562, 1.010549545288086, 25.171043395996094, 3.98455810546875, 10.0531005859375, -2.0240631103515625, 0.1795654296875, 2.177276611328125, -5.752277374267578, 23.410659790039062, 2.4297637939453125, -2.3991737365722656, 0.35379791259765625, -0.3779563903808594, 14.448257446289062, -11.940322875976562, 1.5058212280273438, -1.9188232421875, 16.750350952148438, 21.080501556396484, 3.395570755004883, 5.3771514892578125, 13.236576080322266, -3.9852066040039062, 6.733551025390625, -14.439903259277344, 18.929115295410156, 14.464370727539062, -0.9486007690429688, 25.044509887695312, 19.03119659423828, -10.93408203125, -24.444595336914062, 10.956792831420898, 16.190017700195312, 20.06841278076172, 13.168083190917969, 5.028102874755859, 32.447662353515625, 13.7052001953125, 0.8249969482421875, -19.802597045898438, -0.7813796997070312, 1.76873779296875, 8.08660888671875, 22.2752685546875, 5.2942352294921875, 7.144130706787109, 2.329120635986328, -1.0132331848144531, 6.641448974609375, 11.600631713867188, -3.9684410095214844, 2.6084747314453125, 2.3070945739746094, 10.258232116699219, -7.246826171875, -17.369924545288086, 20.435157775878906, 4.7571258544921875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000200.npy"}
{"epoch": 0.30234315948601664, "step": 201, "batch_size": 64, "mean": 6.098217964172363, "std": 9.234031677246094, "min": -16.298187255859375, "p10": -4.234371185302734, "median": 5.156984329223633, "p90": 19.422109222412114, "max": 25.150604248046875, "pos_frac": 0.6875, "sample": [1.6995086669921875, 4.6692352294921875, 2.7761383056640625, 5.202312469482422, -2.088306427001953, 5.2558441162109375, 16.361587524414062, -0.035190582275390625, 3.064455032348633, -0.6938323974609375, 8.588272094726562, 9.452709197998047, 20.041950225830078, 7.27496337890625, -1.55804443359375, -6.885162353515625, 20.70263671875, 13.781547546386719, 0.8952178955078125, 18.39248275756836, -7.767387390136719, 15.332717895507812, 5.111656188964844, -0.0539093017578125, -2.2840099334716797, 23.287914276123047, 4.356328964233398, 13.746345520019531, -4.553943634033203, 18.62128448486328, -0.2722511291503906, 8.852401733398438, 5.955394744873047, -3.827526092529297, 20.309833526611328, 25.150604248046875, 2.0010986328125, -16.298187255859375, 13.724151611328125, 7.172874450683594, 16.46163558959961, 10.157451629638672, -3.1358118057250977, 6.658481597900391, -9.360044479370117, 16.36895751953125, 8.606956481933594, -0.07037353515625, -4.408733367919922, 19.785667419433594, -2.043886184692383, -1.7756195068359375, 0.7105255126953125, 8.33523941040039, 13.544351577758789, 17.889602661132812, 15.188926696777344, -0.5150833129882812, 3.469268798828125, 1.6550445556640625, 19.76531982421875, 7.937652587890625, 2.2883453369140625, -12.691650390625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000201.npy"}
{"epoch": 0.30385487528344673, "step": 202, "batch_size": 64, "mean": 5.95410680770874, "std": 9.561747550964355, "min": -15.643218994140625, "p10": -4.922231292724608, "median": 6.376410007476807, "p90": 17.084454345703126, "max": 30.49652862548828, "pos_frac": 0.6875, "sample": [-3.370403289794922, 13.275604248046875, 25.819992065429688, 16.716766357421875, 12.90005874633789, -0.05234527587890625, 7.024883270263672, 15.271186828613281, 12.211606979370117, 4.365623474121094, 9.850452423095703, 10.489013671875, 16.300704956054688, 1.9548072814941406, -1.2235679626464844, 21.164344787597656, 11.745498657226562, -3.5151519775390625, 8.507156372070312, -5.7726593017578125, 10.244087219238281, 6.479763031005859, 2.8297119140625, -5.330421447753906, -12.813411712646484, 7.7847900390625, 4.515308380126953, 30.49652862548828, -0.223907470703125, 14.616767883300781, -2.9605712890625, 3.0525779724121094, 11.12955093383789, 0.3160552978515625, 0.2580718994140625, -15.643218994140625, 18.224281311035156, 6.273056983947754, 9.60098648071289, -11.666950225830078, -1.2496337890625, 3.9917755126953125, -0.25494384765625, 17.242034912109375, 13.801803588867188, 1.4570693969726562, 16.31591796875, 2.588176727294922, 11.545799255371094, -3.96978759765625, 0.64776611328125, -2.770793914794922, -0.7821178436279297, -2.3188304901123047, -6.1908111572265625, 9.058685302734375, 23.381866455078125, 11.521957397460938, -0.7334194183349609, 17.531925201416016, -13.831016540527344, 14.457237243652344, 8.835578918457031, 9.939998626708984], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000202.npy"}
{"epoch": 0.30536659108087677, "step": 203, "batch_size": 64, "mean": 5.6292877197265625, "std": 9.066327095031738, "min": -20.327224731445312, "p10": -2.7642753601074217, "median": 4.299568176269531, "p90": 17.663726806640625, "max": 32.52217102050781, "pos_frac": 0.75, "sample": [4.860504150390625, 0.436767578125, -13.690879821777344, -0.091827392578125, -0.3748016357421875, 10.093767166137695, -0.12006378173828125, 14.28863525390625, 3.0480194091796875, -3.6665420532226562, 24.998443603515625, 6.330287933349609, 0.6517181396484375, 32.52217102050781, 0.07775115966796875, 19.000747680664062, 17.372238159179688, -0.0926513671875, 15.96677017211914, 13.893035888671875, -20.327224731445312, -2.1226043701171875, 8.123237609863281, -2.953907012939453, -1.5017852783203125, -2.796600341796875, 2.3406143188476562, 10.509918212890625, 5.166473388671875, 6.462226867675781, 9.770149230957031, 5.7865753173828125, 1.7649765014648438, 2.7931175231933594, 15.174224853515625, 1.2717971801757812, 20.123580932617188, 9.486743927001953, 26.77630615234375, 4.421142578125, 8.604621887207031, 7.725677490234375, 1.33465576171875, 4.358253479003906, 17.788650512695312, 2.2136459350585938, 6.015472412109375, 4.240882873535156, -2.6888504028320312, 24.08154296875, 12.374616622924805, -1.7531318664550781, 0.2693939208984375, 4.826366424560547, 1.627288818359375, -4.496692657470703, 1.8124008178710938, 0.4776878356933594, 11.323455810546875, 4.620548248291016, -1.4741020202636719, -3.0141143798828125, 2.6804237365722656, 11.552675247192383], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000203.npy"}
{"epoch": 0.30687830687830686, "step": 204, "batch_size": 64, "mean": 6.280874252319336, "std": 9.368861198425293, "min": -17.10928726196289, "p10": -3.4342041015625, "median": 5.426334381103516, "p90": 20.381779289245607, "max": 29.960243225097656, "pos_frac": 0.75, "sample": [27.62834930419922, 15.029205322265625, -8.108510971069336, 6.833074569702148, 0.2592754364013672, 29.960243225097656, 5.531707763671875, 12.267133712768555, 0.8769989013671875, 11.216094970703125, 19.595046997070312, -3.2834930419921875, -5.895732879638672, 10.655609130859375, 5.553886413574219, -1.415374755859375, 9.740909576416016, 6.858432769775391, -7.93096923828125, -0.7408828735351562, 3.519001007080078, 10.785568237304688, 5.173370361328125, 15.176490783691406, -0.29692840576171875, -2.3695831298828125, 15.33154296875, 7.284954071044922, 3.1451683044433594, 6.726566314697266, 15.480772018432617, -5.353462219238281, 8.447917938232422, 10.515419006347656, 6.875082015991211, -3.4987945556640625, 6.207061767578125, 22.359027862548828, 2.2644309997558594, 8.287910461425781, 5.1932830810546875, 2.1862106323242188, 17.279769897460938, 20.718950271606445, -9.10494613647461, 0.45088958740234375, 5.160621643066406, 2.649717330932617, 22.014633178710938, 0.870452880859375, 16.405372619628906, 23.110794067382812, 2.3391571044921875, 10.591758728027344, 0.4856414794921875, 23.060012817382812, -2.3967819213867188, 1.3390922546386719, -0.21778106689453125, 5.320960998535156, -17.10928726196289, -2.2330780029296875, 6.276458740234375, -3.10845947265625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000204.npy"}
{"epoch": 0.30839002267573695, "step": 205, "batch_size": 64, "mean": 5.048940658569336, "std": 10.87977409362793, "min": -17.290546417236328, "p10": -7.353438186645507, "median": 3.4586868286132812, "p90": 20.796919822692878, "max": 30.52642822265625, "pos_frac": 0.65625, "sample": [7.16522216796875, 0.6447906494140625, -0.9403877258300781, -7.640972137451172, 4.430179595947266, 24.239810943603516, 0.17757034301757812, 18.328323364257812, -0.556365966796875, 24.527984619140625, 3.937164306640625, 4.915489196777344, 30.52642822265625, 13.113016128540039, 21.58603858947754, -2.739917755126953, 8.468788146972656, -5.4326019287109375, 3.14013671875, 2.2342605590820312, 0.7210178375244141, 9.536170959472656, 0.8610763549804688, 4.0374755859375, 9.90838623046875, -3.4236221313476562, -2.9503555297851562, -2.5613937377929688, -3.868316650390625, 8.752700805664062, -1.3326568603515625, -2.9992332458496094, -6.682525634765625, -0.03018951416015625, -10.496810913085938, 6.82672119140625, 5.886505126953125, 2.6230621337890625, -5.846290588378906, 18.14134979248047, -13.865753173828125, 18.955642700195312, 16.829151153564453, 17.568069458007812, 29.233993530273438, 15.749298095703125, 2.7579193115234375, 11.084556579589844, 23.9364013671875, -3.0571060180664062, -9.94504165649414, 10.444725036621094, 0.66217041015625, 14.337028503417969, -11.0228271484375, 22.00897216796875, 9.18777847290039, -11.975074768066406, 16.75482940673828, -6.584064483642578, 5.5901641845703125, -17.290546417236328, 3.7772369384765625, 0.7666492462158203], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000205.npy"}
{"epoch": 0.30990173847316704, "step": 206, "batch_size": 64, "mean": 3.6308653354644775, "std": 9.90750789642334, "min": -20.42193603515625, "p10": -5.388189315795898, "median": 1.6329784393310547, "p90": 15.768256378173831, "max": 30.021759033203125, "pos_frac": 0.65625, "sample": [-1.970712661743164, -9.964645385742188, 4.270668029785156, -5.5597991943359375, -6.52610969543457, -2.0279693603515625, 5.8736114501953125, 2.4707374572753906, 13.629825592041016, 3.9429092407226562, 1.0050125122070312, 25.010177612304688, 7.409538269042969, -3.634206771850586, 0.03448486328125, 22.15447998046875, -0.6579933166503906, 6.665504455566406, -14.68389892578125, 0.6370887756347656, 30.021759033203125, 0.4243011474609375, 0.6770477294921875, 0.114166259765625, -0.5248336791992188, 14.805999755859375, 8.444826126098633, -16.520519256591797, 10.452644348144531, 13.598876953125, 15.126014709472656, 12.359443664550781, 6.573062896728516, 14.244720458984375, 4.310081481933594, -20.42193603515625, -18.4775390625, -4.5010528564453125, -2.469696044921875, 2.222747802734375, 5.196434020996094, 1.3005905151367188, 11.574996948242188, -0.89605712890625, -4.987766265869141, 1.9653663635253906, 21.63968276977539, 14.81475830078125, -3.3455352783203125, -4.041229248046875, 7.761566162109375, 1.2066192626953125, 8.308265686035156, 0.033535003662109375, 17.398330688476562, 17.806930541992188, -0.32471466064453125, 4.0721282958984375, -4.682594299316406, 16.043502807617188, -4.399993896484375, 8.1751708984375, -1.4343185424804688, 0.6508941650390625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000206.npy"}
{"epoch": 0.31141345427059713, "step": 207, "batch_size": 64, "mean": 6.224671840667725, "std": 10.14179801940918, "min": -16.65233612060547, "p10": -4.279101181030273, "median": 4.264848709106445, "p90": 17.94134330749512, "max": 40.05058288574219, "pos_frac": 0.734375, "sample": [40.05058288574219, -1.7485580444335938, 3.7532196044921875, 4.4552154541015625, 8.299312591552734, 7.525676727294922, 4.30645751953125, 1.4846839904785156, 2.3057098388671875, 10.750625610351562, 2.847339630126953, 14.5423583984375, 1.3037948608398438, 14.590522766113281, -3.5814075469970703, -4.861248016357422, -6.1936798095703125, 17.308685302734375, 11.203290939331055, 13.596275329589844, 9.668617248535156, 4.587821960449219, 2.424266815185547, -4.5191650390625, -1.594573974609375, 0.6127090454101562, -0.7358417510986328, 23.627685546875, -2.583587646484375, 1.0227203369140625, 17.212085723876953, 6.520633697509766, 34.099609375, 0.4445037841796875, 2.7830047607421875, 4.734687805175781, -6.841949462890625, 15.15863037109375, -0.720001220703125, -0.23686504364013672, 11.720001220703125, 4.500816345214844, 4.223239898681641, 4.389827728271484, 14.707145690917969, 16.820327758789062, 1.8949127197265625, -0.6834602355957031, 18.212482452392578, 8.465980529785156, -16.65233612060547, 0.7308502197265625, 6.365142822265625, 6.814643859863281, 1.3289794921875, 24.336959838867188, 21.796661376953125, -3.718952178955078, 11.841873168945312, -3.1820716857910156, -6.098964691162109, 2.9139556884765625, -6.857208251953125, 26.904342651367188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000207.npy"}
{"epoch": 0.3129251700680272, "step": 208, "batch_size": 64, "mean": 3.379854917526245, "std": 10.199172973632812, "min": -17.608673095703125, "p10": -10.60871238708496, "median": 1.0149455070495605, "p90": 17.071334075927734, "max": 24.443817138671875, "pos_frac": 0.59375, "sample": [-7.714496612548828, -0.48023223876953125, 10.221721649169922, -4.498237609863281, 18.62481689453125, 15.69339370727539, 1.183882713317871, 7.0907440185546875, -17.608673095703125, -15.16043472290039, -0.35266876220703125, 0.84600830078125, 10.220527648925781, 7.994392395019531, 7.070960998535156, 7.699235916137695, -0.2607917785644531, 5.99365234375, -1.8853302001953125, -12.485771179199219, 9.80206298828125, -11.010421752929688, 0.5104904174804688, -0.0505828857421875, 7.275449752807617, 14.005434036254883, 12.677547454833984, -3.0430984497070312, -0.3630561828613281, 16.933616638183594, -3.914216995239258, 5.857933044433594, -9.278434753417969, 0.7671661376953125, -8.520553588867188, 13.632453918457031, 22.080184936523438, -1.652252197265625, -1.3180694580078125, -2.250027656555176, 24.443817138671875, 8.068572998046875, 3.0111465454101562, -15.48978042602539, 14.259201049804688, 11.02215576171875, 20.064971923828125, 0.176483154296875, 12.653068542480469, 6.67988395690918, -11.303474426269531, 19.757896423339844, -0.6564254760742188, 18.106468200683594, -4.984016418457031, 8.354049682617188, 3.3636322021484375, 0.7013397216796875, -9.671390533447266, -2.9419174194335938, -16.942626953125, 0.3405303955078125, 15.83245849609375, 17.130355834960938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000208.npy"}
{"epoch": 0.3144368858654573, "step": 209, "batch_size": 64, "mean": 7.482547760009766, "std": 10.869872093200684, "min": -11.3287353515625, "p10": -6.695505142211914, "median": 5.510795593261719, "p90": 20.055722808837892, "max": 36.753265380859375, "pos_frac": 0.75, "sample": [20.3009033203125, 17.368776321411133, -1.4174423217773438, 16.445796966552734, -6.3913726806640625, 29.376983642578125, 8.663421630859375, 16.405540466308594, 1.4441604614257812, 36.753265380859375, 1.5045928955078125, -4.5761871337890625, -11.3287353515625, 17.744766235351562, -9.012130737304688, 2.363433837890625, -0.9560775756835938, 2.1153602600097656, 0.10836410522460938, 32.05308532714844, 3.2344093322753906, -6.825847625732422, -6.051979064941406, 10.83251953125, 14.5098876953125, 0.8079071044921875, 0.8206253051757812, 12.094375610351562, 11.922149658203125, 16.483078002929688, 13.427034378051758, 7.2403564453125, 8.689247131347656, 13.243827819824219, 18.334243774414062, 6.077152252197266, 23.596282958984375, 4.3827667236328125, -8.940788269042969, 8.46160888671875, -1.79559326171875, -0.4799365997314453, 0.22187042236328125, -8.771591186523438, 19.220947265625, 4.948356628417969, 19.48363494873047, 14.294960021972656, 6.073234558105469, 17.420337677001953, 1.9083385467529297, -7.03904914855957, -1.4646244049072266, 3.5263214111328125, -1.9771728515625, 3.313383102416992, 25.507850646972656, -7.08465576171875, 10.402847290039062, 12.531391143798828, 24.270004272460938, 18.746501922607422, 4.088958740234375, 0.23136520385742188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000209.npy"}
{"epoch": 0.31594860166288735, "step": 210, "batch_size": 64, "mean": 7.500221252441406, "std": 9.973087310791016, "min": -11.40093994140625, "p10": -6.776323318481445, "median": 7.718616485595703, "p90": 21.25396766662598, "max": 26.4197998046875, "pos_frac": 0.703125, "sample": [9.74029541015625, -0.000438690185546875, 3.17608642578125, 10.328750610351562, -0.20887184143066406, 13.095756530761719, 24.482803344726562, 11.827247619628906, 20.148780822753906, 10.914093017578125, 20.82019805908203, 26.4197998046875, 10.541038513183594, 11.832992553710938, 20.71295166015625, -0.9369277954101562, -9.345417022705078, -11.40093994140625, 23.657325744628906, -6.893642425537109, 9.777214050292969, 16.738815307617188, 7.930267333984375, -2.0097503662109375, 23.014541625976562, -0.405792236328125, -1.9624805450439453, 9.854484558105469, 16.482337951660156, -8.371452331542969, 14.008560180664062, 6.425148010253906, 7.506965637207031, 4.3447265625, 4.995429992675781, -1.1898460388183594, 0.13457107543945312, -2.9233779907226562, -11.346931457519531, 21.878995895385742, -6.5025787353515625, 5.4263916015625, 10.374221801757812, 2.3624954223632812, 20.364299774169922, -0.42731475830078125, 16.155418395996094, -1.4020652770996094, 13.990509033203125, 5.8796539306640625, 5.143138885498047, 12.80392074584961, 9.006156921386719, -8.553966522216797, 0.492706298828125, 23.72107696533203, 21.439868927001953, 5.9687042236328125, 6.4555511474609375, -4.697132110595703, 14.717727661132812, -7.148399353027344, 12.70849609375, 17.940982818603516], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000210.npy"}
{"epoch": 0.31746031746031744, "step": 211, "batch_size": 64, "mean": 6.012731552124023, "std": 10.129266738891602, "min": -14.712753295898438, "p10": -7.152262115478513, "median": 5.491018295288086, "p90": 18.61661605834961, "max": 35.705291748046875, "pos_frac": 0.71875, "sample": [2.1266746520996094, -13.907842636108398, 12.143600463867188, 7.764629364013672, -3.14739990234375, -0.009403228759765625, -9.670318603515625, 10.975234985351562, 8.3939208984375, 1.9594230651855469, 3.6467132568359375, 10.21860122680664, 0.4336395263671875, 18.46466064453125, 3.811431884765625, 7.398103713989258, 1.6019630432128906, -3.9484214782714844, 9.007591247558594, -3.268218994140625, 16.371410369873047, -0.96728515625, 13.490707397460938, -8.498306274414062, -4.106525421142578, -4.165153503417969, 2.4127426147460938, 1.6750946044921875, 9.889503479003906, 9.765705108642578, 5.794269561767578, -7.9925994873046875, 1.9735946655273438, -0.7350311279296875, 16.1697998046875, 17.671215057373047, 4.982078552246094, 9.782302856445312, -8.740768432617188, 14.203960418701172, 6.112010955810547, 35.705291748046875, 27.226287841796875, 11.748943328857422, 2.7862091064453125, 18.681739807128906, 8.710617065429688, 22.603759765625, -5.191474914550781, 22.448501586914062, -14.712753295898438, 5.187767028808594, 11.695098876953125, 14.50063705444336, 21.46148681640625, 12.357137680053711, 15.069419860839844, 8.402374267578125, -3.3291778564453125, 20.457382202148438, -10.11031723022461, -1.882720947265625, 0.5782470703125, 1.3370513916015625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000211.npy"}
{"epoch": 0.31897203325774753, "step": 212, "batch_size": 64, "mean": 7.620143890380859, "std": 11.514043807983398, "min": -22.28521728515625, "p10": -4.0480510711669915, "median": 7.000949859619141, "p90": 21.932725524902345, "max": 31.40972900390625, "pos_frac": 0.75, "sample": [29.204612731933594, 14.270042419433594, 13.237582206726074, 15.100872039794922, 1.2793655395507812, 2.5444259643554688, 27.80992889404297, 1.8464469909667969, -7.9639434814453125, 20.165206909179688, -3.0778846740722656, -22.28521728515625, 21.953407287597656, 19.460662841796875, 12.953811645507812, 6.678619384765625, 21.88446807861328, 8.330001831054688, -18.967796325683594, 11.5074462890625, 7.323280334472656, 0.13321208953857422, 3.0949535369873047, 5.736932754516602, -8.880302429199219, 0.886810302734375, 12.961456298828125, -2.7541427612304688, 15.059329986572266, 31.40972900390625, 12.387252807617188, 7.615257263183594, 4.0771026611328125, 4.1437530517578125, 17.814163208007812, 12.489761352539062, -2.5135154724121094, 15.1832275390625, -0.287567138671875, 9.696609497070312, 12.938224792480469, -1.390289306640625, 15.053627014160156, -2.4716720581054688, 2.5214309692382812, 11.152027130126953, -0.9018917083740234, 28.0093994140625, -12.252784729003906, -15.418731689453125, 5.268156051635742, 3.50494384765625, 6.115974426269531, -2.667694091796875, 20.235122680664062, 26.31792449951172, 22.696609497070312, 20.081283569335938, 17.429855346679688, 5.7267303466796875, -4.463836669921875, 2.02716064453125, -2.065916061401367, 8.734184265136719], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000212.npy"}
{"epoch": 0.3204837490551776, "step": 213, "batch_size": 64, "mean": 5.9808807373046875, "std": 11.732393264770508, "min": -26.924854278564453, "p10": -6.982440185546875, "median": 4.365625381469727, "p90": 22.468903350830082, "max": 31.978668212890625, "pos_frac": 0.671875, "sample": [15.109378814697266, -3.72930908203125, 6.983390808105469, 27.762771606445312, 7.891254425048828, 3.5998401641845703, -4.0988922119140625, -2.3084259033203125, 6.139251708984375, 18.351078033447266, 7.3965911865234375, 2.1498641967773438, -6.1163482666015625, -2.0918188095092773, -26.924854278564453, -7.04156494140625, 9.202770233154297, 16.92596435546875, -4.437408447265625, 7.93212890625, 6.7531280517578125, -15.351341247558594, 31.255813598632812, 28.46826171875, 23.11846923828125, 4.138465881347656, -3.178272247314453, -6.844482421875, -0.6956825256347656, 1.4627799987792969, -4.839303970336914, -7.807098388671875, -0.7182998657226562, 3.508556365966797, 12.790451049804688, -8.121734619140625, 1.668715476989746, 8.07174301147461, 22.96173858642578, 5.862396240234375, 4.036075592041016, 8.339263916015625, 20.972328186035156, 0.8461723327636719, 21.318954467773438, 0.3190765380859375, 4.592784881591797, 10.792877197265625, 31.01062774658203, -1.3461074829101562, -8.340682983398438, 7.926235198974609, -3.976348876953125, 6.3606109619140625, -2.5319061279296875, 3.455841064453125, -7.495391845703125, 31.978668212890625, 19.723617553710938, 3.6538543701171875, 8.916427612304688, 13.882320404052734, 14.550384521484375, 18.590694427490234], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000213.npy"}
{"epoch": 0.3219954648526077, "step": 214, "batch_size": 64, "mean": 5.750315189361572, "std": 10.73340892791748, "min": -16.07010269165039, "p10": -8.057003784179686, "median": 4.282692909240723, "p90": 19.006333923339845, "max": 39.210296630859375, "pos_frac": 0.703125, "sample": [7.185670852661133, 1.3844451904296875, -1.081817626953125, 0.5253963470458984, -0.9634284973144531, 11.286556243896484, 4.318153381347656, 4.247232437133789, 4.8611297607421875, 11.6676025390625, 22.40625762939453, 4.904535293579102, -0.9408836364746094, -14.153121948242188, 10.477203369140625, 14.284103393554688, 10.39398193359375, -10.27789306640625, 6.01153564453125, 14.917989730834961, 2.6910972595214844, 11.63607406616211, 4.144371032714844, -0.58880615234375, 2.408235549926758, 17.691638946533203, 7.604408264160156, 4.188535690307617, 14.369285583496094, 8.582550048828125, 6.8640594482421875, -0.050933837890625, -8.463827133178711, -11.886016845703125, 8.343490600585938, 21.04668426513672, -1.599273681640625, -4.7832794189453125, 24.1733341217041, 31.750762939453125, -0.37237548828125, -7.5436248779296875, 8.702430725097656, 2.3775596618652344, 18.309783935546875, 39.210296630859375, 2.568866729736328, 0.909912109375, 6.1256103515625, -3.806558609008789, -0.9050140380859375, 19.304855346679688, 1.9471988677978516, -2.8599395751953125, 2.4495887756347656, 9.904182434082031, 31.652069091796875, 9.792655944824219, -16.07010269165039, 6.5993804931640625, 15.373409271240234, 2.9498291015625, -8.277023315429688, -9.899862289428711], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000214.npy"}
{"epoch": 0.3235071806500378, "step": 215, "batch_size": 64, "mean": 8.354536056518555, "std": 11.348723411560059, "min": -12.526756286621094, "p10": -5.244546508789062, "median": 7.462470054626465, "p90": 24.036719131469727, "max": 34.62220764160156, "pos_frac": 0.75, "sample": [3.5193939208984375, -6.543235778808594, -2.2380599975585938, 9.339241027832031, 3.6610107421875, 0.20379638671875, 3.7357635498046875, -1.4182510375976562, 27.956069946289062, 23.4505615234375, -5.409149169921875, -3.625751495361328, -9.449974060058594, 9.974544525146484, 10.117286682128906, 23.39013671875, 4.252037048339844, 13.736335754394531, 14.794441223144531, 3.0438690185546875, 6.89129638671875, 8.217491149902344, -2.6165428161621094, 12.355087280273438, 22.278135299682617, 6.387271881103516, -12.526756286621094, 21.38903045654297, 12.168136596679688, -4.8604736328125, 1.9770278930664062, 19.938796997070312, 1.3377227783203125, -1.7597503662109375, -7.149162292480469, -11.579116821289062, 14.280803680419922, 8.856559753417969, 16.359657287597656, 18.69659423828125, 4.046062469482422, 4.095390319824219, 6.695220947265625, 0.068084716796875, 26.942031860351562, 34.62220764160156, 20.727767944335938, 27.06610679626465, 8.03364372253418, 12.960067749023438, 24.28792953491211, 9.519065856933594, 10.049667358398438, -9.71551513671875, -1.617502212524414, 4.799532890319824, 10.60546875, 33.60235595703125, 25.24007797241211, -1.1810550689697266, 21.798141479492188, 0.332427978515625, 11.577735900878906, -2.9964599609375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000215.npy"}
{"epoch": 0.3250188964474679, "step": 216, "batch_size": 64, "mean": 6.552940845489502, "std": 10.691316604614258, "min": -22.266944885253906, "p10": -6.877248191833496, "median": 7.123008728027344, "p90": 19.795124053955078, "max": 29.01763916015625, "pos_frac": 0.734375, "sample": [3.2388916015625, 19.669166564941406, 3.72686767578125, 14.661117553710938, 12.489578247070312, 4.9312744140625, 13.765922546386719, 10.637016296386719, 17.15338897705078, 19.849105834960938, 7.872161865234375, 8.225807189941406, 11.727203369140625, 3.0416107177734375, 2.733551025390625, -6.582420349121094, 7.900661468505859, 9.840400695800781, -10.222854614257812, 23.642967224121094, 16.159317016601562, 23.822540283203125, 8.091987609863281, 4.5653839111328125, 12.240882873535156, -7.003602981567383, 10.890861511230469, 0.2822685241699219, 8.027008056640625, 20.701904296875, 14.026763916015625, -22.266944885253906, 1.8290748596191406, 8.04122543334961, 6.118133544921875, 5.2218017578125, -6.27813720703125, -1.8208122253417969, 17.29277801513672, -8.717716217041016, -2.2300071716308594, 16.015350341796875, 28.152488708496094, -5.122795104980469, -6.5626678466796875, -5.4669342041015625, 6.1381072998046875, 0.7428607940673828, -11.004287719726562, 6.3738555908203125, 28.717777252197266, -8.347908020019531, -8.956985473632812, 0.04684257507324219, 29.01763916015625, -0.47888851165771484, 8.107391357421875, 14.061607360839844, 0.9471664428710938, 14.68505859375, -4.88545036315918, 18.079208374023438, -0.28745269775390625, 12.120094299316406], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000216.npy"}
{"epoch": 0.32653061224489793, "step": 217, "batch_size": 64, "mean": 7.158425807952881, "std": 12.35999870300293, "min": -18.8577880859375, "p10": -7.543086242675779, "median": 5.871997833251953, "p90": 25.600283813476562, "max": 43.11088562011719, "pos_frac": 0.75, "sample": [4.983436584472656, 25.606365203857422, -8.752548217773438, 0.1507568359375, -2.1083030700683594, -1.5814743041992188, 0.6591682434082031, 43.11088562011719, -4.080841064453125, -10.4859619140625, 23.905807495117188, 26.0521240234375, 6.202278137207031, 5.410919189453125, 13.985649108886719, 5.760963439941406, 11.63482666015625, 5.270103454589844, 5.9830322265625, 2.630218505859375, 0.938079833984375, 25.58609390258789, 7.5594329833984375, 8.297096252441406, 6.181060791015625, 19.518463134765625, -12.077919006347656, 4.9452667236328125, -2.942699432373047, 19.969284057617188, 16.495935440063477, 3.3972854614257812, 11.569061279296875, -9.066337585449219, 1.8561058044433594, -9.310956954956055, 0.387298583984375, 10.958763122558594, -1.6138534545898438, -13.732002258300781, -18.8577880859375, 26.313003540039062, -4.6453094482421875, 1.0583000183105469, 0.7511672973632812, -4.316627502441406, 0.5052566528320312, 8.684051513671875, 6.7996063232421875, 34.82685852050781, 7.301643371582031, 14.350589752197266, 7.485441207885742, 0.7382984161376953, 30.666946411132812, 8.269256591796875, 33.599639892578125, 11.062667846679688, 7.176872253417969, 12.964332580566406, 14.114089965820312, 20.809829711914062, -0.05075836181640625, -4.72100830078125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000217.npy"}
{"epoch": 0.328042328042328, "step": 218, "batch_size": 64, "mean": 4.715145111083984, "std": 12.59392261505127, "min": -22.0726318359375, "p10": -12.27226219177246, "median": 5.223444938659668, "p90": 22.584593200683603, "max": 31.62804412841797, "pos_frac": 0.65625, "sample": [-8.006027221679688, 0.09480094909667969, 9.790435791015625, 6.6298675537109375, -1.156341552734375, 3.8818511962890625, 12.870849609375, -4.9766845703125, 14.317550659179688, 12.357566833496094, 1.3628082275390625, 3.3351783752441406, 16.788188934326172, -12.298088073730469, -12.365406036376953, -5.856494903564453, 8.474857330322266, 23.654945373535156, -22.0726318359375, 3.340179443359375, 6.6268768310546875, 12.048957824707031, -0.1618499755859375, -2.9675216674804688, 17.519468307495117, 24.008819580078125, 15.958084106445312, 30.560394287109375, 4.101421356201172, 16.848175048828125, -4.069506645202637, 10.796127319335938, -9.344161987304688, -12.21200180053711, -9.59415054321289, 6.851295471191406, 0.3081207275390625, 6.512237548828125, 0.7659072875976562, 10.79217529296875, 23.817707061767578, 6.8535003662109375, 3.3093767166137695, -5.3137054443359375, 5.389509201049805, -16.77658462524414, 20.08710479736328, -2.970195770263672, -12.463428497314453, -11.657012939453125, 13.112726211547852, -19.239185333251953, 13.465728759765625, -3.643369674682617, 7.8953094482421875, 14.726043701171875, 5.057380676269531, 31.62804412841797, -19.178329467773438, -1.4365501403808594, 9.97237777709961, 7.641578674316406, 28.070953369140625, 27.904043197631836], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000218.npy"}
{"epoch": 0.3295540438397581, "step": 219, "batch_size": 64, "mean": 6.653696060180664, "std": 10.551820755004883, "min": -13.642852783203125, "p10": -5.345341491699219, "median": 3.8574695587158203, "p90": 21.033586311340333, "max": 30.423187255859375, "pos_frac": 0.71875, "sample": [13.120170593261719, 25.573062896728516, 13.445465087890625, 15.407510757446289, 20.185043334960938, -2.8406753540039062, -1.6965866088867188, 2.3531436920166016, 2.8172988891601562, -8.81866455078125, 1.2825736999511719, 30.423187255859375, 15.73504638671875, 3.2105369567871094, -4.056976318359375, 3.0101165771484375, 1.8802986145019531, 1.3683547973632812, -0.2651634216308594, -1.0126571655273438, -4.385974884033203, -13.642852783203125, 1.6315689086914062, 4.3933563232421875, -5.5763702392578125, 2.0511512756347656, 16.7716064453125, -0.9216146469116211, 25.710647583007812, 3.748016357421875, 19.28729248046875, 17.97846221923828, 11.772438049316406, 21.158496856689453, 19.868072509765625, 28.552589416503906, 3.09674072265625, 4.191803932189941, -9.540519714355469, 3.9669227600097656, 8.32419204711914, 6.077056884765625, 10.166160583496094, -7.25885009765625, 6.9033203125, -6.9027099609375, 20.742128372192383, -3.0726165771484375, 15.684066772460938, 23.675193786621094, -0.0559539794921875, 7.079593658447266, 9.606338500976562, 2.57806396484375, -13.565971374511719, 15.27178955078125, 7.7783050537109375, 0.7299423217773438, 7.5938720703125, -4.225894927978516, 23.994274139404297, -4.8062744140625, 15.069229125976562, 3.218353271484375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000219.npy"}
{"epoch": 0.3310657596371882, "step": 220, "batch_size": 64, "mean": 5.840293884277344, "std": 11.429299354553223, "min": -16.751502990722656, "p10": -7.590653991699218, "median": 4.106866836547852, "p90": 21.046746063232426, "max": 30.889389038085938, "pos_frac": 0.65625, "sample": [-7.50164794921875, 1.0246124267578125, 7.063014984130859, 3.2561302185058594, 6.116342544555664, -0.9802703857421875, 15.699920654296875, 2.8766098022460938, 0.028829574584960938, 3.4388504028320312, -0.8014297485351562, 30.71294403076172, 8.298101425170898, -14.413562774658203, 4.392475128173828, 17.03022575378418, 18.432998657226562, -14.194282531738281, -2.1407814025878906, 12.4835205078125, 15.644023895263672, 17.419410705566406, -3.791177749633789, -4.230445861816406, -7.6287994384765625, -16.751502990722656, -0.3097877502441406, -3.26971435546875, 19.16339111328125, 14.710968017578125, 20.37371063232422, 5.908210754394531, 14.00048828125, 9.635002136230469, 21.335189819335938, -2.177532196044922, 22.024734497070312, 2.230316162109375, 24.563501358032227, -7.724277496337891, 3.821258544921875, 4.997123718261719, -0.2676849365234375, 9.160919189453125, 30.889389038085938, -5.204307556152344, 25.803482055664062, -0.8565311431884766, -5.019207000732422, 16.375831604003906, 3.750457763671875, -10.601581573486328, 15.025938034057617, -1.8605842590332031, 5.262763977050781, 8.015884399414062, -16.60919189453125, 12.891593933105469, 11.616645812988281, 1.4077510833740234, 29.613494873046875, -4.173978805541992, 5.803020477294922, 1.9879932403564453], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000220.npy"}
{"epoch": 0.3325774754346183, "step": 221, "batch_size": 64, "mean": 6.562739849090576, "std": 12.5269193649292, "min": -23.260887145996094, "p10": -7.170808792114258, "median": 4.320318222045898, "p90": 25.875111770629886, "max": 40.824188232421875, "pos_frac": 0.65625, "sample": [5.260982513427734, -5.645160675048828, 12.432174682617188, 9.140335083007812, -3.1892929077148438, 9.735710144042969, 1.0946578979492188, 7.687774658203125, 5.0622100830078125, -5.9438323974609375, -6.732814788818359, -0.8265838623046875, 21.619102478027344, 14.075019836425781, -7.763236999511719, 0.23351478576660156, -0.8915557861328125, -0.0741424560546875, 2.2861461639404297, 2.1008224487304688, 25.256649017333984, 28.388671875, 3.8103179931640625, 12.597579956054688, 26.140167236328125, 0.6243896484375, 15.155071258544922, 27.873523712158203, -0.6421356201171875, -10.125150680541992, 11.251083374023438, 24.706573486328125, -9.337852478027344, 4.830318450927734, 29.030014038085938, 10.592926025390625, 10.5228271484375, -9.074089050292969, 16.4427490234375, 11.519561767578125, 17.47441864013672, -0.3565216064453125, -1.884765625, 5.101339340209961, 3.0628433227539062, -2.2469253540039062, -3.105581283569336, 27.483612060546875, -7.3585205078125, -3.9522056579589844, -20.93267822265625, -23.260887145996094, 3.387176513671875, 18.89056396484375, 12.94757080078125, -1.2518081665039062, 3.014251708984375, 13.575477600097656, 40.824188232421875, 13.162384033203125, -5.94964599609375, 12.515205383300781, 27.448200225830078, 2.2026596069335938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000221.npy"}
{"epoch": 0.3340891912320484, "step": 222, "batch_size": 64, "mean": 8.472888946533203, "std": 11.091513633728027, "min": -12.639816284179688, "p10": -5.051002693176269, "median": 8.850826263427734, "p90": 23.737231445312503, "max": 34.27501678466797, "pos_frac": 0.703125, "sample": [-1.5988044738769531, -5.4533843994140625, 5.902637481689453, 0.1973114013671875, 14.961929321289062, 14.599613189697266, 9.815925598144531, -2.8940467834472656, 23.889739990234375, 27.597431182861328, 13.4508056640625, 3.135417938232422, -12.639816284179688, 5.921939849853516, 9.756792068481445, 0.6206283569335938, -1.8221206665039062, 5.216682434082031, -0.4451255798339844, 6.046195983886719, 6.824928283691406, -0.24622058868408203, 28.832260131835938, -2.7304916381835938, -0.5001068115234375, 26.609676361083984, 18.883407592773438, 32.991424560546875, -9.581733703613281, 8.05377197265625, 9.807640075683594, 14.399261474609375, 14.550422668457031, 22.157182693481445, 12.659652709960938, 11.236701965332031, 10.504657745361328, -7.7184600830078125, -4.112112045288086, -2.7539520263671875, -5.7635650634765625, 15.427642822265625, 0.14495849609375, 8.927215576171875, 2.476602554321289, 23.381378173828125, -0.86224365234375, -6.119148254394531, 8.774436950683594, 16.4737548828125, 16.782691955566406, 7.196529388427734, 9.793571472167969, 34.27501678466797, -1.4250335693359375, 15.78887939453125, 24.367645263671875, 21.814682006835938, -3.5941009521484375, 20.783096313476562, 12.388359069824219, 13.805866241455078, 12.700729370117188, -11.401702880859375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000222.npy"}
{"epoch": 0.3356009070294785, "step": 223, "batch_size": 64, "mean": 8.803045272827148, "std": 11.898703575134277, "min": -21.43120574951172, "p10": -4.634222412109374, "median": 6.456317901611328, "p90": 23.3266960144043, "max": 38.605560302734375, "pos_frac": 0.84375, "sample": [6.486228942871094, 17.38343048095703, 6.882270812988281, 1.3393630981445312, 0.8013763427734375, 17.423614501953125, -5.819732666015625, 4.0896453857421875, -12.873390197753906, 19.13327407836914, 6.4264068603515625, 38.605560302734375, 14.929410934448242, 4.37078857421875, 3.58843994140625, 10.59429931640625, 2.9497146606445312, 3.503997802734375, 9.68115234375, -6.650382995605469, 23.903533935546875, 12.818954467773438, -15.622581481933594, 5.009979248046875, 16.30185317993164, 1.9803695678710938, 1.4970321655273438, 22.053009033203125, -9.584293365478516, -21.43120574951172, 12.044029235839844, 4.973686218261719, 0.2943115234375, 1.8521575927734375, 9.419578552246094, 22.255146026611328, 13.263465881347656, 19.167510986328125, -5.1911468505859375, 0.3050117492675781, -2.678070068359375, 11.136356353759766, 5.811515808105469, 11.397735595703125, 30.785186767578125, 18.888282775878906, 16.510986328125, 30.115337371826172, 1.0191650390625, 19.40574836730957, 33.827552795410156, -3.3347320556640625, 23.36675262451172, 30.083908081054688, 14.712738037109375, 4.339263916015625, 20.06319808959961, 23.233230590820312, 3.6794357299804688, -2.649120330810547, 1.9796142578125, 1.9788665771484375, 7.36993408203125, 4.1961212158203125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000223.npy"}
{"epoch": 0.3371126228269085, "step": 224, "batch_size": 64, "mean": 5.63276481628418, "std": 10.37497329711914, "min": -16.921993255615234, "p10": -7.607929801940917, "median": 4.687859535217285, "p90": 18.370809555053714, "max": 33.20642852783203, "pos_frac": 0.71875, "sample": [2.2703514099121094, 33.20642852783203, 0.6276397705078125, 1.475372314453125, 6.12689208984375, 7.35101318359375, 16.59703826904297, 9.626720428466797, -1.2726116180419922, 4.593879699707031, 14.02825927734375, 10.860275268554688, 11.178813934326172, 24.015296936035156, 4.781839370727539, -1.9171333312988281, 5.768104553222656, 16.655872344970703, 12.071022033691406, 8.721954345703125, -6.5243682861328125, 14.032081604003906, 0.4581451416015625, -16.921993255615234, -4.7033843994140625, -15.66510009765625, 1.991363525390625, 2.6775283813476562, -0.707977294921875, -12.239410400390625, 25.096721649169922, -2.275754928588867, -10.28969955444336, 4.24591064453125, 9.165328979492188, 0.8394699096679688, -2.8024368286132812, -8.659919738769531, 9.999881744384766, 6.4912109375, -0.9681777954101562, 4.515430450439453, 11.141937255859375, -0.9430007934570312, -13.426498413085938, 2.99542236328125, 21.19145965576172, 17.134479522705078, 9.18853759765625, 2.449512481689453, -8.07231330871582, 14.501585006713867, 6.607799530029297, 18.900665283203125, 27.300952911376953, 7.635398864746094, 1.1687202453613281, 6.119384765625, -1.3472671508789062, 27.19632339477539, -0.07684326171875, 15.3720703125, 6.6081085205078125, 4.3286285400390625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000224.npy"}
{"epoch": 0.3386243386243386, "step": 225, "batch_size": 64, "mean": 8.266858100891113, "std": 14.780274391174316, "min": -34.79669189453125, "p10": -7.8062187194824215, "median": 4.520171165466309, "p90": 26.88406505584717, "max": 49.9912109375, "pos_frac": 0.703125, "sample": [-6.310651779174805, 9.345195770263672, 3.9842872619628906, 1.9641876220703125, 11.39837646484375, 27.04559326171875, 2.3048553466796875, -10.044988632202148, -0.6339225769042969, 27.231067657470703, 0.1150054931640625, 22.647064208984375, -1.1497421264648438, 21.582992553710938, -1.519775390625, 3.3095169067382812, -4.720623016357422, 1.4699249267578125, 15.984737396240234, 25.257888793945312, 26.507165908813477, 14.213808059692383, 49.9912109375, 7.306510925292969, 36.4969482421875, 4.466270446777344, 25.6627197265625, 14.657470703125, 23.46912384033203, -5.611452102661133, 4.908531188964844, -34.79669189453125, 19.000473022460938, -6.1352081298828125, 26.195274353027344, 11.773696899414062, 4.574071884155273, -8.017730712890625, 21.251708984375, 0.1351165771484375, 5.614013671875, 1.2366180419921875, 34.52747344970703, -5.760124206542969, 10.132980346679688, -4.688079833984375, -10.0380859375, -7.312690734863281, 17.292518615722656, 0.535919189453125, 15.916862487792969, 3.515766143798828, 14.985435485839844, 20.719375610351562, 28.956443786621094, -13.66426944732666, 0.9109954833984375, 2.6732101440429688, 28.657455444335938, -0.5657577514648438, -8.93438720703125, -11.116073608398438, 22.52745819091797, -2.3541641235351562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000225.npy"}
{"epoch": 0.3401360544217687, "step": 226, "batch_size": 64, "mean": 7.866827487945557, "std": 13.584572792053223, "min": -24.649368286132812, "p10": -5.5772056579589835, "median": 5.829093933105469, "p90": 26.24948196411133, "max": 36.20185852050781, "pos_frac": 0.71875, "sample": [17.571868896484375, 2.153411865234375, -1.8375701904296875, 8.858612060546875, 5.81622314453125, 4.915130615234375, 26.615570068359375, 18.279945373535156, 2.596454620361328, -17.618743896484375, 20.786312103271484, 7.844478607177734, 10.896377563476562, 24.11611557006836, -3.471973419189453, -4.003822326660156, 12.586807250976562, -0.5771026611328125, 21.67055320739746, 8.455070495605469, 26.305953979492188, 1.6311416625976562, 2.1219329833984375, 11.548210144042969, 10.868728637695312, 17.5305118560791, 5.203838348388672, 31.84436798095703, 24.0570068359375, 21.10572052001953, -8.005245208740234, 36.20185852050781, 3.1495819091796875, -3.5433807373046875, 26.474239349365234, 9.548980712890625, 5.532157897949219, -24.349563598632812, -5.8655853271484375, 0.1444263458251953, -1.4280471801757812, 5.8419647216796875, -3.5940093994140625, -16.351608276367188, 8.081520080566406, 28.558029174804688, -24.649368286132812, 7.215789794921875, -14.784666061401367, 19.139907836914062, 4.177757263183594, 3.1832542419433594, 11.3355712890625, 5.6280059814453125, -1.8376922607421875, 34.804931640625, 21.653106689453125, 19.8973388671875, 1.9467506408691406, -3.416635513305664, 26.117713928222656, -4.904319763183594, -1.1291084289550781, 20.832164764404297], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000226.npy"}
{"epoch": 0.3416477702191988, "step": 227, "batch_size": 64, "mean": 8.073860168457031, "std": 15.071760177612305, "min": -25.751296997070312, "p10": -7.581941223144531, "median": 6.289279937744141, "p90": 31.49719276428223, "max": 37.798744201660156, "pos_frac": 0.625, "sample": [31.94293975830078, 17.59572982788086, -0.8564987182617188, 31.830795288085938, 1.173248291015625, -24.819469451904297, 19.65477752685547, 26.344207763671875, -25.751296997070312, -2.36358642578125, 31.20916748046875, -7.709075927734375, -3.0611305236816406, 31.29509735107422, 22.10580825805664, -7.2852935791015625, -0.4671916961669922, -3.2317352294921875, 24.509960174560547, -2.923614501953125, 5.8461151123046875, 13.27093505859375, 15.592941284179688, 13.782527923583984, 31.98046112060547, 16.0302734375, -6.349750518798828, 14.902778625488281, -2.730182647705078, -1.744405746459961, -4.183509826660156, 31.583805084228516, 32.39146423339844, 5.936424255371094, 11.735519409179688, 16.517385482788086, 37.798744201660156, -0.0775909423828125, 3.3319664001464844, 22.37340545654297, 1.665496826171875, 7.740478515625, -8.244682312011719, 8.498977661132812, 18.14862060546875, -3.1719093322753906, 11.205268859863281, 6.6421356201171875, -4.957496643066406, 2.107706069946289, -20.499469757080078, 13.739715576171875, -2.3411407470703125, 31.947341918945312, 1.0849838256835938, -4.3078460693359375, 7.341239929199219, -0.54571533203125, 18.487823486328125, 1.2994232177734375, -8.774208068847656, -19.959365844726562, 21.14330291748047, 21.29425811767578], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000227.npy"}
{"epoch": 0.3431594860166289, "step": 228, "batch_size": 64, "mean": 7.148487567901611, "std": 13.754227638244629, "min": -24.38570785522461, "p10": -9.301760864257812, "median": 7.255659103393555, "p90": 25.213389778137213, "max": 35.6065788269043, "pos_frac": 0.6875, "sample": [-6.020290374755859, 31.526473999023438, 14.6405029296875, 32.94402313232422, 9.031745910644531, 17.445587158203125, 8.390045166015625, 26.800395965576172, 3.157958984375, -2.98486328125, 16.70065689086914, 1.4779891967773438, 12.738067626953125, -24.38570785522461, 1.3264312744140625, 3.0006065368652344, 19.166942596435547, 3.3347320556640625, -3.94281005859375, 23.889219284057617, 6.797908782958984, 18.9400634765625, -4.712810516357422, 1.9980621337890625, -7.933868408203125, 21.920944213867188, 25.78089141845703, 35.6065788269043, 14.801956176757812, 14.882671356201172, -13.219694137573242, -13.517112731933594, 11.383079528808594, -13.444145202636719, 10.151742935180664, -20.73025894165039, 9.586479187011719, -0.7099037170410156, -15.134979248046875, -8.204208374023438, 11.191055297851562, 5.052833557128906, 14.9169921875, -7.519382476806641, 6.954925537109375, 8.39463996887207, 22.53205108642578, 13.119979858398438, -4.346294403076172, 7.1284027099609375, -6.747730255126953, -2.4275588989257812, 19.971450805664062, 7.382915496826172, -4.065155029296875, 8.219383239746094, 32.5223388671875, 30.958866119384766, 23.825668334960938, -9.772140502929688, -3.445343017578125, 0.17285919189453125, 17.754383087158203, 3.2469482421875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000228.npy"}
{"epoch": 0.34467120181405897, "step": 229, "batch_size": 64, "mean": 8.164986610412598, "std": 12.945555686950684, "min": -24.700702667236328, "p10": -7.6788589477539055, "median": 7.680672645568848, "p90": 25.451615142822266, "max": 37.499114990234375, "pos_frac": 0.75, "sample": [4.912605285644531, 12.466373443603516, 20.86522674560547, 14.555618286132812, 17.762779235839844, 9.55279541015625, 29.217693328857422, 14.295732498168945, 18.362499237060547, 4.001682281494141, 1.8749542236328125, 18.555694580078125, 23.428443908691406, -6.7211761474609375, 13.361160278320312, 6.065496444702148, -8.08929443359375, -2.4988250732421875, 25.264572143554688, 14.613861083984375, 5.5561370849609375, 4.140460968017578, 4.8523101806640625, 16.026290893554688, -10.568374633789062, 8.474655151367188, -19.437652587890625, -2.4727659225463867, -3.1861019134521484, 4.6689910888671875, 10.715194702148438, 36.890167236328125, -19.109554290771484, 11.685129165649414, 0.2696685791015625, 3.1850547790527344, 25.610687255859375, -10.421806335449219, -2.9434738159179688, 7.9423065185546875, 32.427345275878906, 7.557222366333008, -0.11533164978027344, -8.873136520385742, 9.096576690673828, 5.3416900634765625, 24.70716094970703, 8.363746643066406, 9.553022384643555, 16.746082305908203, 5.487083435058594, -1.1408061981201172, 8.87545394897461, 4.2165679931640625, 19.992523193359375, 7.8041229248046875, -2.567028045654297, 3.6780223846435547, 29.497529983520508, 37.499114990234375, 25.531776428222656, 2.8704566955566406, -3.0145835876464844, -24.700702667236328], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000229.npy"}
{"epoch": 0.34618291761148906, "step": 230, "batch_size": 64, "mean": 6.360638618469238, "std": 13.560385704040527, "min": -27.638648986816406, "p10": -12.458681488037108, "median": 6.383087158203125, "p90": 25.40067825317383, "max": 33.799156188964844, "pos_frac": 0.6875, "sample": [-12.99700927734375, 13.404342651367188, 8.033466339111328, 2.6209716796875, 17.161102294921875, -16.38604736328125, 6.7339324951171875, -0.39363861083984375, -17.234466552734375, 18.060829162597656, 4.974494934082031, 20.04535675048828, 25.601058959960938, 15.477432250976562, 0.8451461791992188, 9.237272262573242, 17.93454360961914, 26.74846649169922, -20.378707885742188, -5.251468658447266, -27.638648986816406, 7.431301116943359, 1.908538818359375, 10.609798431396484, 9.317337036132812, -3.808521270751953, 20.82691192626953, -0.8250503540039062, 25.573875427246094, 19.116188049316406, 16.011133193969727, -3.4485511779785156, -0.8719272613525391, 3.5139007568359375, -2.7012786865234375, -17.283187866210938, -1.7845687866210938, 5.461734771728516, 31.214324951171875, 18.95941162109375, 12.7501220703125, 6.0322418212890625, -0.8173675537109375, 6.882759094238281, 33.799156188964844, 9.544139862060547, -19.497821807861328, 25.93509292602539, 2.908477783203125, -8.192476272583008, 17.89593505859375, -11.202583312988281, 17.268714904785156, 13.38290786743164, 3.6083202362060547, 24.996551513671875, 4.184474945068359, 0.7333221435546875, 7.906654357910156, -0.5567092895507812, 0.3314352035522461, 31.355010986328125, -7.299144744873047, 9.311851501464844], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000230.npy"}
{"epoch": 0.3476946334089191, "step": 231, "batch_size": 64, "mean": 9.541065216064453, "std": 14.043618202209473, "min": -25.138885498046875, "p10": -6.590239715576171, "median": 7.249624252319336, "p90": 29.45005569458008, "max": 35.180503845214844, "pos_frac": 0.78125, "sample": [25.61986541748047, -2.312835693359375, 26.393287658691406, 1.711944580078125, 1.58245849609375, 3.7727127075195312, 0.202301025390625, 25.894195556640625, 7.9272003173828125, -1.8824138641357422, 10.009010314941406, 2.6562042236328125, 10.65087890625, 4.91412353515625, -13.201461791992188, 6.572048187255859, -2.3496551513671875, 26.78403091430664, 24.325103759765625, -7.050880432128906, 8.704498291015625, -12.76190185546875, 17.01291275024414, 10.718799591064453, 3.463153839111328, 31.63427734375, -2.3345184326171875, -20.894123077392578, 32.72511672973633, 12.197406768798828, 30.356361389160156, -8.942024230957031, 2.3525543212890625, 31.5113525390625, 14.094085693359375, -4.813720703125, 5.361841201782227, 19.682723999023438, 11.8067626953125, -7.8507537841796875, 28.75558853149414, 0.015899658203125, 3.4785308837890625, -5.515411376953125, 26.14080047607422, 0.8967199325561523, 3.3864974975585938, -3.319171905517578, 12.456878662109375, 19.504947662353516, 16.874542236328125, 28.716583251953125, 5.2970428466796875, 14.076042175292969, 26.510112762451172, 35.180503845214844, 0.1872406005859375, 11.290481567382812, 5.4632568359375, 15.2760009765625, -25.138885498046875, 2.7368011474609375, 29.747684478759766, 32.366607666015625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000231.npy"}
{"epoch": 0.3492063492063492, "step": 232, "batch_size": 64, "mean": 6.433259010314941, "std": 14.921506881713867, "min": -23.89525604248047, "p10": -9.577540588378906, "median": 3.3862838745117188, "p90": 26.087162017822266, "max": 43.79388427734375, "pos_frac": 0.625, "sample": [-0.5346450805664062, 27.225910186767578, 14.443580627441406, 40.93205261230469, 2.866220474243164, -13.42718505859375, -5.8021087646484375, 26.23821258544922, -9.502548217773438, 6.2352447509765625, -9.60968017578125, 4.475048065185547, -21.55921173095703, -0.012256622314453125, -4.408359527587891, 23.43244171142578, -3.000225067138672, -0.7568359375, 2.0435028076171875, 7.184268951416016, -0.8289031982421875, 6.4986419677734375, 3.1929779052734375, 12.150787353515625, 16.56885528564453, 6.903890609741211, -8.678699493408203, 29.306434631347656, -11.468727111816406, 17.067420959472656, 1.5080337524414062, 25.734710693359375, -7.376611709594727, -0.9935417175292969, 0.346893310546875, 0.1189727783203125, -1.1128559112548828, -3.3927383422851562, 37.82170104980469, 9.694252014160156, -0.4897041320800781, 17.85455322265625, 10.991195678710938, 43.79388427734375, 23.301849365234375, 40.6414794921875, -20.723922729492188, 2.674407958984375, -3.7008056640625, 3.57958984375, 24.108154296875, 8.597248077392578, 21.304977416992188, 6.374011993408203, 8.204444885253906, -23.89525604248047, -0.678074836730957, 7.356266021728516, 0.740478515625, 11.015937805175781, -5.095916748046875, -16.7593994140625, 21.658973693847656, 11.349286079406738], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000232.npy"}
{"epoch": 0.3507180650037793, "step": 233, "batch_size": 64, "mean": 6.6543354988098145, "std": 14.119437217712402, "min": -19.35712432861328, "p10": -10.408314514160155, "median": 4.205394744873047, "p90": 26.622649383544925, "max": 40.32415771484375, "pos_frac": 0.625, "sample": [40.32415771484375, 19.981876373291016, 23.55944061279297, -5.915151596069336, 4.390331268310547, 7.223048210144043, -3.1870193481445312, 26.134136199951172, -2.792285919189453, 6.7708282470703125, 17.75536346435547, 36.42204284667969, -1.0940475463867188, -3.5409011840820312, -2.71014404296875, 28.207839965820312, 8.634237289428711, 0.11318397521972656, -1.517303466796875, -4.142036437988281, 22.279800415039062, 5.289913177490234, 20.6929931640625, 0.1765289306640625, -17.0528564453125, 8.011871337890625, -10.116195678710938, 6.675056457519531, -10.580543518066406, -8.820770263671875, 12.528610229492188, -1.6728019714355469, 14.945663452148438, 8.04541015625, -0.9595794677734375, -14.124786376953125, 17.061012268066406, 4.467437744140625, 4.020458221435547, 0.3968791961669922, 7.031219482421875, -19.35712432861328, 30.91217041015625, 0.9188919067382812, 6.511474609375, -3.8304481506347656, 9.242950439453125, -10.53350830078125, 26.832012176513672, 16.791330337524414, -0.854156494140625, -11.418651580810547, 1.6192131042480469, 25.799575805664062, 35.00929260253906, 34.29255676269531, 1.9005136489868164, -10.582908630371094, 22.147964477539062, -9.388565063476562, -0.46462249755859375, 0.39124298095703125, -3.9434547424316406, 20.96880340576172], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000233.npy"}
{"epoch": 0.35222978080120937, "step": 234, "batch_size": 64, "mean": 6.796961784362793, "std": 13.529097557067871, "min": -17.829269409179688, "p10": -9.454720306396483, "median": 4.407646179199219, "p90": 27.955224609375, "max": 33.45579528808594, "pos_frac": 0.625, "sample": [-3.093597412109375, -15.534194946289062, 8.118568420410156, -4.235687255859375, 2.796802520751953, -3.5996170043945312, 25.27680778503418, -1.5232048034667969, 14.455581665039062, -3.716094970703125, 4.3734893798828125, 12.116456985473633, 14.742408752441406, 18.943275451660156, 12.325599670410156, -0.9670372009277344, -0.06450653076171875, -17.829269409179688, 22.814361572265625, -11.702095031738281, 3.3894004821777344, 11.60693359375, 17.637588500976562, 32.17011260986328, 2.0598621368408203, -1.5903892517089844, 30.43041229248047, 4.48588752746582, -9.5926513671875, 1.0298652648925781, 22.0380859375, 29.82059097290039, 28.38235855102539, 23.90904998779297, 8.703611373901367, -9.132881164550781, 4.441802978515625, -4.101844787597656, 1.2861270904541016, -2.8913230895996094, 9.709770202636719, 20.643165588378906, -0.22418212890625, 33.45579528808594, 28.17626953125, 7.0145111083984375, 0.0567474365234375, 14.069408416748047, 27.439453125, -12.385879516601562, -6.5616455078125, 5.035556793212891, -15.400321960449219, 30.539871215820312, -0.3422069549560547, -8.525993347167969, 12.708694458007812, 9.552074432373047, -5.469451904296875, 2.5961990356445312, -6.361362457275391, 25.53224754333496, 7.9084320068359375, -11.942230224609375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000234.npy"}
{"epoch": 0.35374149659863946, "step": 235, "batch_size": 64, "mean": 8.745163917541504, "std": 12.232616424560547, "min": -21.088417053222656, "p10": -4.180493927001953, "median": 7.129840850830078, "p90": 24.0785831451416, "max": 34.4947624206543, "pos_frac": 0.8125, "sample": [-1.6541824340820312, 10.022254943847656, 8.224044799804688, 8.86009407043457, 19.916778564453125, -3.5175094604492188, -5.430397033691406, 14.088218688964844, 6.48870849609375, 5.472442626953125, 2.3937835693359375, 24.027759552001953, 1.9762630462646484, 18.67562484741211, 11.775146484375, 19.510211944580078, 7.7436065673828125, 0.8516845703125, 18.049850463867188, 9.157905578613281, 3.8025283813476562, 4.448905944824219, -12.038421630859375, 21.784332275390625, 23.952838897705078, 7.538753509521484, 0.09654998779296875, -7.4533843994140625, 34.4947624206543, 1.4216156005859375, 15.107070922851562, 0.7366399765014648, 23.170839309692383, -4.464630126953125, 24.100364685058594, 4.015032768249512, 21.551349639892578, 5.060966491699219, 31.57720947265625, -19.16126251220703, 5.807579040527344, 32.882835388183594, -1.727783203125, 1.7658004760742188, 18.75885772705078, 8.164470672607422, 0.9181671142578125, 9.450077056884766, 30.3006591796875, 6.720928192138672, 4.175376892089844, 1.956024169921875, 27.11627197265625, 17.42296600341797, 4.554370880126953, -21.088417053222656, 32.238319396972656, 13.60635757446289, 14.713302612304688, 1.6633453369140625, 8.574077606201172, -13.494514465332031, -0.9167442321777344, -0.2462005615234375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000235.npy"}
{"epoch": 0.35525321239606955, "step": 236, "batch_size": 64, "mean": 7.980992317199707, "std": 15.904292106628418, "min": -32.95228576660156, "p10": -8.933319664001463, "median": 7.668188095092773, "p90": 28.20691375732422, "max": 37.34796142578125, "pos_frac": 0.65625, "sample": [8.99859619140625, -5.1790618896484375, 13.188285827636719, 20.79743194580078, 4.408348083496094, 9.128875732421875, 23.819480895996094, 26.72625732421875, -2.4507522583007812, 9.409225463867188, 18.485790252685547, 21.251953125, -2.958993911743164, 4.201225280761719, -4.399539947509766, -2.0712432861328125, 18.737018585205078, 19.956192016601562, 37.04443359375, 31.3258056640625, -0.30834388732910156, -3.5865325927734375, 2.7404632568359375, -9.636123657226562, 23.488758087158203, 27.681625366210938, 15.020771026611328, 28.432037353515625, 25.580673217773438, 15.4854736328125, 29.277618408203125, 2.5341415405273438, -15.729080200195312, 18.882244110107422, 13.0684814453125, 2.3293609619140625, -4.35064697265625, 37.34796142578125, -17.467185974121094, 15.698612213134766, -7.29344367980957, 5.886360168457031, 13.464950561523438, 25.06903839111328, 1.1357574462890625, -2.048633575439453, -4.2454681396484375, -29.31188201904297, 9.891357421875, 6.337779998779297, -2.3881988525390625, -32.95228576660156, -26.197647094726562, 26.622100830078125, 36.346641540527344, 12.824462890625, 9.327896118164062, -1.68292236328125, -4.316099166870117, -14.701889038085938, 30.368427276611328, -2.4463882446289062, 0.5258331298828125, 3.6581192016601562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000236.npy"}
{"epoch": 0.35676492819349964, "step": 237, "batch_size": 64, "mean": 5.391643524169922, "std": 12.6112060546875, "min": -21.487640380859375, "p10": -10.10698127746582, "median": 4.48394775390625, "p90": 20.884644317626954, "max": 42.29583740234375, "pos_frac": 0.625, "sample": [23.651016235351562, 15.697433471679688, -2.50189208984375, 32.74066162109375, 3.0723495483398438, 5.0142974853515625, 8.175811767578125, 3.159229278564453, 10.522464752197266, 8.964160919189453, 4.123649597167969, -5.4102935791015625, -9.248992919921875, 14.196548461914062, -0.384979248046875, 21.081741333007812, -1.2146263122558594, 24.862884521484375, -10.06341552734375, 5.5502471923828125, -21.487640380859375, 12.064656257629395, -11.1761474609375, 14.332508087158203, 31.0830078125, -4.425069808959961, 0.3764801025390625, 11.316499710083008, -14.727005004882812, -1.3427448272705078, 15.928327560424805, 6.811321258544922, 20.42475128173828, -10.460128784179688, 5.46502685546875, -3.7604141235351562, 3.7645339965820312, -19.826629638671875, 25.878829956054688, -10.125652313232422, 2.2535858154296875, 6.2935028076171875, 0.9532585144042969, -3.1735763549804688, -1.9057769775390625, -4.606977462768555, 10.345211029052734, 15.93719482421875, -2.088897705078125, 19.348648071289062, 17.216751098632812, 6.408653259277344, 7.8381805419921875, 42.29583740234375, -7.703948974609375, -1.4202194213867188, 1.684225082397461, 16.066150665283203, 4.844245910644531, 19.87862777709961, -2.2577438354492188, -10.162910461425781, -5.696941375732422, 10.615306854248047], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000237.npy"}
{"epoch": 0.35827664399092973, "step": 238, "batch_size": 64, "mean": 8.181037902832031, "std": 13.067937850952148, "min": -23.95110321044922, "p10": -6.720365905761718, "median": 6.000694274902344, "p90": 25.74182929992676, "max": 37.45446014404297, "pos_frac": 0.71875, "sample": [-3.0287551879882812, 6.607662200927734, 28.64505958557129, -14.62628173828125, 10.295188903808594, 32.534912109375, -3.8335418701171875, 11.907928466796875, 36.266754150390625, -8.005977630615234, -2.1534576416015625, -6.9456634521484375, 10.2774658203125, 4.270839691162109, 21.56416893005371, 14.708477020263672, 1.8093032836914062, -6.194671630859375, 19.622589111328125, 22.533363342285156, 24.68560028076172, 37.45446014404297, 15.974903106689453, 6.219764709472656, 4.468889236450195, 28.786537170410156, 12.369346618652344, -2.274993896484375, -6.018009185791016, 7.6418609619140625, 31.55999755859375, 3.882542610168457, 5.010368347167969, -1.0459213256835938, -4.383806228637695, 3.1658782958984375, 20.45654296875, 3.064960479736328, 25.645877838134766, 25.78295135498047, 0.6767807006835938, -4.236660003662109, 1.7433547973632812, -0.6738433837890625, 10.948516845703125, 15.163726806640625, 18.485260009765625, 24.051616668701172, 3.305622100830078, 13.6829833984375, 12.510604858398438, 5.781623840332031, 18.017959594726562, -23.95110321044922, -0.8702545166015625, 5.2024383544921875, -10.517837524414062, 8.737762451171875, 9.011470794677734, 2.5552520751953125, 14.91949462890625, -11.083053588867188, -12.808303833007812, 4.229866027832031], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000238.npy"}
{"epoch": 0.35978835978835977, "step": 239, "batch_size": 64, "mean": 8.40644645690918, "std": 15.070500373840332, "min": -22.0296630859375, "p10": -12.234173583984374, "median": 6.034017562866211, "p90": 35.174712371826175, "max": 40.21565246582031, "pos_frac": 0.734375, "sample": [6.061367034912109, -1.576812744140625, 24.03160858154297, 3.4968414306640625, 36.16093444824219, 6.141944885253906, -12.883922576904297, 19.710227966308594, -17.233322143554688, 7.54888916015625, -17.214019775390625, -0.6343841552734375, -12.78329086303711, 40.21565246582031, -0.6599884033203125, 2.9709625244140625, -3.5712547302246094, -10.952899932861328, 3.79144287109375, 36.904296875, -0.9991188049316406, 14.27215576171875, 16.705677032470703, 5.5311431884765625, 36.202972412109375, 3.885247230529785, 4.131582260131836, 19.206104278564453, 2.124469757080078, 6.808624267578125, 34.433082580566406, 7.161251068115234, -18.01959991455078, 0.24366378784179688, -0.26685333251953125, 35.4925537109375, 0.10911750793457031, 12.278099060058594, 3.787109375, 17.561172485351562, -1.6694869995117188, 15.323379516601562, 9.183609008789062, 14.380722045898438, -7.199527740478516, 22.797821044921875, 22.56690216064453, 39.879547119140625, 37.25927734375, -22.0296630859375, -0.22270488739013672, 0.9106979370117188, -14.38345718383789, 6.0066680908203125, 3.198883056640625, 7.481666564941406, 26.44361114501953, 1.3479862213134766, 11.324020385742188, 19.086761474609375, 1.6600475311279297, 7.9874114990234375, 15.806007385253906, 10.699695587158203], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000239.npy"}
{"epoch": 0.36130007558578986, "step": 240, "batch_size": 64, "mean": 12.479536056518555, "std": 15.309183120727539, "min": -33.20550537109375, "p10": -4.302278518676757, "median": 14.48248291015625, "p90": 30.884319305419925, "max": 39.036041259765625, "pos_frac": 0.78125, "sample": [23.690567016601562, -6.6786651611328125, 20.21630859375, -3.526092529296875, 19.649581909179688, 39.036041259765625, 0.1374359130859375, 25.69793701171875, -12.178308486938477, -4.634929656982422, 9.367149353027344, 0.195831298828125, 14.697502136230469, 31.1173095703125, 25.583106994628906, 24.38695526123047, 24.320480346679688, -7.156976699829102, 18.69915008544922, 32.21125793457031, -2.7214508056640625, 18.661216735839844, 1.9296197891235352, -33.20550537109375, 22.47736358642578, 37.09617614746094, 3.2445144653320312, 12.40008544921875, 34.5079345703125, 24.3720703125, 18.144630432128906, 21.630172729492188, 5.032367706298828, -2.797027587890625, -1.0426483154296875, 19.048629760742188, 0.7237586975097656, 25.35748291015625, 32.456634521484375, 5.391378402709961, 14.267463684082031, 7.750461578369141, -8.588218688964844, 28.19574737548828, 38.32405090332031, -27.327651977539062, -0.9166450500488281, 4.707275390625, 1.6759490966796875, 20.941619873046875, 6.052684783935547, 30.336891174316406, -0.3489837646484375, 18.988235473632812, 0.9945793151855469, 15.117881774902344, 27.634174346923828, 30.340675354003906, 15.809379577636719, 7.759647369384766, 29.179702758789062, -0.21942901611328125, 6.986686706542969, 13.489112854003906], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000240.npy"}
{"epoch": 0.36281179138321995, "step": 241, "batch_size": 64, "mean": 6.35860013961792, "std": 13.402851104736328, "min": -24.952796936035156, "p10": -8.972261428833008, "median": 3.8206939697265625, "p90": 26.727293205261237, "max": 33.5836181640625, "pos_frac": 0.625, "sample": [-5.858154296875, -0.3274574279785156, 28.440837860107422, -13.224861145019531, 25.085609436035156, -10.518402099609375, -0.2078399658203125, 12.440277099609375, 8.564716339111328, 15.636611938476562, -1.48541259765625, 32.152809143066406, -9.065364837646484, -1.1284103393554688, -0.73736572265625, 24.44091796875, 4.203685760498047, 5.464422225952148, 3.505279541015625, 24.358253479003906, 16.184329986572266, 12.271774291992188, -6.5382843017578125, 1.0967025756835938, 7.882240295410156, 0.4781827926635742, 12.881233215332031, 2.3512115478515625, -2.9375, -2.8109664916992188, 8.559898376464844, 11.227493286132812, 29.909591674804688, 0.8900947570800781, -2.2819442749023438, -12.872161865234375, -3.5587234497070312, 23.047195434570312, 5.901519775390625, 5.915779113769531, -24.952796936035156, 0.5864486694335938, -1.140268325805664, 33.5836181640625, 19.01421546936035, 2.939472198486328, -8.755020141601562, 30.799888610839844, 6.564605712890625, -20.116378784179688, 16.500732421875, 4.1361083984375, 2.5241622924804688, 18.05877685546875, 27.430871963500977, -11.32098388671875, -2.4922943115234375, 17.736194610595703, 28.8984375, 5.392827987670898, -6.489145278930664, 23.510690689086914, -4.368326187133789, -0.4292449951171875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000241.npy"}
{"epoch": 0.36432350718065004, "step": 242, "batch_size": 64, "mean": 10.372234344482422, "std": 15.015217781066895, "min": -27.931777954101562, "p10": -5.434691619873044, "median": 7.508628845214844, "p90": 32.07325553894043, "max": 47.0008544921875, "pos_frac": 0.8125, "sample": [-9.233114242553711, 4.678764343261719, 19.697250366210938, 10.000171661376953, 15.716300964355469, -23.339881896972656, 34.22929763793945, 12.714790344238281, 28.5743408203125, 27.640655517578125, 12.071250915527344, 5.6225738525390625, -2.8921127319335938, 32.406044006347656, 39.16002655029297, -0.194366455078125, 31.296749114990234, -8.362823486328125, 3.5419769287109375, 12.936351776123047, -0.27801513671875, 2.9149398803710938, 31.092100143432617, 14.64509391784668, 4.8133697509765625, -6.5243682861328125, 6.332489013671875, 14.203243255615234, 17.325748443603516, 16.99090576171875, 7.260932922363281, 0.4516181945800781, 47.0008544921875, -0.44643402099609375, 7.8441619873046875, 4.7853240966796875, 5.834868431091309, 1.5211868286132812, 39.10929870605469, 0.3196239471435547, -10.808479309082031, 19.354934692382812, 7.756324768066406, 1.2493934631347656, 29.02142333984375, 1.5377197265625, 9.192180633544922, 6.948699951171875, 16.67913818359375, 8.491464614868164, 36.710289001464844, 13.549858093261719, 35.503692626953125, 2.3597888946533203, -0.19854736328125, 6.377105712890625, 16.991455078125, 2.2048988342285156, 6.4812164306640625, -27.931777954101562, 23.81146240234375, 2.5737838745117188, -19.96595001220703, 14.471733093261719], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000242.npy"}
{"epoch": 0.36583522297808013, "step": 243, "batch_size": 64, "mean": 7.908749580383301, "std": 14.902213096618652, "min": -24.72003173828125, "p10": -9.875531387329099, "median": 6.175165176391602, "p90": 28.661443328857423, "max": 33.838985443115234, "pos_frac": 0.703125, "sample": [6.776454925537109, 28.691200256347656, -2.2349014282226562, -23.507339477539062, 33.838985443115234, 12.441566467285156, -10.907730102539062, 9.145431518554688, 23.42876434326172, -5.5941925048828125, 26.785263061523438, 10.284568786621094, 2.5318603515625, 5.200981140136719, 4.643089294433594, 26.404396057128906, 0.6851711273193359, -18.593093872070312, 30.506183624267578, 7.1927337646484375, -3.9576339721679688, 31.257461547851562, 8.689311981201172, -0.953094482421875, -4.815521240234375, 8.748306274414062, -1.6713027954101562, 27.765687942504883, 27.93993377685547, 22.31244659423828, -3.437957763671875, -17.104049682617188, 2.8071746826171875, -17.081390380859375, 12.361106872558594, 7.68316650390625, 30.59991455078125, 33.23253631591797, -24.72003173828125, 13.858070373535156, 28.16498565673828, -1.5775604248046875, 13.195327758789062, 33.5791015625, -4.052581787109375, 2.5615234375, -3.469493865966797, 2.055652618408203, 28.592010498046875, 22.469131469726562, 9.200958251953125, 6.60614013671875, 1.1051712036132812, 24.606277465820312, -7.371227264404297, 14.538345336914062, 5.744190216064453, 3.7754440307617188, 5.223234176635742, -7.467067718505859, -11.591171264648438, 13.490293502807617, 2.5752639770507812, 2.9724884033203125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000243.npy"}
{"epoch": 0.3673469387755102, "step": 244, "batch_size": 64, "mean": 13.635225296020508, "std": 14.014012336730957, "min": -8.46500015258789, "p10": -2.8255424499511714, "median": 11.074703216552734, "p90": 35.718017578125014, "max": 44.425750732421875, "pos_frac": 0.84375, "sample": [10.128074645996094, 37.01976013183594, 6.456443786621094, 2.3626556396484375, -4.290252685546875, 14.429040908813477, -1.8672733306884766, 25.182586669921875, 5.419319152832031, -6.442848205566406, -2.5324249267578125, 19.812965393066406, 5.0991668701171875, 32.68061828613281, 11.44781494140625, -8.085968017578125, 25.441200256347656, 38.156768798828125, -2.9511642456054688, 23.868410110473633, 17.431686401367188, 25.134967803955078, -8.32675552368164, 26.272884368896484, 4.232292175292969, 38.55705261230469, 14.510990142822266, 14.154487609863281, 42.58988952636719, 6.441080093383789, 41.49700927734375, 3.9868087768554688, 4.134004592895508, -5.875572204589844, -8.46500015258789, 6.087398529052734, -1.9819145202636719, 28.533573150634766, 30.890029907226562, 44.425750732421875, 0.8146047592163086, 43.0439453125, 1.4991331100463867, 3.9318923950195312, 3.215667724609375, 25.112083435058594, 16.759429931640625, 10.701591491699219, 3.3799819946289062, 14.164627075195312, 6.579492568969727, 21.652450561523438, 5.098821640014648, 15.459678649902344, 11.538703918457031, 29.215572357177734, 21.80081558227539, 5.5459747314453125, 19.329505920410156, 7.734352111816406, 10.577896118164062, 7.137655258178711, 18.370365142822266, 14.424636840820312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000244.npy"}
{"epoch": 0.3688586545729403, "step": 245, "batch_size": 64, "mean": 12.468271255493164, "std": 15.683525085449219, "min": -25.29840087890625, "p10": -4.067625045776365, "median": 10.43648910522461, "p90": 34.81762466430664, "max": 43.545677185058594, "pos_frac": 0.828125, "sample": [11.754295349121094, 10.2750244140625, 10.903533935546875, 39.787567138671875, 25.635440826416016, -4.643314361572266, 1.400482177734375, 16.55374526977539, 3.195903778076172, -0.8274688720703125, 0.7766189575195312, 4.50250244140625, 33.53644561767578, 12.091079711914062, 25.804412841796875, 26.44949722290039, -0.05344963073730469, 34.5325927734375, -25.29840087890625, 41.57536697387695, 36.14740753173828, 3.799346923828125, 8.063461303710938, 9.524456024169922, 32.46186828613281, 24.363235473632812, 8.935562133789062, -0.9311332702636719, 18.051101684570312, 1.8303508758544922, 13.983978271484375, -23.255935668945312, 1.9478378295898438, 22.269119262695312, 15.557991027832031, 16.99272918701172, -2.7243499755859375, -8.075675964355469, 20.41307830810547, 43.545677185058594, 17.31879425048828, 2.2121429443359375, 34.939781188964844, 1.1469497680664062, -6.571830749511719, 3.8987350463867188, 38.20759582519531, 37.27095413208008, 6.0277252197265625, -4.8929901123046875, 14.360794067382812, 19.53101348876953, 2.73333740234375, 8.163009643554688, 32.793922424316406, 0.9393081665039062, -17.776702880859375, 33.294761657714844, 1.7773208618164062, 1.2449188232421875, 29.951812744140625, 8.360275268554688, 11.587844848632812, 10.597953796386719], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000245.npy"}
{"epoch": 0.37037037037037035, "step": 246, "batch_size": 64, "mean": 11.289131164550781, "std": 16.662206649780273, "min": -28.314041137695312, "p10": -6.177889823913573, "median": 8.562347412109375, "p90": 32.31844253540039, "max": 52.916534423828125, "pos_frac": 0.71875, "sample": [26.080673217773438, 22.319950103759766, 40.928436279296875, 32.861549377441406, 30.603668212890625, 5.590705871582031, -5.087404251098633, 9.6295166015625, 4.47882080078125, -6.645240783691406, 12.861175537109375, 19.483642578125, 10.120925903320312, 3.4797096252441406, 29.945297241210938, -0.3589363098144531, 9.310966491699219, -9.03782844543457, -4.039924621582031, -1.03656005859375, 4.731842041015625, 35.838748931884766, -0.9178619384765625, 23.27259635925293, 45.43183898925781, 11.9139404296875, 30.9461669921875, 52.916534423828125, 0.697113037109375, 6.7868804931640625, 10.334270477294922, 24.54688262939453, 19.174415588378906, 5.668060302734375, -16.431640625, 16.08104705810547, -10.964385986328125, 10.317047119140625, 31.051193237304688, -4.903076171875, 30.146514892578125, -3.1507492065429688, -13.098548889160156, -2.3798446655273438, 6.154125213623047, -13.00042724609375, 20.12004852294922, 24.572650909423828, 5.5714874267578125, -28.314041137695312, -4.732025146484375, 14.769735336303711, 23.638771057128906, 46.27442932128906, 4.620708465576172, -2.6053466796875, 5.07396125793457, 42.00022888183594, -1.31036376953125, 9.11358642578125, 3.555389404296875, 8.0111083984375, 17.866596221923828, 1.6256256103515625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000246.npy"}
{"epoch": 0.37188208616780044, "step": 247, "batch_size": 64, "mean": 5.2804951667785645, "std": 15.90892219543457, "min": -36.50348663330078, "p10": -12.054013061523431, "median": 2.9738121032714844, "p90": 25.221311187744146, "max": 48.332763671875, "pos_frac": 0.59375, "sample": [-17.653900146484375, 14.170074462890625, 2.9506988525390625, 0.6537895202636719, 9.835197448730469, -4.507423400878906, 0.6194839477539062, 24.290496826171875, 39.033546447753906, -5.308990478515625, 15.414779663085938, 11.384628295898438, 30.136882781982422, 18.25269317626953, 2.0381622314453125, 21.93719482421875, 8.663280487060547, 17.532516479492188, -1.6795587539672852, -5.6659088134765625, 48.332763671875, 38.55755615234375, -27.26766586303711, -6.461322784423828, 33.60858917236328, -14.306869506835938, 8.585977554321289, -0.06629180908203125, 2.1056137084960938, 22.95555877685547, 10.837066650390625, -1.7709426879882812, 5.248504638671875, 3.7051239013671875, -4.337554931640625, -2.247425079345703, 19.40188980102539, -36.50348663330078, -2.1023788452148438, -1.454833984375, -19.480072021484375, -0.9348621368408203, -4.6709136962890625, 27.71392822265625, 14.13922119140625, 2.9969253540039062, -6.7973480224609375, -5.295654296875, 8.75125503540039, 3.3220367431640625, 1.4782257080078125, 10.833892822265625, -2.7083892822265625, 9.933578491210938, -20.264781951904297, 9.066768646240234, -0.7844753265380859, 11.924530029296875, -4.901145935058594, -25.16302490234375, -5.3535308837890625, 25.62023162841797, 9.373565673828125, 20.23421287536621], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000247.npy"}
{"epoch": 0.37339380196523053, "step": 248, "batch_size": 64, "mean": 9.78189468383789, "std": 17.25150489807129, "min": -21.355119705200195, "p10": -11.509613037109375, "median": 6.39500617980957, "p90": 34.942040634155276, "max": 45.138038635253906, "pos_frac": 0.71875, "sample": [3.355825424194336, 4.426475524902344, -8.5091552734375, -18.018798828125, 18.39063262939453, 6.505340576171875, 24.203826904296875, 17.54084587097168, 2.713756561279297, -4.687719345092773, -1.5115432739257812, 44.97522735595703, 5.006935119628906, -9.351261138916016, 19.780715942382812, 21.599910736083984, 26.13201904296875, -19.458267211914062, -20.541231155395508, 6.896076202392578, -11.98504638671875, 2.842731475830078, 25.56158447265625, 15.234151840209961, 24.460235595703125, 20.700286865234375, 3.3059234619140625, -0.11725616455078125, 9.709556579589844, 35.152862548828125, 0.086822509765625, 41.366912841796875, 5.106353759765625, -21.355119705200195, 36.59828186035156, 34.45012283325195, -3.0179519653320312, -11.1077880859375, 6.284671783447266, 9.23916244506836, 12.497169494628906, 29.140880584716797, 4.057531356811523, 45.13172912597656, 29.104705810546875, 14.777172088623047, 7.449892044067383, 1.4924850463867188, 14.580707550048828, 22.1429443359375, 45.138038635253906, 9.790054321289062, 21.03076171875, -8.142044067382812, -10.788494110107422, 2.1482467651367188, -11.68182373046875, 32.890235900878906, 37.18885803222656, 4.806610107421875, -0.01045989990234375, 0.35173797607421875, -15.776885986328125, -3.2449188232421875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000248.npy"}
{"epoch": 0.3749055177626606, "step": 249, "batch_size": 64, "mean": 3.5041306018829346, "std": 17.517961502075195, "min": -43.305999755859375, "p10": -22.274894142150877, "median": 5.695558547973633, "p90": 20.682758712768557, "max": 41.817466735839844, "pos_frac": 0.65625, "sample": [20.467525482177734, -10.6346435546875, 17.18334197998047, 15.469558715820312, 18.17265510559082, -11.2430419921875, -31.08551788330078, 30.918548583984375, 20.523895263671875, 6.876922607421875, 20.750843048095703, 5.761970520019531, 3.808818817138672, 5.629146575927734, -22.99988555908203, -21.9154052734375, -1.5415802001953125, -36.15509033203125, -9.701095581054688, -22.4289608001709, 4.544208526611328, -0.686614990234375, -20.729522705078125, 8.267049789428711, 4.759330749511719, 6.721855163574219, 0.13338851928710938, -9.0164794921875, 23.090179443359375, -1.23358154296875, -9.188667297363281, 16.496490478515625, 2.323789596557617, -1.5393218994140625, 32.743927001953125, -28.526641845703125, 11.020011901855469, -11.406814575195312, 26.143388748168945, -15.305679321289062, 19.595752716064453, 41.817466735839844, 33.485450744628906, 18.961578369140625, 19.780792236328125, 11.85239028930664, -24.359031677246094, 7.1871490478515625, 8.965408325195312, 12.382278442382812, 4.969003677368164, 20.44415283203125, 7.9597625732421875, 0.8802986145019531, 16.080371856689453, 8.378829956054688, -12.497634887695312, 9.657501220703125, 15.289199829101562, -43.305999755859375, 4.431529998779297, -7.3159942626953125, 0.0352630615234375, 13.120567321777344], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000249.npy"}
{"epoch": 0.3764172335600907, "step": 250, "batch_size": 64, "mean": 11.716409683227539, "std": 18.000612258911133, "min": -50.65618896484375, "p10": -5.026941680908203, "median": 8.817682266235352, "p90": 35.84766464233399, "max": 53.89111328125, "pos_frac": 0.765625, "sample": [25.68427276611328, 1.2983627319335938, -3.617706298828125, 20.814987182617188, 8.426712036132812, 18.64453125, 27.18295669555664, 30.163360595703125, 36.597808837890625, 4.7704620361328125, 37.670555114746094, -25.57280731201172, 9.709648132324219, 3.8023529052734375, 12.246715545654297, 11.237668991088867, -2.142974853515625, -50.65618896484375, 17.424762725830078, 5.7008514404296875, 14.422447204589844, -9.742984771728516, 5.731163024902344, -23.697725296020508, 17.06610107421875, 8.5992431640625, 40.17626953125, 26.356319427490234, 6.448783874511719, 33.51277160644531, -1.5034599304199219, 5.85443115234375, 29.39110565185547, -7.281059265136719, 6.9417266845703125, 12.27993392944336, 9.036121368408203, 22.965496063232422, 50.8690185546875, -0.130706787109375, 12.192695617675781, -1.435760498046875, 29.058818817138672, -5.078163146972656, 40.45970153808594, 5.428150177001953, 42.03668212890625, 30.233285903930664, 7.603336334228516, 34.097328186035156, -0.4398956298828125, 5.1923065185546875, 2.5963058471679688, 17.190067291259766, -7.71051025390625, 10.385505676269531, 53.89111328125, 14.970130920410156, 4.247657775878906, -1.232217788696289, -4.9074249267578125, 15.407455444335938, 4.05718994140625, 4.925128936767578], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000250.npy"}
{"epoch": 0.3779289493575208, "step": 251, "batch_size": 64, "mean": 7.54291296005249, "std": 16.114498138427734, "min": -26.218780517578125, "p10": -11.122315979003906, "median": 5.775676727294922, "p90": 25.914469146728518, "max": 49.468299865722656, "pos_frac": 0.703125, "sample": [2.8571348190307617, 32.75855255126953, 49.468299865722656, 1.0011577606201172, 8.2054443359375, 4.228832244873047, -21.544876098632812, 1.3665847778320312, -1.0321273803710938, 47.34141540527344, -2.163482666015625, -26.218780517578125, 12.699371337890625, -5.519813537597656, 1.511444091796875, -25.394548416137695, -1.5012931823730469, 25.5792236328125, -3.4042739868164062, -11.210678100585938, 5.3141632080078125, 44.80110168457031, 1.7248382568359375, -10.9161376953125, 18.593276977539062, -2.5940170288085938, 16.979896545410156, 23.312103271484375, -6.428306579589844, 2.586193084716797, 26.02881622314453, 6.440040588378906, 2.0377731323242188, 6.605839729309082, 16.68822479248047, 15.552734375, 19.913719177246094, 20.9476318359375, -16.554161071777344, -10.14358901977539, 10.347763061523438, 4.04949951171875, 14.784072875976562, 9.08742904663086, 2.8340835571289062, 6.237190246582031, 7.1891632080078125, 18.312576293945312, 9.57781982421875, 37.90069580078125, 20.653579711914062, 0.9162826538085938, 22.359092712402344, -16.82317352294922, 26.19293212890625, -2.128538131713867, 5.144220352172852, -18.235450744628906, 14.64462661743164, -0.36884307861328125, -4.462650299072266, 7.946563720703125, 25.647659301757812, 11.022087097167969], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000251.npy"}
{"epoch": 0.3794406651549509, "step": 252, "batch_size": 64, "mean": 7.735276222229004, "std": 17.40585708618164, "min": -42.80322265625, "p10": -10.671715927124021, "median": 8.317535400390625, "p90": 32.30663890838624, "max": 45.397857666015625, "pos_frac": 0.765625, "sample": [8.949705123901367, -27.985565185546875, -2.508014678955078, -18.1282958984375, 34.818519592285156, 11.153465270996094, 3.2549285888671875, 24.0887451171875, 12.613923072814941, 28.276657104492188, 36.800933837890625, 14.169818878173828, 8.2982177734375, 4.5802459716796875, 37.505916595458984, 8.477005004882812, 33.51637268066406, 2.972665786743164, 2.9327774047851562, 0.029626846313476562, -5.276023864746094, 21.635704040527344, 20.093048095703125, 14.219915390014648, 16.519668579101562, 1.0050201416015625, 0.0277252197265625, -33.49993896484375, 17.05420684814453, 11.48733901977539, 10.257471084594727, 11.247322082519531, 29.48392677307129, 3.1105117797851562, 12.377609252929688, 4.9259490966796875, 5.05865478515625, 16.599132537841797, 10.214656829833984, -4.5966796875, 16.917282104492188, -2.2619400024414062, -1.5201797485351562, 17.394914627075195, 1.4398422241210938, 45.397857666015625, -5.023017883300781, -5.871818542480469, -42.80322265625, -23.75415802001953, -7.685415267944336, 9.403884887695312, 0.58197021484375, -11.951559066772461, 4.407096862792969, 10.373481750488281, 8.33685302734375, 42.644142150878906, 5.692268371582031, 36.72306823730469, 29.1163330078125, 2.95123291015625, -23.85858726501465, 2.6444854736328125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000252.npy"}
{"epoch": 0.38095238095238093, "step": 253, "batch_size": 64, "mean": 6.770212173461914, "std": 16.735166549682617, "min": -22.672809600830078, "p10": -11.190742111206054, "median": 3.5984153747558594, "p90": 34.72695293426516, "max": 48.6522216796875, "pos_frac": 0.640625, "sample": [19.689430236816406, -3.5415802001953125, 45.198638916015625, 6.494926452636719, 7.898525238037109, 5.61175537109375, 1.7557525634765625, 29.966001510620117, 25.29047393798828, -22.672809600830078, 5.759532928466797, -0.567413330078125, 36.76736068725586, -0.4211902618408203, 2.7004852294921875, -15.316436767578125, -13.943168640136719, 28.318519592285156, -8.929649353027344, 3.498443603515625, 1.043243408203125, 14.32059097290039, -8.441825866699219, 23.10821533203125, 13.312446594238281, -9.410629272460938, 38.78812789916992, 6.937690734863281, 1.9058685302734375, 15.609893798828125, 2.1964263916015625, -0.8713455200195312, 11.979133605957031, 3.6983871459960938, 4.37103271484375, 48.6522216796875, -5.1741943359375, -6.071308135986328, 12.322479248046875, 11.571868896484375, 10.484184265136719, -19.505630493164062, 2.8805999755859375, -17.928937911987305, 15.487777709960938, -11.44036865234375, -10.608280181884766, 5.415885925292969, 44.0612907409668, -0.12627792358398438, 14.541595458984375, 8.470817565917969, 45.21217346191406, 0.7697296142578125, 13.066535949707031, -9.222862243652344, -17.646034240722656, -5.544948577880859, 37.439537048339844, -5.864463806152344, 0.6113128662109375, -7.019191741943359, 6.8273468017578125, -0.47414398193359375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000253.npy"}
{"epoch": 0.382464096749811, "step": 254, "batch_size": 64, "mean": 9.789339065551758, "std": 15.839370727539062, "min": -19.214099884033203, "p10": -7.361006546020508, "median": 7.197768211364746, "p90": 33.14683837890625, "max": 46.977970123291016, "pos_frac": 0.65625, "sample": [-8.71072769165039, -3.0764007568359375, 23.23822021484375, -3.288238525390625, 5.07269287109375, 19.221275329589844, 37.07083511352539, 24.19434356689453, 5.381551742553711, 1.4958877563476562, -0.8155059814453125, 32.13676452636719, 9.838066101074219, 20.830373764038086, 14.991851806640625, 20.933494567871094, 45.210968017578125, 29.776626586914062, -7.0575408935546875, 7.276947021484375, -3.9181900024414062, 45.667633056640625, -1.6371688842773438, 33.57972717285156, -2.815521240234375, -13.068656921386719, 14.320159912109375, 34.976417541503906, 25.74765396118164, -6.610221862792969, 46.977970123291016, -3.2235565185546875, 0.569244384765625, -0.5672836303710938, -19.214099884033203, 24.71373748779297, -6.201877593994141, 25.84469223022461, 1.156463623046875, 14.729560852050781, 14.943946838378906, -5.642110824584961, 8.686614990234375, -9.565727233886719, -7.191173553466797, 41.478851318359375, -1.0138587951660156, 11.732242584228516, 13.817276000976562, 3.6118526458740234, -8.8729248046875, -9.004104614257812, 5.197746276855469, 21.8436279296875, 19.409530639648438, 8.973976135253906, 3.040283203125, 7.118589401245117, 15.70452880859375, 3.1067123413085938, 8.905128479003906, 9.410675048828125, -7.4337921142578125, -6.488395690917969], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000254.npy"}
{"epoch": 0.3839758125472411, "step": 255, "batch_size": 64, "mean": 9.045727729797363, "std": 17.449893951416016, "min": -33.977962493896484, "p10": -10.957308959960937, "median": 9.517522811889648, "p90": 30.967522430419923, "max": 50.0877685546875, "pos_frac": 0.78125, "sample": [31.109756469726562, -31.585861206054688, 34.288978576660156, 21.275665283203125, 21.842140197753906, 34.994110107421875, -0.18273162841796875, -2.6485977172851562, 1.7514801025390625, 24.215164184570312, 0.09136962890625, -4.552900314331055, 19.906795501708984, 26.380783081054688, -33.977962493896484, 2.94268798828125, 9.056648254394531, 9.018539428710938, 9.415931701660156, -31.87009048461914, 29.215858459472656, 17.898208618164062, 8.866737365722656, 4.628623962402344, 13.938278198242188, 2.0439453125, -18.759475708007812, 6.073726654052734, 39.59667205810547, 4.255546569824219, 18.665298461914062, 12.094963073730469, -11.465385437011719, 9.201248168945312, 12.38958740234375, 15.306591033935547, 9.627655029296875, 14.479522705078125, 36.9492301940918, 9.90755844116211, 3.597728729248047, 11.996627807617188, 23.411651611328125, -3.9180755615234375, 6.977134704589844, -8.250951766967773, 5.73565673828125, 50.0877685546875, -22.1683349609375, 19.746246337890625, 34.6614990234375, 10.147003173828125, 4.4217376708984375, 30.635643005371094, 16.496261596679688, 7.127372741699219, 12.317741394042969, 24.646533966064453, 13.835372924804688, -30.08544921875, -9.361122131347656, 0.634918212890625, 9.61911392211914, -9.771797180175781], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000255.npy"}
{"epoch": 0.3854875283446712, "step": 256, "batch_size": 64, "mean": 10.635353088378906, "std": 18.025094985961914, "min": -28.619163513183594, "p10": -6.417193603515625, "median": 5.570352554321289, "p90": 38.0760856628418, "max": 42.046844482421875, "pos_frac": 0.71875, "sample": [-5.275428771972656, -0.9393310546875, 12.748128890991211, 33.82561492919922, 25.507492065429688, 3.6393585205078125, -28.619163513183594, 34.78889465332031, 1.3140907287597656, -5.309726715087891, -19.295387268066406, 8.071523666381836, -6.399932861328125, 38.26697540283203, -2.07452392578125, 40.81146240234375, 17.50609588623047, 3.583709716796875, -18.601097106933594, 0.27819061279296875, -8.24892807006836, 36.37693786621094, 3.216724395751953, 9.990829467773438, -6.424591064453125, 4.259361267089844, 19.157291412353516, -2.2354660034179688, 8.706966400146484, 12.139068603515625, -2.5431747436523438, 23.039939880371094, 4.585624694824219, 33.69348907470703, 6.743694305419922, 16.700546264648438, 39.08885192871094, 27.116886138916016, 2.630615234375, 39.300025939941406, 2.18487548828125, 0.3283958435058594, 42.046844482421875, -19.399795532226562, 40.028099060058594, 12.074626922607422, 32.83882141113281, 2.3968887329101562, 2.632892608642578, 30.63360595703125, -0.6785049438476562, 32.94532012939453, 6.34368896484375, 8.292861938476562, 30.205368041992188, -24.96414566040039, 37.63067626953125, -3.1447830200195312, 1.5675716400146484, 38.82977294921875, -4.654697418212891, 8.194366455078125, -1.588775634765625, 4.797016143798828], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000256.npy"}
{"epoch": 0.3869992441421013, "step": 257, "batch_size": 64, "mean": 8.97639274597168, "std": 19.02106475830078, "min": -41.835304260253906, "p10": -13.765117263793945, "median": 4.859977722167969, "p90": 37.29662170410157, "max": 52.41889572143555, "pos_frac": 0.734375, "sample": [0.24330806732177734, 2.8239974975585938, -1.644775390625, 32.003662109375, 43.519981384277344, 13.056526184082031, -5.4809112548828125, 38.66352844238281, 15.394126892089844, 42.017677307128906, 30.317981719970703, 4.567283630371094, 15.701190948486328, -1.5363082885742188, 24.298629760742188, 33.14903259277344, 38.4849853515625, 5.078117370605469, -16.3128662109375, 1.5087203979492188, 0.32563018798828125, 4.641838073730469, -13.097686767578125, -20.97591209411621, 28.708404541015625, -25.556838989257812, 0.5663032531738281, 10.540328979492188, 15.875877380371094, -5.646781921386719, 17.92023468017578, 37.5592041015625, 2.0915069580078125, -7.679435729980469, 15.28887939453125, 4.624855041503906, -7.9459228515625, 20.960182189941406, 33.50444793701172, -23.291587829589844, -41.835304260253906, 36.683929443359375, 14.725276947021484, 5.586151123046875, -9.643630981445312, -14.051158905029297, 7.443275451660156, 0.21443939208984375, 18.936378479003906, 5.591339111328125, 20.164283752441406, 5.334327697753906, 38.488372802734375, 0.4405250549316406, -11.51824951171875, 52.41889572143555, 2.1556739807128906, -17.162132263183594, 24.90814208984375, 2.4828338623046875, 21.776735305786133, 4.452293395996094, 3.3346099853515625, -0.705291748046875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000257.npy"}
{"epoch": 0.3885109599395314, "step": 258, "batch_size": 64, "mean": 10.682682991027832, "std": 14.06686019897461, "min": -25.127357482910156, "p10": -5.085285568237304, "median": 7.623222351074219, "p90": 30.864097595214858, "max": 45.78828430175781, "pos_frac": 0.8125, "sample": [12.481704711914062, 5.370429992675781, 13.857925415039062, -1.6679916381835938, 6.1002349853515625, 40.096336364746094, 34.836448669433594, 4.487060546875, 23.16796875, -4.236011505126953, 17.577075958251953, 17.419105529785156, -7.9078826904296875, 14.683492660522461, 16.158287048339844, 7.0960693359375, -10.303516387939453, -6.698753356933594, 3.1922073364257812, -4.4505767822265625, 13.37106704711914, -0.5835723876953125, -25.127357482910156, 27.610458374023438, 4.8251953125, 9.81631088256836, 7.536048889160156, 5.207013130187988, 4.409782409667969, 6.630088806152344, 13.570083618164062, 1.6511497497558594, 37.51898193359375, 24.90245819091797, 4.672492980957031, -5.489311218261719, -0.13927841186523438, 7.710395812988281, 6.8824462890625, 45.78828430175781, 10.380577087402344, 2.2704010009765625, 15.655082702636719, 5.4985504150390625, 18.501190185546875, 32.85171890258789, 22.867385864257812, 13.348312377929688, 4.0229949951171875, 27.205814361572266, 4.129631042480469, 3.5121078491210938, 20.159217834472656, 8.23675537109375, 17.87429428100586, 7.340930938720703, 17.13629150390625, -24.667861938476562, 5.200428009033203, -5.357303619384766, 32.258514404296875, 17.463603973388672, 43.03439712524414, 12.746307373046875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000258.npy"}
{"epoch": 0.3900226757369615, "step": 259, "batch_size": 64, "mean": 15.510651588439941, "std": 19.52655601501465, "min": -26.56768798828125, "p10": -10.581482124328613, "median": 15.270456314086914, "p90": 40.61719360351563, "max": 55.751373291015625, "pos_frac": 0.796875, "sample": [-11.184150695800781, -26.56768798828125, -22.252975463867188, 24.0416259765625, 9.577423095703125, 36.35222244262695, 36.356468200683594, 48.49180603027344, 17.13907241821289, -10.087234497070312, 37.24005889892578, 4.266670227050781, 18.748558044433594, 14.169784545898438, 19.883590698242188, -2.34783935546875, -10.793302536010742, -3.4379425048828125, 55.751373291015625, 17.914520263671875, 0.43896484375, 39.711090087890625, 43.9320182800293, 37.487693786621094, 12.8990478515625, 1.2949600219726562, 1.6861038208007812, 25.615779876708984, 10.449790954589844, 9.195976257324219, 15.951953887939453, 16.09649658203125, -12.200923919677734, 17.256210327148438, 15.003910064697266, -9.397941589355469, -1.1600341796875, 6.0600433349609375, 34.233856201171875, 16.594268798828125, 39.03520584106445, 17.929065704345703, 4.488422393798828, 44.97978210449219, 5.162162780761719, 8.020233154296875, -20.318130493164062, -3.4231529235839844, -19.844573974609375, 5.7961578369140625, 46.15156555175781, 33.5555419921875, 37.16729736328125, 9.906005859375, 41.005523681640625, 38.52748107910156, 18.906280517578125, 4.469841003417969, 5.299106597900391, 13.95330810546875, 42.1583251953125, 37.829017639160156, 31.978919982910156, 15.537002563476562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000259.npy"}
{"epoch": 0.3915343915343915, "step": 260, "batch_size": 64, "mean": 9.910329818725586, "std": 21.76373291015625, "min": -35.57168197631836, "p10": -20.0091194152832, "median": 7.507120132446289, "p90": 41.71364135742188, "max": 50.429664611816406, "pos_frac": 0.71875, "sample": [31.189746856689453, -32.673545837402344, 23.576946258544922, -5.246978759765625, -7.644134521484375, 36.063690185546875, 2.962005615234375, 3.1226425170898438, 46.811492919921875, -14.423355102539062, 43.17451477050781, 26.101221084594727, 40.444976806640625, 0.334381103515625, 10.439674377441406, -2.390167236328125, 24.695579528808594, 5.625659942626953, 8.870759963989258, 50.429664611816406, 21.71856689453125, 5.006195068359375, 45.63359069824219, 32.67094421386719, 0.2735939025878906, 5.039760589599609, 26.890869140625, 45.76021957397461, -2.4373550415039062, 0.9242095947265625, 23.587078094482422, 18.219573974609375, 6.7465057373046875, -21.145904541015625, -1.9867057800292969, 25.74036407470703, -0.052032470703125, -17.35662078857422, -34.898582458496094, 5.7378387451171875, -0.6979446411132812, 5.671886444091797, 8.26773452758789, 9.820945739746094, 4.783599853515625, 16.159198760986328, 20.860469818115234, -16.768970489501953, 43.400169372558594, 42.257354736328125, -25.056133270263672, 19.085203170776367, 8.930068969726562, 8.28619384765625, 29.74407196044922, 30.652626037597656, -0.4752197265625, 29.673690795898438, 18.632186889648438, -35.57168197631836, 2.764263153076172, -35.23919677734375, -31.037261962890625, 2.5809707641601562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000260.npy"}
{"epoch": 0.3930461073318216, "step": 261, "batch_size": 64, "mean": 7.944557189941406, "std": 19.290058135986328, "min": -41.23272705078125, "p10": -17.33945846557617, "median": 4.591554641723633, "p90": 34.7197853088379, "max": 51.961570739746094, "pos_frac": 0.6875, "sample": [-1.11572265625, 5.39727783203125, 5.310131072998047, -3.3093948364257812, -15.075881958007812, 30.60161781311035, 17.322664260864258, 32.40548324584961, 4.341907501220703, -1.9473838806152344, -19.157878875732422, -21.206558227539062, -0.2991485595703125, 35.36827850341797, 26.683120727539062, -5.137054443359375, 1.5292282104492188, 46.709381103515625, 0.4746742248535156, 51.40654754638672, 25.070877075195312, 12.686332702636719, -5.088020324707031, -14.916183471679688, 17.763641357421875, -7.3560943603515625, 51.961570739746094, 15.367107391357422, -2.0894317626953125, 4.8412017822265625, 8.017967224121094, 14.683055877685547, 21.109283447265625, 38.637725830078125, 2.5347862243652344, 0.16309356689453125, 44.27079772949219, -41.23272705078125, -22.63990020751953, 16.10809326171875, 16.764915466308594, 2.6143951416015625, 21.51927947998047, 19.579193115234375, 16.541545867919922, -0.15183258056640625, -2.743000030517578, -29.915725708007812, -22.617900848388672, 7.035606384277344, 7.8858489990234375, -2.0952529907226562, 33.206634521484375, 41.24073028564453, 0.162628173828125, -18.30956268310547, 7.647544860839844, 2.5482559204101562, 0.88555908203125, 2.7321929931640625, 10.519821166992188, 2.868816375732422, 17.615402221679688, 2.7220935821533203], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000261.npy"}
{"epoch": 0.3945578231292517, "step": 262, "batch_size": 64, "mean": 10.531760215759277, "std": 16.439531326293945, "min": -23.57681655883789, "p10": -6.923040580749512, "median": 7.692777633666992, "p90": 33.61746673583985, "max": 51.352691650390625, "pos_frac": 0.734375, "sample": [-19.29205322265625, 15.723564147949219, 12.710552215576172, -0.24308204650878906, 4.107818603515625, 20.43016242980957, 9.849761962890625, 0.11008453369140625, 28.683719635009766, 33.31840515136719, -0.00617218017578125, 40.52439880371094, -7.003929138183594, 11.241527557373047, 37.51659393310547, -23.57681655883789, 9.531135559082031, 23.099037170410156, 8.914093017578125, 13.93038558959961, -15.442779541015625, 3.7102127075195312, -0.25933837890625, 3.438323974609375, 27.417648315429688, 29.988609313964844, 4.907806396484375, 25.22420883178711, -13.383895874023438, -0.5588245391845703, 33.745635986328125, -2.1615371704101562, 3.3038673400878906, 33.836605072021484, -18.614776611328125, 11.656875610351562, 5.967105865478516, 18.439361572265625, 30.247650146484375, 3.1083984375, -5.996639251708984, 0.774627685546875, -6.73430061340332, 28.375350952148438, 51.352691650390625, 23.987205505371094, 10.131202697753906, 6.471462249755859, -4.8997344970703125, 4.745674133300781, 14.304962158203125, 1.7586898803710938, 42.850433349609375, 11.221065521240234, 5.9785003662109375, -0.6948013305664062, -1.41729736328125, 45.9644775390625, 5.2605438232421875, 13.06500244140625, 13.969810485839844, 20.268783569335938, 3.5152969360351562, -14.3607177734375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000262.npy"}
{"epoch": 0.3960695389266818, "step": 263, "batch_size": 64, "mean": 11.31068229675293, "std": 18.651948928833008, "min": -38.427032470703125, "p10": -10.323153877258301, "median": 8.121967315673828, "p90": 35.957148742675784, "max": 53.234107971191406, "pos_frac": 0.734375, "sample": [-1.3913078308105469, 36.10139465332031, -11.458518981933594, 1.0475311279296875, 7.152885437011719, 2.0911407470703125, 17.495223999023438, 12.0849609375, 27.76239776611328, 20.921585083007812, 31.58319091796875, 7.544239044189453, -7.259769439697266, 16.638954162597656, 1.4002532958984375, -14.93661880493164, 26.6551513671875, 1.3588714599609375, 0.4737548828125, -5.563697814941406, 18.695648193359375, 29.59540557861328, 24.262451171875, 19.208740234375, -22.131938934326172, 4.781246185302734, 50.08392333984375, 28.068199157714844, 14.707427978515625, 19.22710418701172, 46.65217590332031, 10.709625244140625, -5.2873077392578125, 11.927841186523438, -9.973478317260742, 1.6925048828125, -12.79935073852539, 43.38652038574219, 3.785114288330078, 23.868682861328125, -1.3067626953125, -3.25958251953125, 16.553314208984375, 35.620574951171875, -6.45184326171875, 4.312080383300781, 4.7946624755859375, 41.08043670654297, 18.453140258789062, 6.776008605957031, -17.259307861328125, 53.234107971191406, 34.66246032714844, -4.188232421875, 45.551727294921875, -0.20670700073242188, 13.916534423828125, 8.699695587158203, 2.8271141052246094, -10.473014831542969, 0.4177093505859375, -38.427032470703125, 25.10210418701172, 23.292339324951172], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000263.npy"}
{"epoch": 0.3975812547241119, "step": 264, "batch_size": 64, "mean": 11.853877067565918, "std": 17.446550369262695, "min": -20.775854110717773, "p10": -6.05789566040039, "median": 7.796092987060547, "p90": 37.14591827392578, "max": 61.741851806640625, "pos_frac": 0.71875, "sample": [29.0672607421875, -6.440826416015625, 1.6707305908203125, 17.05301856994629, 37.70708084106445, -2.1457176208496094, 5.9235382080078125, 0.05336761474609375, 25.723094940185547, 18.89191436767578, -20.775854110717773, -2.212902069091797, -17.675048828125, 14.285293579101562, -0.49399566650390625, -1.84039306640625, 1.9556312561035156, 22.664825439453125, 2.048877716064453, 16.734397888183594, 14.806045532226562, 27.32186508178711, 17.497711181640625, 17.546951293945312, 0.9883804321289062, 51.29147720336914, -1.159317970275879, 21.393203735351562, 0.1474456787109375, 37.124855041503906, 22.71009063720703, 7.7494049072265625, 36.97393798828125, 5.514148712158203, 9.420204162597656, 1.9412593841552734, 37.154945373535156, -4.3160858154296875, -2.4763107299804688, 46.962677001953125, 7.842781066894531, -13.0423583984375, -2.599945068359375, 18.864730834960938, 31.56743621826172, 19.25141143798828, -3.54876708984375, 21.7125244140625, -15.465362548828125, 5.0516357421875, -0.0839691162109375, 4.138618469238281, 46.4575080871582, 10.71075439453125, 5.917579650878906, -7.092353820800781, 5.992927551269531, 8.323150634765625, 16.663742065429688, -5.164390563964844, -8.994308471679688, 61.741851806640625, 38.87499237060547, 20.74078369140625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000264.npy"}
{"epoch": 0.39909297052154197, "step": 265, "batch_size": 64, "mean": 14.966060638427734, "std": 21.78093719482422, "min": -37.956703186035156, "p10": -16.380084991455078, "median": 15.236894607543945, "p90": 43.56260757446289, "max": 53.629547119140625, "pos_frac": 0.78125, "sample": [-20.3497314453125, 11.302389144897461, -10.34170150756836, 0.7539520263671875, 22.63127899169922, 1.637847900390625, 24.25050926208496, 15.525482177734375, -14.759674072265625, 53.56085205078125, 5.883705139160156, 30.81964111328125, 13.59832763671875, 17.608386993408203, -4.3560791015625, 1.1108551025390625, 16.658145904541016, 6.904933929443359, 12.99502182006836, -37.956703186035156, -18.148921966552734, 34.7799072265625, 45.83268737792969, 14.948307037353516, 2.1563358306884766, 8.416007995605469, 28.86347198486328, 3.081623077392578, 24.407379150390625, 8.573600769042969, 42.3514289855957, 34.34052276611328, -13.687629699707031, 47.346763610839844, 25.347427368164062, 53.629547119140625, 9.191354751586914, 10.095993041992188, -9.795116424560547, -0.23287200927734375, 17.392562866210938, 40.26573181152344, 24.73552703857422, 40.97991943359375, 53.24082946777344, 29.216232299804688, 43.71056365966797, 20.902130126953125, 17.537109375, -24.394123077392578, 26.715301513671875, 7.7864532470703125, -0.9193572998046875, 41.20367431640625, 36.92041015625, 33.501869201660156, 43.62425231933594, -17.074546813964844, 8.80429458618164, -21.169876098632812, 43.41876983642578, 0.07495880126953125, 15.976585388183594, -23.596664428710938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000265.npy"}
{"epoch": 0.40060468631897206, "step": 266, "batch_size": 64, "mean": 12.542583465576172, "std": 19.502246856689453, "min": -40.78274154663086, "p10": -3.114633178710937, "median": 9.044336318969727, "p90": 38.965388488769534, "max": 59.78666687011719, "pos_frac": 0.796875, "sample": [-6.324577331542969, 30.213359832763672, 56.562232971191406, 5.048664093017578, 28.33129119873047, 0.6616172790527344, 3.8409271240234375, 2.243865966796875, 2.332019805908203, -24.048221588134766, 12.568721771240234, -2.3915557861328125, 1.5892524719238281, 9.83519172668457, 5.096477508544922, 22.649375915527344, -2.659637451171875, 56.345848083496094, -0.6720733642578125, 0.9417915344238281, 50.39475631713867, 12.365936279296875, 4.129173278808594, 35.53236770629883, 8.84835433959961, 39.99591064453125, -8.513885498046875, 10.738540649414062, 43.187828063964844, 24.61339569091797, 26.84478759765625, 3.04400634765625, 19.03899383544922, -3.30963134765625, 7.98695182800293, -2.5396575927734375, 12.244644165039062, 38.008758544921875, 38.99864959716797, -1.7594528198242188, 12.10772705078125, -6.7603759765625, 10.921356201171875, 8.18545150756836, -35.035308837890625, 8.381141662597656, 11.579818725585938, 3.854328155517578, 8.789047241210938, 25.077552795410156, -40.78274154663086, 10.00826644897461, -0.9523162841796875, 20.429229736328125, 22.547958374023438, 25.13625717163086, 59.78666687011719, 1.179586410522461, 32.3963623046875, 9.240318298339844, 38.887779235839844, 2.9702720642089844, 12.413314819335938, 0.34860992431640625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000266.npy"}
{"epoch": 0.4021164021164021, "step": 267, "batch_size": 64, "mean": 10.98253059387207, "std": 18.291147232055664, "min": -42.81858825683594, "p10": -12.729879760742186, "median": 9.421411514282227, "p90": 34.11010665893555, "max": 50.258819580078125, "pos_frac": 0.703125, "sample": [22.406280517578125, -0.7317028045654297, 4.09889030456543, -13.485248565673828, 10.123878479003906, -2.3785400390625, 23.787609100341797, 30.789230346679688, 34.74462890625, 8.718944549560547, 3.211406707763672, 2.888641357421875, 5.210334777832031, 39.37982177734375, 4.256032943725586, -21.359031677246094, 4.013092041015625, -1.249969482421875, 1.91717529296875, -18.26553726196289, -8.508522033691406, -13.349334716796875, 4.8430328369140625, 7.22625732421875, 32.782562255859375, 47.845462799072266, -0.32623291015625, -7.814409255981445, 27.653533935546875, -17.076183319091797, 22.53778839111328, 13.171855926513672, 39.27680587768555, 11.302352905273438, -0.5274391174316406, 2.57208251953125, 31.079376220703125, -12.00140380859375, 2.8755035400390625, 44.95054626464844, 26.44317626953125, -0.0938568115234375, 50.258819580078125, 1.6919898986816406, 27.41100311279297, 18.682504653930664, 34.679054260253906, 14.522514343261719, -42.81858825683594, 26.483257293701172, 13.1900634765625, -13.042083740234375, -2.8374156951904297, 25.994827270507812, 21.042858123779297, -0.40927886962890625, 25.418563842773438, 24.91431427001953, 23.548690795898438, -6.196380615234375, 14.783815383911133, 22.11648178100586, 17.873992919921875, 12.634042739868164], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000267.npy"}
{"epoch": 0.4036281179138322, "step": 268, "batch_size": 64, "mean": 7.760472297668457, "std": 18.307710647583008, "min": -24.080230712890625, "p10": -8.542910957336426, "median": 1.9557952880859375, "p90": 37.392390441894534, "max": 56.5894775390625, "pos_frac": 0.59375, "sample": [-3.2699127197265625, 47.86500549316406, 7.551059722900391, -20.745254516601562, 1.784637451171875, 25.591224670410156, -4.2635650634765625, -24.080230712890625, -16.090076446533203, 36.808067321777344, 25.725082397460938, 0.868621826171875, -3.6004295349121094, 4.764514923095703, 11.105056762695312, -5.931854248046875, 1.9065055847167969, -18.434539794921875, -8.65803337097168, 37.64281463623047, 31.098987579345703, 4.983894348144531, -3.2757740020751953, -4.286140441894531, -4.6897735595703125, 7.5433349609375, -0.8039321899414062, 24.76685333251953, 9.106315612792969, -4.137397766113281, 11.79876708984375, -2.4062423706054688, -2.204376220703125, 3.1686553955078125, -0.1155548095703125, 43.52888488769531, -1.0257797241210938, -16.07964324951172, -1.4923248291015625, 1.3180341720581055, 13.212661743164062, -1.6232223510742188, 14.114044189453125, 56.5894775390625, 51.55455017089844, 16.961395263671875, 29.48255157470703, -4.0760040283203125, 2.834197998046875, -23.524147033691406, -5.2040557861328125, 1.61187744140625, 16.673831939697266, 1.0514163970947266, 38.3201904296875, 14.507881164550781, -8.2742919921875, 6.751922607421875, 30.14788055419922, 2.9077720642089844, -3.370738983154297, 41.64421081542969, 9.036251068115234, 2.005084991455078], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000268.npy"}
{"epoch": 0.4051398337112623, "step": 269, "batch_size": 64, "mean": 11.029000282287598, "std": 18.439836502075195, "min": -34.25373840332031, "p10": -7.370617294311523, "median": 9.399742126464844, "p90": 36.90402183532715, "max": 53.62168884277344, "pos_frac": 0.71875, "sample": [3.0415725708007812, -32.317962646484375, 12.761810302734375, 16.61929702758789, 5.488494873046875, -2.299652099609375, 13.460067749023438, 53.35420608520508, 24.368629455566406, -5.706695556640625, -3.4588241577148438, 9.734848022460938, 19.905746459960938, 3.6341171264648438, 19.99932098388672, 43.0430908203125, 0.34234619140625, 6.778472900390625, 12.030158996582031, -10.616851806640625, -4.751983642578125, -0.25347900390625, 14.45489501953125, 16.66314697265625, -33.435577392578125, 36.71363830566406, 17.797271728515625, 53.62168884277344, 25.695510864257812, 34.50548553466797, 8.638483047485352, 19.08523178100586, -0.2600898742675781, 32.061607360839844, 15.741897583007812, 3.159799575805664, -0.6429290771484375, 4.521759033203125, -2.6600799560546875, 40.57109832763672, 40.008060455322266, 11.381645202636719, -3.4487991333007812, 10.669528007507324, 36.98561477661133, -10.685615539550781, 25.301746368408203, -7.589328765869141, -11.567169189453125, 8.373275756835938, 9.06463623046875, 2.3434696197509766, 28.49029541015625, 8.842826843261719, -1.4175796508789062, -34.25373840332031, -6.86029052734375, 28.5595703125, 1.688018798828125, 20.925003051757812, 37.60329818725586, 20.49901580810547, 18.547100067138672, 1.0058937072753906], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000269.npy"}
{"epoch": 0.40665154950869237, "step": 270, "batch_size": 64, "mean": 9.902159690856934, "std": 17.924068450927734, "min": -42.423500061035156, "p10": -7.847921752929684, "median": 8.926350593566895, "p90": 35.869943237304696, "max": 56.135169982910156, "pos_frac": 0.765625, "sample": [30.181617736816406, 33.258262634277344, -2.9689693450927734, 50.61973571777344, 9.090642929077148, 37.601898193359375, 22.644073486328125, 19.248363494873047, 22.041305541992188, 2.464385986328125, 56.135169982910156, 32.55303955078125, -0.3733062744140625, 23.454116821289062, 2.9917259216308594, -4.425018310546875, 7.5036468505859375, 7.311038970947266, 1.0494003295898438, 37.16328430175781, 10.84716796875, 3.9016265869140625, -15.231056213378906, 6.519386291503906, 24.24231719970703, 12.719833374023438, -3.742918014526367, -24.140335083007812, 37.96204376220703, 4.924434661865234, -0.20355224609375, 5.68084716796875, 3.2687454223632812, 14.721450805664062, -0.6797714233398438, 10.819271087646484, 13.402027130126953, -13.955169677734375, 21.70758819580078, -42.423500061035156, 10.300262451171875, 36.989234924316406, 25.899795532226562, -3.0951061248779297, 13.73640251159668, 9.944259643554688, -9.31488037109375, -1.6675949096679688, 12.888442993164062, 4.400730133056641, 19.120513916015625, 0.7907371520996094, -18.669586181640625, 15.232620239257812, 3.726360321044922, 2.0005645751953125, 5.802764892578125, -34.81254577636719, 37.361907958984375, 8.76205825805664, 10.599716186523438, 0.595794677734375, 15.92715072631836, 9.333755493164062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000270.npy"}
{"epoch": 0.40816326530612246, "step": 271, "batch_size": 64, "mean": 7.860054016113281, "std": 21.25482177734375, "min": -26.556217193603516, "p10": -14.684622955322265, "median": 3.045191764831543, "p90": 40.27345275878907, "max": 66.08013916015625, "pos_frac": 0.578125, "sample": [9.761253356933594, 18.043128967285156, -1.853790283203125, -15.07110595703125, 40.80107116699219, 44.52710723876953, 11.479269027709961, 5.1162872314453125, 17.086288452148438, -10.947872161865234, 29.360427856445312, -9.130117416381836, 39.04234313964844, -6.301979064941406, 16.117385864257812, 28.845703125, 4.438411712646484, -1.372406005859375, 1.6519718170166016, 57.16111755371094, -8.088249206542969, -5.037029266357422, 28.010643005371094, -16.313907623291016, -7.1058349609375, -2.668201446533203, -13.782829284667969, 5.9161224365234375, -0.9386138916015625, -5.294727325439453, 12.634490966796875, 15.116378784179688, -21.60089111328125, -24.837665557861328, 8.001800537109375, 42.325828552246094, 10.011749267578125, -0.8850326538085938, 31.02593231201172, 49.79620361328125, -2.2800827026367188, -12.670974731445312, 66.08013916015625, -6.90252685546875, 0.08377838134765625, 8.608734130859375, -2.237457275390625, -21.954980850219727, -26.556217193603516, -20.965377807617188, 7.968841552734375, 1.3212890625, 38.66419982910156, 31.929676055908203, 4.481971740722656, 0.982421875, 0.1738433837890625, 14.001983642578125, 7.021659851074219, -11.583038330078125, -6.245012283325195, -13.396865844726562, 55.01223373413086, 16.464553833007812], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000271.npy"}
{"epoch": 0.40967498110355255, "step": 272, "batch_size": 64, "mean": 16.469074249267578, "std": 18.66221046447754, "min": -27.703414916992188, "p10": -2.4536369323730467, "median": 9.94532585144043, "p90": 44.201819610595706, "max": 54.90522384643555, "pos_frac": 0.828125, "sample": [9.769035339355469, -0.9021759033203125, -3.521148681640625, 5.614265441894531, 18.457870483398438, 49.31861114501953, 8.313095092773438, 50.818695068359375, 54.14472961425781, 10.12161636352539, 17.178131103515625, 18.093936920166016, 1.4696846008300781, 1.5795059204101562, 32.57069396972656, 7.090763092041016, 7.3539276123046875, -5.005073547363281, 44.43595886230469, 48.24091720581055, 31.166038513183594, 8.90445327758789, 37.431705474853516, 21.592483520507812, 37.749351501464844, 17.390762329101562, 33.120704650878906, -18.124847412109375, 6.92291259765625, 38.783203125, 13.606613159179688, 54.90522384643555, 3.9775314331054688, 0.06013298034667969, 39.73918151855469, 8.08327865600586, 0.5469436645507812, -2.101837158203125, 31.266616821289062, -27.703414916992188, 23.252731323242188, 43.655494689941406, 32.17035675048828, 29.864883422851562, 6.492668151855469, 3.749431610107422, 5.453094482421875, -2.0075225830078125, 27.049163818359375, 33.464088439941406, 12.939861297607422, 20.554668426513672, 1.132598876953125, -6.205108642578125, 19.963462829589844, 25.035400390625, 3.8412704467773438, 9.73602294921875, -2.6044082641601562, 0.8620471954345703, -6.204826354980469, -0.005245208740234375, 49.806846618652344, 9.563636779785156], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000272.npy"}
{"epoch": 0.41118669690098264, "step": 273, "batch_size": 64, "mean": 11.581703186035156, "std": 23.39433479309082, "min": -45.920921325683594, "p10": -22.602706146240227, "median": 13.889758110046387, "p90": 41.374541473388675, "max": 57.605682373046875, "pos_frac": 0.765625, "sample": [10.983940124511719, 20.790298461914062, 39.5177001953125, -28.091346740722656, 20.7796630859375, 7.41815185546875, -8.725311279296875, 23.077316284179688, -37.66510009765625, 5.950614929199219, -0.503021240234375, -25.594253540039062, 20.594482421875, -8.605459213256836, 14.137149810791016, 32.07793426513672, 39.52789306640625, 10.126953125, 48.98475646972656, 6.055717468261719, 16.56798553466797, 14.967105865478516, -15.622428894042969, -11.555747985839844, -45.920921325683594, 16.648208618164062, 27.204727172851562, 3.3050613403320312, -10.391393661499023, 1.8260345458984375, 2.6440277099609375, 45.969451904296875, 19.67975616455078, -38.03539276123047, 24.182510375976562, 17.019027709960938, 31.556724548339844, 24.26282501220703, 0.9295272827148438, 16.568519592285156, -29.733482360839844, 15.321182250976562, 31.656227111816406, -13.122966766357422, 2.3083114624023438, 57.605682373046875, 53.101139068603516, 40.71638488769531, 2.8905792236328125, 41.65660858154297, 12.350435256958008, 4.36212158203125, 13.642366409301758, -8.049850463867188, 9.087181091308594, 27.94487762451172, 5.246189117431641, 28.238800048828125, 1.5625801086425781, 15.512367248535156, 47.44294738769531, 47.84950256347656, -32.12834930419922, 33.12449645996094], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000273.npy"}
{"epoch": 0.4126984126984127, "step": 274, "batch_size": 64, "mean": 13.344884872436523, "std": 19.98091697692871, "min": -24.874343872070312, "p10": -10.224048233032226, "median": 9.46828842163086, "p90": 42.17303619384766, "max": 50.9539794921875, "pos_frac": 0.71875, "sample": [5.374626159667969, -5.97772216796875, -8.657526016235352, 1.9764404296875, -1.2258758544921875, -19.573837280273438, 41.659881591796875, 26.1854248046875, 17.728641510009766, -8.707378387451172, 34.32831573486328, 42.39295959472656, -10.590599060058594, -16.64007568359375, 12.447769165039062, 50.77693176269531, 37.585567474365234, -15.82276725769043, 18.400592803955078, -5.433067321777344, 21.524581909179688, 39.641029357910156, 21.94512176513672, 45.35371017456055, 7.5619659423828125, 0.6990623474121094, 41.40235900878906, 41.1806640625, 3.082134246826172, 6.841217041015625, -12.103931427001953, 0.690704345703125, 32.565696716308594, 0.6026840209960938, -8.942707061767578, 12.2451171875, -1.3115119934082031, 0.5010147094726562, 18.501827239990234, 8.345535278320312, -9.368762969970703, 31.583175659179688, 2.4041595458984375, 6.458251953125, -0.5528507232666016, 30.382781982421875, 10.591041564941406, -24.874343872070312, 12.249908447265625, -7.2140655517578125, 45.293540954589844, 50.0172119140625, 20.17695426940918, 29.051681518554688, 4.94781494140625, -1.2310047149658203, -12.701065063476562, 22.455833435058594, 50.9539794921875, 22.560588836669922, 7.554256439208984, 17.804603576660156, 44.1832275390625, 24.791168212890625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000274.npy"}
{"epoch": 0.41421012849584277, "step": 275, "batch_size": 64, "mean": 17.002704620361328, "std": 20.68329620361328, "min": -23.57080078125, "p10": -5.575213813781738, "median": 12.638839721679688, "p90": 48.83003158569336, "max": 61.0315055847168, "pos_frac": 0.796875, "sample": [1.866729736328125, 4.9893341064453125, 9.273368835449219, 25.077136993408203, 44.84527587890625, 0.10543441772460938, -2.1113662719726562, -16.610034942626953, 51.572265625, 42.29148864746094, 5.8807373046875, 20.682968139648438, 6.530029296875, 52.91297912597656, -2.574798583984375, 0.47022247314453125, 16.925704956054688, 6.098470687866211, 43.397239685058594, -5.903785705566406, -23.57080078125, 61.0315055847168, 0.9816131591796875, 8.244735717773438, 39.62309265136719, 2.5054855346679688, 1.2237777709960938, -11.075691223144531, 22.01233673095703, 13.428314208984375, 19.493301391601562, 51.237518310546875, -0.3071746826171875, 22.692955017089844, 52.61016082763672, 49.13951873779297, 6.121103286743164, 15.343502044677734, -12.905715942382812, -3.1488189697265625, 35.41743469238281, 15.14181900024414, 6.043846130371094, 1.3649835586547852, 7.736087799072266, 38.714256286621094, 43.70164489746094, 26.7850341796875, -4.753910064697266, 28.988662719726562, 2.85711669921875, -10.968452453613281, -6.905231475830078, 23.13127899169922, 49.5059928894043, 32.56512451171875, 11.849365234375, 18.645126342773438, 29.18268394470215, 31.78765869140625, 48.10789489746094, 5.372161865234375, -4.80854606628418, 38.31089782714844], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000275.npy"}
{"epoch": 0.41572184429327286, "step": 276, "batch_size": 64, "mean": 13.906017303466797, "std": 21.865917205810547, "min": -37.81895446777344, "p10": -13.40870895385742, "median": 10.291315078735352, "p90": 47.50452423095704, "max": 56.70277404785156, "pos_frac": 0.796875, "sample": [19.221546173095703, 2.4098434448242188, -17.505382537841797, 4.593940734863281, 2.7749462127685547, 11.113250732421875, 54.49928283691406, -37.704010009765625, 22.263580322265625, 13.435047149658203, 23.471885681152344, 10.052616119384766, 8.01025390625, 29.272262573242188, 22.83538818359375, 36.085269927978516, 15.716480255126953, 2.0237693786621094, 22.899810791015625, 2.9820327758789062, 5.7740631103515625, 50.58326721191406, 16.031030654907227, 48.05461883544922, -3.0746841430664062, -37.81895446777344, 39.59515380859375, 9.652145385742188, 20.02145767211914, 48.3349494934082, 1.4022979736328125, 0.073394775390625, -19.322418212890625, -15.316459655761719, 36.32912063598633, -12.468612670898438, 54.85166931152344, -7.860317230224609, 23.285232543945312, 33.76909637451172, -14.412887573242188, 2.0429840087890625, 54.22229766845703, 0.8186454772949219, 2.9873580932617188, -13.811607360839844, 33.16429901123047, 38.35737609863281, 56.70277404785156, 26.341651916503906, -1.6902847290039062, 4.2950439453125, 7.348930358886719, 10.530014038085938, 1.8460235595703125, 46.220970153808594, 37.8906135559082, 6.991424560546875, -10.509559631347656, 8.6668701171875, 12.23904800415039, 18.5867919921875, 26.49819564819336, -5.689689636230469], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000276.npy"}
{"epoch": 0.41723356009070295, "step": 277, "batch_size": 64, "mean": 7.397580146789551, "std": 22.040964126586914, "min": -41.12565612792969, "p10": -19.006427764892578, "median": 5.195046424865723, "p90": 35.69664268493653, "max": 64.32305908203125, "pos_frac": 0.625, "sample": [42.050472259521484, -9.390472412109375, 36.06466293334961, 2.8235092163085938, 13.810993194580078, -9.009185791015625, 15.835784912109375, -1.1212997436523438, 40.149688720703125, 21.605667114257812, -19.45891571044922, -2.8044281005859375, 25.308074951171875, 1.1944808959960938, -17.839130401611328, 10.721733093261719, 0.5624618530273438, 7.0526123046875, -1.1415824890136719, -2.337371826171875, 11.565305709838867, 22.47339630126953, 5.605892181396484, 15.112617492675781, -17.490882873535156, 19.5570068359375, 7.841575622558594, -17.067962646484375, -17.3497314453125, 24.406890869140625, 0.6598377227783203, 34.73530197143555, -5.938671112060547, 28.09774398803711, 4.784200668334961, 25.41101837158203, -21.051071166992188, -38.699684143066406, -9.936294555664062, -17.95062255859375, -1.0266876220703125, -10.100852966308594, 29.90777587890625, 34.837928771972656, 4.2411346435546875, 7.571527481079102, 7.43206787109375, 16.826766967773438, -41.12565612792969, 64.32305908203125, 34.249305725097656, -0.5596027374267578, 34.811920166015625, 37.894920349121094, -5.8971099853515625, 2.0417613983154297, 46.663848876953125, -24.974365234375, 9.787445068359375, 52.52647018432617, -25.550304412841797, 1.5072517395019531, -24.442272186279297, 13.655136108398438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000277.npy"}
{"epoch": 0.41874527588813304, "step": 278, "batch_size": 64, "mean": 10.672148704528809, "std": 23.223318099975586, "min": -45.268165588378906, "p10": -18.22612152099609, "median": 9.498937606811523, "p90": 45.501056289672874, "max": 60.620574951171875, "pos_frac": 0.671875, "sample": [-7.538719177246094, -18.556961059570312, 27.641197204589844, 1.791259765625, 8.92050552368164, -5.3085432052612305, -33.66288757324219, 38.999961853027344, 47.64668273925781, 54.909027099609375, 18.144309997558594, 1.5097312927246094, 40.49459457397461, 51.72799301147461, 51.82582092285156, -6.043910980224609, 8.332174301147461, 22.038375854492188, 14.452625274658203, 16.784069061279297, 15.836343765258789, 11.996253967285156, 2.145343780517578, -28.890792846679688, -45.268165588378906, 37.343353271484375, -17.45416259765625, 10.372211456298828, 19.791534423828125, 8.210159301757812, 10.077369689941406, 10.290153503417969, -11.271270751953125, -24.249717712402344, 37.18711853027344, 6.585662841796875, 2.958271026611328, -0.1905670166015625, 28.787647247314453, -4.355499267578125, 51.45515441894531, 60.620574951171875, -0.05448150634765625, 16.363433837890625, 38.635921478271484, 21.797779083251953, 8.730144500732422, -2.2519283294677734, -16.511688232421875, 6.090118408203125, 11.683853149414062, 28.873779296875, -25.727951049804688, -1.7528648376464844, -3.5654449462890625, 54.454071044921875, -27.42376708984375, 13.527313232421875, 15.414813995361328, -5.336235046386719, -1.2627182006835938, 0.5179939270019531, 22.480228424072266, 12.250839233398438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000278.npy"}
{"epoch": 0.42025699168556313, "step": 279, "batch_size": 64, "mean": 19.376266479492188, "std": 21.287954330444336, "min": -35.56385040283203, "p10": -2.9583374023437488, "median": 15.290428161621094, "p90": 51.455464172363286, "max": 58.66388702392578, "pos_frac": 0.84375, "sample": [-3.888315200805664, 9.238399505615234, 54.60256576538086, 9.0257568359375, -35.56385040283203, 51.65531921386719, 14.380453109741211, -7.264396667480469, -12.124343872070312, 31.219226837158203, 13.795074462890625, 54.39600372314453, 11.286514282226562, 54.474300384521484, 13.611618041992188, 50.9891357421875, 34.74320983886719, 49.38447570800781, -17.266162872314453, -0.9023666381835938, -3.437469482421875, 48.167930603027344, 8.537059783935547, 27.925182342529297, 17.579925537109375, 30.917953491210938, 27.15561294555664, 1.8434791564941406, -0.569244384765625, 33.270591735839844, 30.39319610595703, 43.473697662353516, 54.85911560058594, -1.840362548828125, 0.37732696533203125, 25.053760528564453, 21.15886688232422, 14.176620483398438, 17.52973175048828, 50.50030517578125, 39.466758728027344, 8.12115478515625, 32.61357116699219, 10.088661193847656, -14.702987670898438, 23.005897521972656, 1.0832958221435547, 4.536540985107422, 4.834503173828125, 36.83427429199219, 32.00880432128906, 0.04925537109375, 2.3310890197753906, 53.04414749145508, 1.1083850860595703, 2.0009841918945312, 15.809463500976562, 14.771392822265625, 34.3401985168457, 15.980796813964844, 34.146209716796875, 2.8937301635742188, 58.66388702392578, 4.1851348876953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000279.npy"}
{"epoch": 0.4217687074829932, "step": 280, "batch_size": 64, "mean": 18.94228744506836, "std": 21.46267318725586, "min": -44.06956481933594, "p10": -0.568588638305664, "median": 17.158126831054688, "p90": 51.38313293457032, "max": 66.67772674560547, "pos_frac": 0.875, "sample": [2.9958114624023438, 66.67772674560547, 8.456459045410156, 32.04717254638672, -1.9967575073242188, 53.30266189575195, 1.2848529815673828, 0.2995758056640625, 10.542137145996094, 27.358848571777344, 18.73963165283203, 11.98428726196289, 26.803741455078125, 23.239349365234375, 42.526947021484375, 45.39665222167969, 35.35377502441406, 27.08736801147461, 46.847076416015625, -13.92498779296875, 1.23492431640625, 2.316967010498047, 2.25726318359375, -0.4498863220214844, 58.94005584716797, 33.447265625, 21.497079849243164, 34.92231750488281, 51.988616943359375, 19.000213623046875, 6.216547012329102, 3.309051513671875, 1.567779541015625, 49.9703369140625, 2.2886734008789062, 4.1612548828125, 63.71263122558594, -0.6194610595703125, 56.145782470703125, 29.73688507080078, -17.849876403808594, -44.06956481933594, 16.535903930664062, 0.1529693603515625, 4.773796081542969, 25.429271697998047, 5.646240234375, 0.30267906188964844, 6.766838073730469, 22.465572357177734, 47.19636535644531, 26.701435089111328, 10.638015747070312, 26.572261810302734, -0.8044586181640625, 14.632644653320312, 19.76171875, -1.4898223876953125, 33.57514190673828, 2.02978515625, 11.710716247558594, 53.68279266357422, 23.49895477294922, 17.780349731445312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000280.npy"}
{"epoch": 0.42328042328042326, "step": 281, "batch_size": 64, "mean": 14.368467330932617, "std": 22.588090896606445, "min": -35.77863311767578, "p10": -12.949679183959956, "median": 11.181511878967285, "p90": 44.513265228271486, "max": 65.80857849121094, "pos_frac": 0.765625, "sample": [-18.957759857177734, 46.00113296508789, 8.501285552978516, 14.98297119140625, 34.4085693359375, 23.137359619140625, 20.49771499633789, 10.575401306152344, 4.297393798828125, 38.756622314453125, 19.821685791015625, 35.95985794067383, 25.991130828857422, 28.55707550048828, -7.0563812255859375, 8.482841491699219, 11.787622451782227, 43.328025817871094, -6.447269439697266, 12.798210144042969, 36.3253059387207, 6.3334503173828125, 0.69000244140625, 40.81459045410156, 19.356597900390625, 2.4448623657226562, 12.728137969970703, -35.77863311767578, 9.830062866210938, -14.801414489746094, 1.0998916625976562, -2.2333850860595703, -6.132129669189453, 2.6284093856811523, 56.6392822265625, -34.32536315917969, 12.240921020507812, 65.80857849121094, 1.2834625244140625, -16.254539489746094, 25.148479461669922, 5.09478759765625, 43.1968879699707, 33.558197021484375, 56.12131881713867, -0.6886062622070312, 27.543167114257812, 17.12099838256836, 49.304378509521484, 9.61184310913086, -8.520217895507812, 1.8192901611328125, 51.49831008911133, -28.705997467041016, 0.1816539764404297, 26.027210235595703, 45.02122497558594, -3.0334625244140625, 38.92543029785156, 30.62905502319336, 1.6665191650390625, -8.628963470458984, 7.584697723388672, -15.015861511230469], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000281.npy"}
{"epoch": 0.42479213907785335, "step": 282, "batch_size": 64, "mean": 16.517818450927734, "std": 24.738178253173828, "min": -23.21142578125, "p10": -15.498136520385742, "median": 9.605125427246094, "p90": 55.334113693237306, "max": 66.29889678955078, "pos_frac": 0.75, "sample": [4.6353912353515625, 32.451011657714844, -2.6695404052734375, 4.3913421630859375, 7.323707580566406, 9.826026916503906, 29.434188842773438, 8.205184936523438, -11.720069885253906, 56.277069091796875, 38.38496017456055, 34.72119903564453, 4.9403791427612305, 9.384223937988281, 55.11761474609375, -0.63775634765625, 61.60581970214844, 10.29404067993164, -1.4212646484375, 22.561737060546875, 8.351055145263672, 17.3995361328125, 1.5527191162109375, 63.60884094238281, 1.3693084716796875, -15.709674835205078, 0.661041259765625, 10.395126342773438, 55.42689895629883, 33.283447265625, -21.80205535888672, -13.48866081237793, -18.136138916015625, 50.35176086425781, -8.681079864501953, -23.21142578125, 27.542221069335938, 3.948331832885742, 5.75860595703125, 6.3856201171875, 5.808757781982422, 36.089271545410156, -5.5884246826171875, 33.467369079589844, 45.77061462402344, 62.96321105957031, 23.085540771484375, 36.35430145263672, -2.0318527221679688, 54.30097961425781, -16.481063842773438, 4.428321838378906, -21.31182861328125, 22.508651733398438, -16.85650634765625, 66.29889678955078, 38.75873565673828, 1.2137718200683594, 20.259750366210938, -15.004547119140625, 11.656234741210938, 38.6624755859375, 60.90617370605469, 13.77090072631836], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000282.npy"}
{"epoch": 0.42630385487528344, "step": 283, "batch_size": 64, "mean": 13.276298522949219, "std": 24.394081115722656, "min": -54.36906433105469, "p10": -19.312433624267577, "median": 13.902684211730957, "p90": 46.106577682495114, "max": 63.211700439453125, "pos_frac": 0.75, "sample": [63.211700439453125, -20.126060485839844, -30.872276306152344, -9.635383605957031, 35.97715759277344, 18.49826431274414, 14.964332580566406, -4.337760925292969, 46.00804138183594, -8.813323974609375, 0.1485748291015625, -10.946281433105469, 20.014202117919922, 10.76763916015625, 12.971118927001953, 10.520614624023438, 15.271240234375, 37.38616943359375, -34.09117126464844, 8.34283447265625, 25.619606018066406, 7.5146026611328125, 27.14773178100586, -4.80937385559082, 15.237823486328125, 53.321739196777344, 49.840789794921875, 15.537849426269531, 55.887786865234375, 6.9737548828125, 30.217758178710938, 43.36931228637695, -4.543464660644531, 13.70284652709961, -8.601364135742188, 21.276016235351562, 1.7725601196289062, -17.413970947265625, 26.401107788085938, 30.878097534179688, 53.44610595703125, -2.5215072631835938, 46.634613037109375, 46.148807525634766, 4.67266845703125, 4.817558288574219, 26.99810028076172, -54.36906433105469, 1.2851533889770508, -21.0137939453125, 32.9693489074707, 8.749958038330078, 35.81304931640625, 14.102521896362305, 22.311920166015625, 39.98332214355469, -31.416648864746094, 22.623062133789062, 1.9376296997070312, 6.068145751953125, -30.815841674804688, 35.5324821472168, 4.32421875, 16.810440063476562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000283.npy"}
{"epoch": 0.42781557067271353, "step": 284, "batch_size": 64, "mean": 18.640045166015625, "std": 23.86236572265625, "min": -38.75469207763672, "p10": -2.4968223571777344, "median": 14.652908325195312, "p90": 52.13600921630861, "max": 69.63805389404297, "pos_frac": 0.796875, "sample": [-1.09674072265625, 29.622085571289062, 17.80644989013672, 8.904937744140625, 68.01763916015625, 14.53143310546875, -1.5573501586914062, -31.429779052734375, 66.02507019042969, 3.4447860717773438, -0.4543914794921875, 37.04777908325195, 3.2003021240234375, 7.2010040283203125, 48.11204528808594, -1.0719375610351562, 34.66926574707031, 28.954675674438477, -2.4234695434570312, 26.99609375, 17.702789306640625, -4.052886962890625, 5.0242919921875, -23.725120544433594, 0.0167236328125, 21.077484130859375, 7.237091064453125, -2.67095947265625, 24.54814910888672, 40.583805084228516, 22.363685607910156, 44.361671447753906, 53.860565185546875, 28.13982391357422, 12.923906326293945, 2.5232715606689453, 14.774383544921875, 63.678260803222656, 69.178466796875, 4.004547119140625, 18.395462036132812, 64.66425323486328, 24.154739379882812, 24.327796936035156, -0.18666839599609375, 6.143739700317383, 29.751983642578125, 11.195014953613281, 28.989540100097656, 44.12318420410156, 2.551969528198242, 42.36573791503906, 37.20170593261719, 22.618499755859375, 4.395660400390625, 9.686538696289062, 69.63805389404297, -2.52825927734375, 37.34821701049805, -18.40186309814453, 5.2529754638671875, -38.75469207763672, 11.7999267578125, 0.17961692810058594], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000284.npy"}
{"epoch": 0.4293272864701436, "step": 285, "batch_size": 64, "mean": 11.305547714233398, "std": 25.35523223876953, "min": -63.80281066894531, "p10": -15.806154251098631, "median": 7.390510559082031, "p90": 48.97710533142091, "max": 65.31028747558594, "pos_frac": 0.625, "sample": [-5.001091003417969, 2.263721466064453, -12.303192138671875, 50.10863494873047, -2.8249969482421875, -16.99004364013672, -39.94800567626953, 15.438396453857422, -9.432209014892578, -63.80281066894531, 42.417396545410156, 29.916088104248047, 46.83879470825195, 17.09271240234375, 28.8218994140625, -16.691417694091797, 26.025840759277344, 56.916664123535156, -0.7848052978515625, -0.5004730224609375, 1.9625625610351562, 2.913999557495117, -18.08332061767578, 1.3354568481445312, -3.6768722534179688, -26.879913330078125, -8.371749877929688, 3.1547698974609375, 8.06304931640625, -6.231658935546875, -9.679550170898438, 31.595611572265625, -6.174736022949219, 49.893524169921875, 36.62931823730469, 8.535186767578125, 43.102386474609375, 6.7179718017578125, 15.195812225341797, 16.87143898010254, 35.78936004638672, 14.280082702636719, -2.987518310546875, 57.57232666015625, 56.279083251953125, 17.005508422851562, 30.011924743652344, 65.31028747558594, -12.150543212890625, -22.672042846679688, 8.503824234008789, 23.034080505371094, 21.441543579101562, 33.86940002441406, 5.83758544921875, 9.196144104003906, 55.516822814941406, 36.87699890136719, -2.5938072204589844, -0.50714111328125, -8.583572387695312, -13.74053955078125, 5.8978424072265625, 15.933059692382812], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000285.npy"}
{"epoch": 0.4308390022675737, "step": 286, "batch_size": 64, "mean": 19.008071899414062, "std": 26.276195526123047, "min": -42.37608337402344, "p10": -6.449374389648437, "median": 9.78506088256836, "p90": 57.339228820800784, "max": 76.31436157226562, "pos_frac": 0.78125, "sample": [39.82122802734375, 7.414642333984375, 1.5765495300292969, 7.067626953125, 8.499069213867188, 46.40751647949219, 5.433906555175781, -3.6165828704833984, 8.140804290771484, -9.909950256347656, 42.907745361328125, 8.82513427734375, 9.452430725097656, 15.532909393310547, 27.087238311767578, 55.80401611328125, 28.737689971923828, 27.736482620239258, -39.15998458862305, 26.861061096191406, 5.1426849365234375, -42.37608337402344, -1.2788619995117188, 60.035037994384766, -6.1685028076171875, -7.769111633300781, 40.84416198730469, 26.119354248046875, 11.232421875, 10.117691040039062, 46.97595977783203, 35.344032287597656, 58.228416442871094, 57.99717712402344, 28.717601776123047, 76.31436157226562, -2.8653717041015625, 50.830352783203125, -0.168182373046875, 16.077072143554688, 67.05802154541016, 70.58514404296875, 2.9293441772460938, -8.965789794921875, 26.350139617919922, 4.86383056640625, 8.909856796264648, 5.22607421875, 60.39988708496094, 48.18896484375, 49.53515625, -3.9121017456054688, 4.497222900390625, 5.703891754150391, 0.83642578125, 9.043441772460938, 39.35675811767578, 12.184185028076172, -40.78900146484375, 46.333984375, 9.337800979614258, -6.5697479248046875, 28.573959350585938, -1.1306533813476562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000286.npy"}
{"epoch": 0.4323507180650038, "step": 287, "batch_size": 64, "mean": 24.37060546875, "std": 27.119731903076172, "min": -45.41996765136719, "p10": -6.694028472900389, "median": 19.536219596862793, "p90": 64.76662902832031, "max": 74.31672668457031, "pos_frac": 0.8125, "sample": [18.44495391845703, 1.69482421875, 6.531524658203125, 64.88914489746094, 69.41763305664062, 57.53137969970703, 9.009368896484375, 24.179161071777344, 56.136138916015625, -9.741775512695312, 17.105987548828125, 33.181339263916016, 43.64195251464844, 32.24310302734375, 65.0416259765625, 69.42532348632812, 26.725746154785156, 4.571941375732422, 20.25600242614746, 4.3152313232421875, 25.159805297851562, 0.281768798828125, 41.240081787109375, 13.130611419677734, 15.29690170288086, 55.59968185424805, -16.351890563964844, -5.101676940917969, 59.44964599609375, 1.9475212097167969, 45.84230041503906, 2.7137527465820312, -45.41996765136719, -2.5005264282226562, -0.4121408462524414, 25.929580688476562, 36.07033920288086, -7.37646484375, 20.63196563720703, 46.430660247802734, 40.948570251464844, -1.387725830078125, 49.33066177368164, 69.99305725097656, 18.816436767578125, 20.778671264648438, 74.31672668457031, -8.539346694946289, 53.00386047363281, 54.60155487060547, -0.24424266815185547, 5.4438629150390625, 3.6655120849609375, 10.46160888671875, -10.994258880615234, 15.925430297851562, 52.42569351196289, 16.359962463378906, 24.98595428466797, 10.724529266357422, -20.23822021484375, 71.80586242675781, 64.48075866699219, 15.891250610351562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000287.npy"}
{"epoch": 0.43386243386243384, "step": 288, "batch_size": 64, "mean": 13.595813751220703, "std": 25.153051376342773, "min": -29.244264602661133, "p10": -20.600857925415035, "median": 11.572198867797852, "p90": 46.1343002319336, "max": 73.58329772949219, "pos_frac": 0.703125, "sample": [22.240798950195312, 5.653541564941406, -29.0614013671875, -23.07635498046875, 14.605751037597656, 25.63799285888672, 25.828929901123047, -12.404006958007812, 29.584556579589844, 4.590824127197266, 16.437301635742188, 4.20538330078125, 22.171188354492188, 46.63909912109375, 1.0286865234375, 21.727920532226562, -22.85150146484375, 12.428108215332031, -5.1043243408203125, 35.563419342041016, -10.97186279296875, 31.994224548339844, 53.49177551269531, 67.93785858154297, 6.659477233886719, 44.69575500488281, 18.496368408203125, 54.830196380615234, -4.1650848388671875, 11.104972839355469, 73.58329772949219, -12.912551879882812, 8.357757568359375, 44.95643615722656, 60.206573486328125, 20.28973388671875, 15.280715942382812, -12.909713745117188, 25.437000274658203, -3.6878585815429688, -29.244264602661133, -25.432235717773438, 70.55425262451172, 5.3072357177734375, 9.429454803466797, -22.55649185180664, 8.188026428222656, -3.51031494140625, 12.039424896240234, 44.755523681640625, 35.36748504638672, -1.186431884765625, 40.026763916015625, 3.3347015380859375, -23.047027587890625, -12.52059555053711, 12.533554077148438, 3.6464309692382812, 14.014057159423828, 42.288848876953125, 0.7324676513671875, -1.5930862426757812, 14.52105712890625, -16.03771209716797], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000288.npy"}
{"epoch": 0.43537414965986393, "step": 289, "batch_size": 64, "mean": 17.713703155517578, "std": 26.52002716064453, "min": -50.53925323486328, "p10": -8.527880096435547, "median": 10.945306777954102, "p90": 55.72762184143067, "max": 82.7879638671875, "pos_frac": 0.671875, "sample": [-16.05987548828125, -5.769683837890625, 31.942825317382812, -4.8579559326171875, -1.0296173095703125, 8.959304809570312, -0.8519935607910156, 16.755027770996094, 5.509368896484375, 3.2838897705078125, -10.884696960449219, 35.18701171875, 14.704261779785156, 25.704307556152344, 36.38831329345703, -1.0040111541748047, 46.886505126953125, -14.606277465820312, 10.41146469116211, 38.39030456542969, 3.2561187744140625, 65.91251373291016, 6.690753936767578, 6.2602386474609375, 31.924171447753906, 6.514339447021484, 42.643089294433594, 6.1673736572265625, 53.947044372558594, 68.74573516845703, -5.35894775390625, 26.95073890686035, -6.111278533935547, -26.590911865234375, 62.02747344970703, -4.847902297973633, -8.672645568847656, 82.7879638671875, 31.037765502929688, 26.303985595703125, 9.534645080566406, 15.784500122070312, 21.57427215576172, -15.734977722167969, 34.80320358276367, -50.53925323486328, 13.108556747436523, -2.902578353881836, -0.4483184814453125, -0.4612236022949219, 63.132545471191406, 21.6143798828125, -0.08025360107421875, 6.024658203125, 29.946372985839844, 63.4488525390625, 53.510414123535156, 56.142356872558594, 54.25470733642578, 11.479148864746094, 16.745040893554688, -8.190093994140625, 54.75990676879883, -2.4758872985839844], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000289.npy"}
{"epoch": 0.436885865457294, "step": 290, "batch_size": 64, "mean": 14.920310974121094, "std": 26.356142044067383, "min": -51.5684814453125, "p10": -17.456614303588864, "median": 14.51231575012207, "p90": 51.39491195678711, "max": 71.7770767211914, "pos_frac": 0.765625, "sample": [53.75018310546875, 12.816959381103516, 25.752199172973633, 61.661651611328125, 26.334217071533203, 2.384510040283203, 2.755584716796875, -2.3963890075683594, 6.0298614501953125, 32.58930969238281, 21.272682189941406, 19.226856231689453, -27.79242706298828, -9.141777038574219, 32.53031921386719, 0.1340484619140625, 18.07537841796875, -23.875831604003906, -11.095390319824219, 19.53448486328125, 2.2056922912597656, 35.42634582519531, 53.114479064941406, 45.504638671875, 20.363975524902344, -18.54000473022461, 15.746997833251953, -0.3704833984375, 16.93109893798828, -32.85923767089844, 51.56517028808594, 13.277633666992188, 36.682289123535156, 33.263816833496094, -14.928703308105469, 67.64938354492188, 50.997642517089844, 19.914932250976562, 36.432472229003906, -51.5684814453125, 20.31329345703125, 26.364437103271484, 0.77130126953125, -40.688690185546875, 8.677616119384766, 1.6862411499023438, -1.244729995727539, 71.7770767211914, 64.25855255126953, 37.54326629638672, 9.751235961914062, 25.9468994140625, -1.2795944213867188, 1.9613456726074219, -39.7941780090332, 20.659698486328125, 4.9118194580078125, 8.902629852294922, 6.4340057373046875, 10.233009338378906, 36.55195617675781, 41.30274963378906, 3.1882858276367188, -4.684425354003906], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000290.npy"}
{"epoch": 0.4383975812547241, "step": 291, "batch_size": 64, "mean": 17.954689025878906, "std": 31.742738723754883, "min": -66.09810638427734, "p10": -11.366353607177734, "median": 14.015933990478516, "p90": 62.972586441040036, "max": 85.7567138671875, "pos_frac": 0.703125, "sample": [51.328948974609375, 53.65071105957031, -1.0500564575195312, 8.119430541992188, 36.619415283203125, 35.15705108642578, -50.32489776611328, 10.171554565429688, 70.80931091308594, -11.662368774414062, 62.863582611083984, 34.04182815551758, -35.68939208984375, -7.390323638916016, 63.01930236816406, 66.03400421142578, 2.8906402587890625, -23.788986206054688, 41.452972412109375, 28.305339813232422, -0.36647796630859375, -1.7463455200195312, 43.25212860107422, 3.1175155639648438, 25.013011932373047, 32.57597351074219, 58.74317932128906, 1.2480792999267578, 77.78677368164062, 17.850147247314453, 31.996322631835938, 69.44312286376953, -6.62469482421875, -2.356487274169922, -0.5694961547851562, 15.723922729492188, 31.05005645751953, 0.31620025634765625, 12.307945251464844, 3.958110809326172, 18.19671630859375, 30.072669982910156, -66.09810638427734, 2.3252716064453125, 58.554779052734375, 19.740371704101562, 22.551353454589844, 29.088668823242188, 40.187496185302734, 85.7567138671875, -39.19317626953125, 2.0389862060546875, -34.82489013671875, 9.190147399902344, -9.517730712890625, 6.2933349609375, -0.9741058349609375, 68.54757690429688, 41.16307830810547, 36.590267181396484, 3.0139846801757812, -2.29913330078125, -10.675651550292969, -7.9056243896484375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000291.npy"}
{"epoch": 0.4399092970521542, "step": 292, "batch_size": 64, "mean": 16.948894500732422, "std": 28.798877716064453, "min": -65.67361450195312, "p10": -13.094911193847652, "median": 11.829614639282227, "p90": 54.52745361328125, "max": 71.730224609375, "pos_frac": 0.703125, "sample": [-16.236083984375, 5.458366394042969, 16.60858154296875, -1.7244434356689453, 35.92688751220703, 16.607099533081055, 14.454872131347656, 53.8218994140625, 41.624786376953125, -0.9250946044921875, -4.360466003417969, -3.538999557495117, 18.783607482910156, 3.494171142578125, -34.24750518798828, 9.204357147216797, 35.788848876953125, 50.899322509765625, 59.77631378173828, 7.8163299560546875, 2.0557098388671875, 42.193572998046875, 47.04779052734375, -6.586185455322266, -33.448394775390625, 0.879302978515625, -9.503128051757812, 34.17740249633789, -14.634246826171875, 3.1867332458496094, 5.212432861328125, 40.10717010498047, 16.56723403930664, 0.44771575927734375, 1.9619560241699219, -5.154388427734375, 53.761722564697266, -65.67361450195312, -9.199077606201172, -8.696243286132812, -6.9510040283203125, 29.173568725585938, 38.46236801147461, -2.7090835571289062, 71.16991424560547, 71.730224609375, 28.030502319335938, 35.42689895629883, 8.902778625488281, 15.137680053710938, 64.83941650390625, 54.829833984375, 0.8891258239746094, 63.63301086425781, 19.92281150817871, -2.135345458984375, 24.004371643066406, 5.4443817138671875, 47.10736083984375, 46.99565124511719, 65.15582275390625, -16.993377685546875, -29.549209594726562, 48.275184631347656], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000292.npy"}
{"epoch": 0.4414210128495843, "step": 293, "batch_size": 64, "mean": 21.632625579833984, "std": 36.265655517578125, "min": -76.96505737304688, "p10": -27.510531616210937, "median": 15.602603912353516, "p90": 67.89606857299805, "max": 89.66044616699219, "pos_frac": 0.765625, "sample": [5.225616455078125, 26.993572235107422, 18.188934326171875, 67.64311218261719, 29.155349731445312, 11.855064392089844, 2.5978736877441406, 10.272754669189453, 1.10369873046875, 4.1843719482421875, 49.309051513671875, -26.3250732421875, 74.77161407470703, 69.24942016601562, 53.04302215576172, 2.0816192626953125, 62.92993927001953, 38.844749450683594, 58.807861328125, 58.121498107910156, 36.84642791748047, -32.1097297668457, -0.6407089233398438, -28.018585205078125, 45.141876220703125, 6.3339080810546875, 16.5706844329834, -0.7758445739746094, 21.26592254638672, 89.66044616699219, 3.888216018676758, 80.10494995117188, 56.687286376953125, -42.99598693847656, 15.621139526367188, -3.908557891845703, 68.00447845458984, -76.96505737304688, 4.124053955078125, 15.584068298339844, -29.19605255126953, 85.15714263916016, -53.365447998046875, -25.387535095214844, -3.65704345703125, -0.923675537109375, 31.069091796875, 49.08530807495117, 73.00958251953125, 17.16967010498047, 60.251922607421875, -10.14642333984375, 38.8093147277832, 3.3860931396484375, 53.37785339355469, -31.727569580078125, 64.3797607421875, 8.186386108398438, 9.725547790527344, 61.22480010986328, 14.90814208984375, 3.194793701171875, 10.093521118164062, 63.38983154296875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000293.npy"}
{"epoch": 0.4429327286470144, "step": 294, "batch_size": 64, "mean": 18.27361488342285, "std": 28.048770904541016, "min": -55.444026947021484, "p10": -10.838282775878906, "median": 15.039247512817383, "p90": 51.3643295288086, "max": 88.32058715820312, "pos_frac": 0.6875, "sample": [73.50215148925781, 25.209251403808594, 19.82794952392578, 6.183929443359375, 52.01655578613281, 32.47075653076172, 36.12755584716797, 6.6365966796875, -55.444026947021484, 10.911396026611328, 47.49760437011719, -31.473907470703125, -1.1759071350097656, 18.383453369140625, 33.4136962890625, -11.0057373046875, 67.09410095214844, 34.512550354003906, 44.12445831298828, 38.20552444458008, -5.4703369140625, 7.463493347167969, 9.221549987792969, 33.46588134765625, -2.8773117065429688, 65.95917510986328, -9.595352172851562, 34.60323715209961, 10.008216857910156, 88.32058715820312, 48.812835693359375, -5.307107925415039, -7.451923370361328, 43.157981872558594, -6.528583526611328, -1.6533966064453125, -15.218154907226562, -2.8538055419921875, 14.214981079101562, 39.608619689941406, -10.008903503417969, 2.2992172241210938, 63.15283203125, 16.551883697509766, 29.247920989990234, 49.84246826171875, 43.054954528808594, 15.863513946533203, -3.219240188598633, 66.37411499023438, -24.933284759521484, 38.710243225097656, 39.325439453125, 32.24822998046875, 33.18461608886719, -11.13210678100586, -3.7445068359375, -26.122848510742188, 1.8843631744384766, 1.53118896484375, 27.087257385253906, 6.236286163330078, 7.626747131347656, -10.447555541992188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000294.npy"}
{"epoch": 0.4444444444444444, "step": 295, "batch_size": 64, "mean": 14.720478057861328, "std": 29.42466163635254, "min": -62.72453308105469, "p10": -20.89017372131347, "median": 9.264568328857422, "p90": 53.94102172851565, "max": 85.5917739868164, "pos_frac": 0.765625, "sample": [43.583740234375, 6.087841033935547, -5.133537292480469, -31.935165405273438, 57.532657623291016, 40.72181701660156, 44.773651123046875, 16.744014739990234, 62.882503509521484, 4.9037628173828125, 85.5917739868164, 9.722663879394531, 56.76768493652344, 19.950546264648438, 1.3162117004394531, 8.806472778320312, 0.8894386291503906, 12.359016418457031, 3.345062255859375, 35.384552001953125, 28.5927734375, 8.285411834716797, 2.3290634155273438, 21.00467300415039, 28.50396728515625, -32.13783264160156, -2.897003173828125, -31.995763778686523, 27.069595336914062, 0.15789794921875, -2.6450042724609375, -5.379634857177734, 76.43121337890625, 36.2935791015625, 26.541786193847656, -14.213939666748047, 47.34547424316406, 5.793460845947266, 38.452877044677734, 66.0047607421875, 28.02545166015625, -62.72453308105469, 3.5260391235351562, 6.35498046875, 4.545166015625, 29.16747283935547, -10.219127655029297, 15.356155395507812, -23.50267791748047, 17.666765213012695, 0.8945693969726562, 3.166961669921875, -58.49267578125, -0.0712432861328125, -24.231151580810547, 31.556058883666992, 28.341472625732422, 28.797348022460938, 4.599029541015625, 2.3906707763671875, 29.60516357421875, 26.53082275390625, 77.79011535644531, -14.794330596923828], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000295.npy"}
{"epoch": 0.4459561602418745, "step": 296, "batch_size": 64, "mean": 22.944568634033203, "std": 33.72062301635742, "min": -48.554351806640625, "p10": -16.991185188293453, "median": 18.52495574951172, "p90": 72.01793365478517, "max": 80.44465637207031, "pos_frac": 0.71875, "sample": [32.75202941894531, 70.9039535522461, 21.703933715820312, 17.68968391418457, -1.679046630859375, 18.389846801757812, 7.575893402099609, 77.11544799804688, -12.920928955078125, 36.71583557128906, 21.499935150146484, 72.67111206054688, -14.414190292358398, 79.93220520019531, 5.834831237792969, 15.4384765625, 60.582725524902344, 56.20793151855469, 6.662494659423828, 25.055274963378906, -1.4405364990234375, -22.96289825439453, 56.199424743652344, -3.9493408203125, -7.699981689453125, -33.982513427734375, 4.135231018066406, 21.84445571899414, 2.9357452392578125, 10.4661865234375, -42.63278579711914, -26.17193603515625, -0.6683197021484375, 31.551239013671875, 68.79632568359375, 13.924530029296875, 0.5542373657226562, -4.9074554443359375, 51.654232025146484, -18.095611572265625, 25.222885131835938, 17.697357177734375, 80.44465637207031, -48.554351806640625, 43.053253173828125, 18.660064697265625, 43.63933563232422, 54.01097869873047, 50.646793365478516, 69.58119201660156, 44.43165588378906, -13.0439453125, 72.49535369873047, 30.702552795410156, 78.42790985107422, 15.865615844726562, -40.89793395996094, 78.9665756225586, 4.9620208740234375, -4.145294189453125, -0.8154449462890625, 61.487518310546875, 39.43841552734375, 48.907466888427734], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000296.npy"}
{"epoch": 0.4474678760393046, "step": 297, "batch_size": 64, "mean": 24.06473159790039, "std": 28.845727920532227, "min": -54.531585693359375, "p10": -5.729334259033203, "median": 16.2528133392334, "p90": 66.78395767211914, "max": 76.4254150390625, "pos_frac": 0.75, "sample": [46.120994567871094, -5.9415740966796875, 12.877487182617188, 34.658790588378906, 21.081085205078125, 53.641326904296875, 14.223445892333984, 67.66253662109375, -6.963920593261719, -9.637893676757812, 28.754608154296875, 20.087440490722656, 7.200389862060547, 26.49871063232422, 13.329160690307617, 76.4254150390625, 11.006866455078125, 46.511688232421875, 10.91363525390625, 35.636199951171875, 7.096260070800781, -10.00286865234375, -13.420623779296875, -54.531585693359375, 3.320493698120117, 76.16387939453125, 13.59625244140625, -0.2661590576171875, 63.41120529174805, 5.586650848388672, 55.953277587890625, 73.30982971191406, -1.2818145751953125, -0.7004623413085938, 22.165035247802734, 15.463115692138672, 4.514926910400391, -3.0790252685546875, 39.03621292114258, 17.042510986328125, 67.97152709960938, -1.1513938903808594, 42.32329559326172, 74.17028045654297, 12.377464294433594, 34.108245849609375, -29.67554473876953, -1.4193830490112305, 6.364418029785156, 55.023590087890625, 10.619476318359375, 34.53394317626953, 45.57373809814453, 32.72735595703125, -1.1250572204589844, -5.234107971191406, 64.73394012451172, 57.48565673828125, -1.6166534423828125, 75.15018463134766, 6.7678985595703125, 26.05915641784668, 64.09837341308594, 52.812965393066406], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000297.npy"}
{"epoch": 0.4489795918367347, "step": 298, "batch_size": 64, "mean": 19.878402709960938, "std": 36.7667236328125, "min": -46.42131423950195, "p10": -25.77819595336914, "median": 9.734533309936523, "p90": 72.80496749877929, "max": 93.13105773925781, "pos_frac": 0.734375, "sample": [7.441993713378906, -39.367332458496094, 84.509033203125, 7.3958587646484375, 75.62460327148438, -2.8265380859375, -14.094221115112305, 13.515899658203125, 9.572956085205078, 15.918699264526367, 69.34591674804688, 76.11349487304688, -24.400802612304688, -3.41986083984375, -11.086776733398438, 72.67288970947266, 45.123443603515625, -26.368507385253906, 65.20805358886719, 32.284820556640625, 4.923358917236328, -15.925664901733398, 50.307193756103516, 55.18577575683594, 8.337638854980469, -2.0673370361328125, -44.668922424316406, 29.015132904052734, 39.53546142578125, 82.90194702148438, 0.7169151306152344, 63.31489944458008, 8.479652404785156, 56.97306823730469, 26.4737548828125, 75.99042510986328, 24.69252586364746, 56.40685272216797, 11.4482421875, 3.9480209350585938, 8.554439544677734, 2.0665626525878906, -0.20926284790039062, 12.324440002441406, 9.006401062011719, 93.13105773925781, 62.37812805175781, -21.845123291015625, 1.6826362609863281, 9.896110534667969, 0.21727371215820312, 4.583648681640625, -13.19329833984375, -32.841590881347656, -44.41204833984375, 47.42402648925781, -43.915496826171875, 57.40748596191406, 3.744171142578125, -46.42131423950195, 69.19203186035156, 20.007896423339844, 11.425556182861328, 72.861572265625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000298.npy"}
{"epoch": 0.4504913076341648, "step": 299, "batch_size": 64, "mean": 16.102848052978516, "std": 34.262107849121094, "min": -66.75738525390625, "p10": -25.468308639526363, "median": 10.354565620422363, "p90": 64.94452514648438, "max": 96.85224914550781, "pos_frac": 0.671875, "sample": [40.6213264465332, 11.12445068359375, -6.981929779052734, -5.7612152099609375, 5.7640380859375, 29.415918350219727, 1.0290985107421875, 15.944793701171875, 68.69747924804688, -2.5836925506591797, -3.8705291748046875, 38.49774169921875, -53.36486053466797, 6.35601806640625, -34.11231994628906, 5.409507751464844, 17.501663208007812, 65.04194641113281, 38.31090545654297, 13.227455139160156, 81.55683898925781, -8.435653686523438, -27.90321922302246, 39.58982849121094, -29.39361572265625, -18.48284912109375, -23.01260757446289, 4.980224609375, 94.85086059570312, -3.263080596923828, 64.71720886230469, 26.461196899414062, 38.50397872924805, -66.75738525390625, -1.3593292236328125, -26.520751953125, 43.27796936035156, 57.56296920776367, 0.916717529296875, 8.454185485839844, 30.23212432861328, -5.7021484375, 53.264400482177734, 25.037612915039062, 96.85224914550781, -38.434906005859375, 9.802392959594727, 65.77052307128906, 38.30126953125, 56.44334030151367, 21.609115600585938, 5.306182861328125, 49.70609664916992, -22.311172485351562, -7.934326171875, 41.35753631591797, 22.500335693359375, 10.90673828125, 0.6891727447509766, 9.552688598632812, 17.92507553100586, -4.831092834472656, -19.237655639648438, 67.7655029296875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000299.npy"}
{"epoch": 0.4520030234315949, "step": 300, "batch_size": 64, "mean": 18.007240295410156, "std": 35.59879684448242, "min": -67.64913940429688, "p10": -29.317571830749507, "median": 16.548803329467773, "p90": 65.82204742431642, "max": 89.20433044433594, "pos_frac": 0.703125, "sample": [27.196704864501953, 13.632484436035156, 5.021209716796875, 47.48149108886719, 1.5273818969726562, -26.48162841796875, 15.349296569824219, 11.079353332519531, -39.10466003417969, -53.28477478027344, 89.20433044433594, 53.88616943359375, 49.371646881103516, 39.3729248046875, 81.8064193725586, 17.833656311035156, 49.911834716796875, -24.231491088867188, -1.3685302734375, 60.245086669921875, 6.875080108642578, 68.21217346191406, 24.050567626953125, 0.5572490692138672, -31.252830505371094, 19.703048706054688, 30.499801635742188, 19.459564208984375, 52.12439727783203, 17.118144989013672, -3.105113983154297, -67.64913940429688, 51.97431945800781, -1.5029296875, -39.74676513671875, 25.49028778076172, -0.12208175659179688, 83.52665710449219, 26.225387573242188, 51.308998107910156, 15.979461669921875, 5.030429840087891, -6.724365234375, 79.45880126953125, 42.769710540771484, -66.26303100585938, 1.759490966796875, 34.92569351196289, -5.1822052001953125, 54.557071685791016, 72.28282165527344, -1.8870811462402344, -6.8956756591796875, 18.567974090576172, 2.31866455078125, -30.532976150512695, 6.373344421386719, 36.327186584472656, 45.53129577636719, -12.793037414550781, 70.06309509277344, 39.0450439453125, 12.386337280273438, -6.830347061157227], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000300.npy"}
{"epoch": 0.45351473922902497, "step": 301, "batch_size": 64, "mean": 18.749427795410156, "std": 34.96540832519531, "min": -54.84343719482422, "p10": -23.59882011413574, "median": 10.96214771270752, "p90": 70.99088058471679, "max": 94.84483337402344, "pos_frac": 0.640625, "sample": [62.85032653808594, 19.204803466796875, 3.4179229736328125, 10.781352996826172, 28.209022521972656, -33.30481719970703, 85.47313690185547, 41.31199264526367, 33.61187744140625, 48.079185485839844, -1.98040771484375, 16.459983825683594, 48.96287536621094, -14.666671752929688, -43.884071350097656, 10.101776123046875, -4.210968017578125, 13.045822143554688, 41.18287658691406, 0.5899734497070312, -5.837467193603516, 28.39075469970703, 71.09600830078125, -14.711620330810547, 45.19451904296875, 36.750526428222656, 0.7663249969482422, 70.7455825805664, -0.301422119140625, -16.98187255859375, -25.76593017578125, -14.805217742919922, 50.690513610839844, 81.1053466796875, -4.202156066894531, -2.071868896484375, -20.35662841796875, 94.84483337402344, -24.988330841064453, 27.758113861083984, -1.4223785400390625, 44.17840576171875, 9.023639678955078, -54.84343719482422, 1.6790800094604492, 37.30284881591797, 22.31793975830078, 24.065643310546875, -8.493316650390625, 46.12065124511719, 11.142942428588867, 72.79322814941406, 57.62809371948242, -6.501674652099609, -6.176750183105469, 33.175228118896484, 73.10775756835938, 89.1160888671875, -5.5517120361328125, -28.699691772460938, 5.437705993652344, -29.234458923339844, 69.39350891113281, 1.84796142578125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000301.npy"}
{"epoch": 0.455026455026455, "step": 302, "batch_size": 64, "mean": 9.814539909362793, "std": 31.486413955688477, "min": -65.33845520019531, "p10": -30.20276527404785, "median": 7.4741363525390625, "p90": 48.87826538085939, "max": 91.80548095703125, "pos_frac": 0.6875, "sample": [27.25564193725586, 45.90965270996094, 1.147735595703125, 91.80548095703125, 4.278377532958984, 42.340980529785156, 67.25188446044922, -23.061119079589844, 28.325225830078125, 6.175586700439453, 17.728797912597656, 8.077789306640625, 18.292694091796875, 13.006542205810547, 2.957489013671875, 75.4759750366211, 8.331466674804688, 1.8082504272460938, -8.536766052246094, 10.66131591796875, 2.1384506225585938, -43.16316223144531, -65.33845520019531, 7.5102386474609375, -3.83050537109375, -10.41891098022461, 18.349849700927734, 73.14339447021484, -5.132083892822266, 1.9111175537109375, -27.741104125976562, 39.25672912597656, 16.67486572265625, -0.7570829391479492, -31.257762908935547, 50.15052795410156, 74.60742950439453, -36.719512939453125, 3.4438018798828125, -10.348945617675781, 25.470918655395508, 26.43488311767578, -1.3673744201660156, 27.505210876464844, 7.4380340576171875, 5.30950927734375, -57.90599060058594, 11.179611206054688, -3.5245513916015625, 13.484085083007812, 17.351755142211914, 15.041969299316406, 34.729278564453125, -11.145332336425781, 14.52850341796875, -41.91014862060547, 30.12249755859375, 67.04181671142578, -10.533660888671875, 4.677204132080078, -55.49608612060547, -2.834442138671875, 17.668655395507812, 3.1522865295410156], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000302.npy"}
{"epoch": 0.4565381708238851, "step": 303, "batch_size": 64, "mean": 24.73933982849121, "std": 39.56116485595703, "min": -67.75694274902344, "p10": -23.779420089721672, "median": 22.085346221923828, "p90": 76.5884979248047, "max": 96.93994140625, "pos_frac": 0.734375, "sample": [1.668914794921875, 47.44557189941406, 45.320343017578125, 22.496826171875, 90.86729431152344, 61.936920166015625, 49.81660461425781, -3.101522445678711, 30.675382614135742, 4.06842041015625, 43.00830078125, -31.32888412475586, 5.79034423828125, -42.71004867553711, -2.0869140625, -37.35819625854492, -52.5198974609375, 77.2607421875, -17.02185821533203, 6.144317626953125, -0.07694053649902344, 48.67768859863281, 96.93994140625, 72.85367584228516, -67.03138732910156, 8.997779846191406, 81.79127502441406, -67.75694274902344, -2.768829345703125, 0.26666259765625, 38.37788391113281, 13.065818786621094, 9.743518829345703, 1.9301948547363281, -1.2029876708984375, -4.202392578125, 80.77336120605469, 8.736839294433594, 73.33037567138672, 58.6705322265625, 23.308929443359375, 59.608612060546875, 39.819862365722656, 48.299591064453125, -9.0133056640625, 45.595909118652344, 96.40425109863281, 28.631546020507812, 26.888137817382812, 44.521728515625, -16.397186279296875, 21.673866271972656, -4.397064208984375, 8.208892822265625, 75.01992797851562, 49.916259765625, 10.017410278320312, 8.217498779296875, 66.43510437011719, 19.759689331054688, 90.7874755859375, -26.675518035888672, 59.8414306640625, 65.35591888427734], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000303.npy"}
{"epoch": 0.4580498866213152, "step": 304, "batch_size": 64, "mean": 22.9644832611084, "std": 31.50005340576172, "min": -62.12651062011719, "p10": -3.7144783020019525, "median": 16.52537250518799, "p90": 72.41739501953126, "max": 90.21497344970703, "pos_frac": 0.828125, "sample": [62.42132568359375, 0.6729698181152344, 9.106704711914062, 83.12762451171875, -8.181053161621094, 44.60285568237305, 8.623493194580078, 30.599946975708008, -17.015905380249023, -0.2526512145996094, 22.368759155273438, 53.98393249511719, 30.13787841796875, 0.33431243896484375, 13.900581359863281, 7.51829719543457, -14.710521697998047, 10.33547592163086, 90.21497344970703, -42.58543395996094, 9.317054748535156, 21.98932647705078, 6.9871673583984375, 15.660852432250977, -0.20484161376953125, -31.125640869140625, 20.278799057006836, 21.926307678222656, 7.7088165283203125, 61.67228317260742, 48.08909606933594, -3.077411651611328, 72.61026000976562, 20.048564910888672, 12.572608947753906, -62.12651062011719, 36.41968536376953, 9.410087585449219, 21.437049865722656, 26.396240234375, 22.46980094909668, -3.987506866455078, 79.70442199707031, 24.972335815429688, 7.192142486572266, 0.6480178833007812, 7.714727401733398, 38.984169006347656, 60.9150390625, 71.96737670898438, 7.842124938964844, 8.092247009277344, 81.14217376708984, 3.3038501739501953, 48.59929656982422, -2.069000244140625, 35.29759979248047, 84.1689682006836, 84.28065490722656, 5.441062927246094, 4.1638031005859375, 17.389892578125, 19.249298095703125, 61.05107879638672], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000304.npy"}
{"epoch": 0.4595616024187453, "step": 305, "batch_size": 64, "mean": 20.260169982910156, "std": 38.28538131713867, "min": -85.55868530273438, "p10": -24.35029296875, "median": 14.856353759765625, "p90": 75.43520660400391, "max": 99.04432678222656, "pos_frac": 0.71875, "sample": [76.69985961914062, -24.357315063476562, 4.071380615234375, -4.287330627441406, 5.979461669921875, 4.578563690185547, 84.48737335205078, -29.016124725341797, 23.970260620117188, 58.68952941894531, 0.9163818359375, 68.81068420410156, -8.83758544921875, 57.92977523803711, -11.5811767578125, -41.191158294677734, -21.341461181640625, 3.4465103149414062, -0.8668975830078125, 81.2447509765625, 2.8202781677246094, 26.59407615661621, 47.120574951171875, 41.635833740234375, -13.35654067993164, 18.025665283203125, 63.136783599853516, 1.4860687255859375, 21.002471923828125, -85.55868530273438, 4.5859222412109375, 28.00674057006836, 64.79684448242188, 27.64977264404297, -37.780670166015625, 4.164777755737305, 99.04432678222656, 3.34844970703125, 3.5199975967407227, 55.97428894042969, -2.3370132446289062, 30.915611267089844, 73.55670166015625, 80.85063171386719, 27.729270935058594, -37.89385223388672, 4.4404296875, -43.47441101074219, 69.69673919677734, 76.24028015136719, 11.687042236328125, 51.366294860839844, 18.05567169189453, 87.94829559326172, 26.160171508789062, 36.97959899902344, -4.0771026611328125, 18.829345703125, -24.333908081054688, 62.32036590576172, -7.302146911621094, 1.6982479095458984, 41.936279296875, -9.904075622558594], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000305.npy"}
{"epoch": 0.46107331821617537, "step": 306, "batch_size": 64, "mean": 21.383968353271484, "std": 32.39569854736328, "min": -45.05817413330078, "p10": -10.485234069824218, "median": 16.264976501464844, "p90": 62.32926750183106, "max": 101.82681274414062, "pos_frac": 0.71875, "sample": [82.90148162841797, 1.5999679565429688, -2.880084991455078, 60.031925201416016, 0.32532501220703125, 29.063365936279297, 6.992340087890625, 6.471672058105469, 20.296619415283203, 23.75570297241211, 50.92365646362305, 28.63420867919922, -45.05817413330078, 4.00872802734375, -0.2963981628417969, -8.219512939453125, 58.2508544921875, 11.898956298828125, 49.0323600769043, 29.084304809570312, 36.488433837890625, 81.35554504394531, 47.63233947753906, -6.436515808105469, 39.71173095703125, -3.9564895629882812, -41.849830627441406, 29.39588165283203, 3.9706039428710938, 21.83563232421875, -9.252426147460938, 73.09297943115234, -6.02410888671875, 7.011177062988281, 49.229427337646484, 14.349960327148438, 63.3138427734375, 99.3544921875, 1.9864654541015625, -13.855175018310547, -1.2570419311523438, 1.4337158203125, 28.566070556640625, -2.5806503295898438, -31.590625762939453, 25.23989486694336, 0.8666038513183594, 44.61305236816406, 5.312984466552734, 55.61073684692383, 50.22492980957031, 41.49393081665039, -11.013580322265625, -6.108116149902344, 54.63154602050781, 18.17999267578125, 27.185943603515625, -13.171211242675781, 69.51582336425781, 101.82681274414062, -33.859134674072266, 44.039093017578125, 6.89764404296875, -1.655655860900879], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000306.npy"}
{"epoch": 0.46258503401360546, "step": 307, "batch_size": 64, "mean": 13.357641220092773, "std": 33.411258697509766, "min": -49.662017822265625, "p10": -23.92884140014648, "median": 6.0253190994262695, "p90": 62.92711029052735, "max": 98.93318939208984, "pos_frac": 0.578125, "sample": [37.766845703125, -25.51659393310547, 15.573356628417969, 5.191253662109375, -24.461570739746094, -0.670135498046875, 98.93318939208984, -1.1150360107421875, 63.706878662109375, 79.12654113769531, -12.740119934082031, 4.501445770263672, 34.922584533691406, 61.10765075683594, -49.662017822265625, 32.070953369140625, 14.096420288085938, -2.9750823974609375, 43.82501983642578, -14.112464904785156, 14.485397338867188, -9.291305541992188, -8.310188293457031, -47.57963562011719, 2.827838897705078, -3.2249412536621094, -8.556404113769531, 23.531639099121094, 23.37939453125, -0.1655731201171875, -12.093376159667969, -0.1841888427734375, 1.9547576904296875, -35.997276306152344, -22.685806274414062, 31.24853515625, -19.36572265625, 10.927780151367188, 45.51447296142578, -21.497650146484375, 22.28622055053711, 6.099628448486328, 12.662830352783203, 49.59629821777344, 84.6333236694336, -42.71168518066406, 5.951009750366211, 29.235450744628906, 89.01439666748047, 47.858009338378906, -29.50665283203125, 66.77944946289062, 36.26478576660156, -0.9017925262451172, 11.556694030761719, 17.12187957763672, -5.5056610107421875, 39.041168212890625, 38.43555450439453, -22.201583862304688, -7.312189102172852, 19.996665954589844, 68.72469329833984, -6.716259002685547], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000307.npy"}
{"epoch": 0.46409674981103555, "step": 308, "batch_size": 64, "mean": 21.821016311645508, "std": 35.97514343261719, "min": -48.66758728027344, "p10": -21.838931465148924, "median": 17.602224349975586, "p90": 73.28706741333008, "max": 99.05943298339844, "pos_frac": 0.6875, "sample": [68.41523742675781, 43.08409118652344, -2.25067138671875, -15.382270812988281, 56.35563659667969, 47.18426513671875, 2.706209182739258, 6.411407470703125, 41.498260498046875, -47.344200134277344, 99.05943298339844, 37.00331115722656, 58.73081970214844, 19.561115264892578, -12.184280395507812, -36.899593353271484, 39.82696533203125, 6.517543792724609, 58.048179626464844, 9.616920471191406, 74.64886474609375, -8.829334259033203, 71.82872772216797, 38.07904052734375, 13.390388488769531, 52.68714141845703, -23.176557540893555, 80.80374145507812, -18.717803955078125, -45.43659973144531, 79.49592590332031, 14.324386596679688, 22.867324829101562, -9.061981201171875, 34.78205108642578, -1.9358596801757812, 30.660064697265625, -8.205554962158203, 8.648773193359375, -48.66758728027344, -2.9005661010742188, 8.656837463378906, 31.511764526367188, 72.42902374267578, 13.000900268554688, 34.25727081298828, 27.102203369140625, -14.514877319335938, 17.609783172607422, 34.732887268066406, 73.65480041503906, 17.59466552734375, -15.530548095703125, 49.51643371582031, 96.8748779296875, -4.7645721435546875, 76.88983917236328, -28.96033477783203, 51.523040771484375, -3.0848846435546875, 15.287443161010742, -27.297836303710938, 12.709674835205078, 22.103710174560547], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000308.npy"}
{"epoch": 0.4656084656084656, "step": 309, "batch_size": 64, "mean": 21.988815307617188, "std": 31.484336853027344, "min": -37.487548828125, "p10": -10.695086669921873, "median": 14.996929168701172, "p90": 64.74521865844729, "max": 94.97908020019531, "pos_frac": 0.765625, "sample": [38.51305389404297, 94.97908020019531, 13.413673400878906, 77.38078308105469, 55.018226623535156, -8.38970947265625, 59.381309509277344, 5.23712158203125, 16.22662353515625, -4.384624481201172, 38.36114501953125, -1.3561382293701172, -19.724170684814453, 8.32845687866211, 6.085479736328125, -35.81549072265625, 16.537151336669922, 29.819137573242188, -13.46636962890625, 48.117431640625, 4.181859970092773, 88.56747436523438, 58.771820068359375, -8.236648559570312, 67.04403686523438, 56.93832015991211, -5.407146453857422, 10.632091522216797, 9.101455688476562, 0.00144195556640625, 0.3093109130859375, 18.89300537109375, -5.8757171630859375, 19.162864685058594, 57.987030029296875, 13.767234802246094, 2.4955368041992188, 34.112144470214844, -15.036483764648438, 83.70893859863281, 2.2032470703125, 34.20039367675781, 58.705406188964844, 16.311309814453125, 78.12057495117188, -37.487548828125, 45.045135498046875, 31.04236602783203, -5.898220062255859, 42.84798812866211, -11.68310546875, 0.956024169921875, 9.719825744628906, 51.019615173339844, 93.89976501464844, 2.9798660278320312, 28.261730194091797, 24.214332580566406, -19.933029174804688, -4.55322265625, 4.932952880859375, 2.2877044677734375, 16.97952651977539, 27.730697631835938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000309.npy"}
{"epoch": 0.4671201814058957, "step": 310, "batch_size": 64, "mean": 10.018726348876953, "std": 37.817935943603516, "min": -56.72772979736328, "p10": -38.441063690185544, "median": 4.91572380065918, "p90": 66.64852142333984, "max": 93.41748046875, "pos_frac": 0.625, "sample": [93.41748046875, -3.000885009765625, 26.02161407470703, 37.61460876464844, 36.552490234375, -36.62113952636719, -20.172531127929688, 12.171577453613281, 64.37088012695312, -11.54083251953125, -8.46920394897461, -47.717567443847656, 1.1478328704833984, 66.31169128417969, 1.568603515625, 2.155853271484375, -2.069242477416992, 35.980804443359375, 79.7884292602539, -39.221031188964844, 5.054149627685547, -7.761741638183594, -51.397987365722656, 50.14129638671875, -56.72772979736328, -17.760040283203125, 4.7772979736328125, -53.37902069091797, -20.439300537109375, 69.68006896972656, -31.646209716796875, 8.147262573242188, 1.6461639404296875, 68.02915954589844, 8.813156127929688, 20.1221923828125, 6.774192810058594, 43.727760314941406, 12.8648681640625, 24.795761108398438, -22.64185333251953, 11.723129272460938, -33.89849853515625, -54.568267822265625, 25.51641845703125, 2.4830474853515625, -7.090568542480469, 28.32733154296875, 4.714271545410156, -36.205963134765625, 23.946990966796875, 81.2948989868164, 66.79287719726562, -34.207210540771484, 1.4295482635498047, 11.058284759521484, -3.779541015625, 35.85282516479492, 64.645263671875, -5.678947448730469, -44.96476745605469, 81.75989532470703, 19.18339729309082, 51.75520706176758], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000310.npy"}
{"epoch": 0.46863189720332576, "step": 311, "batch_size": 64, "mean": 18.61467933654785, "std": 34.73135757446289, "min": -55.47068786621094, "p10": -18.005738449096675, "median": 11.539545059204102, "p90": 66.42808227539068, "max": 128.82257080078125, "pos_frac": 0.75, "sample": [54.65132141113281, -1.6558189392089844, 54.519554138183594, -43.299278259277344, -19.434463500976562, -5.832332611083984, 38.904666900634766, 34.63124084472656, -21.974773406982422, 22.704021453857422, 11.059158325195312, -49.47822570800781, 4.134120941162109, 6.302223205566406, 24.565704345703125, -1.2739067077636719, 39.03089904785156, 10.437820434570312, 50.092529296875, 29.990478515625, 0.973175048828125, 44.589935302734375, 1.5132122039794922, 15.473129272460938, 26.583770751953125, 2.963733673095703, -55.47068786621094, -14.672046661376953, -8.158920288085938, 35.051177978515625, 0.66748046875, 72.53378295898438, 107.84503173828125, 3.2276039123535156, 5.23492431640625, 71.47526550292969, 16.749473571777344, 24.093414306640625, 27.35943603515625, 32.43449401855469, 128.82257080078125, 15.615188598632812, -8.569198608398438, 83.49313354492188, -22.718280792236328, -4.2946929931640625, 9.028732299804688, 0.07384490966796875, 9.556327819824219, 13.599048614501953, 97.85060119628906, 6.449920654296875, 1.0381622314453125, -5.068357467651367, 12.01993179321289, 35.228355407714844, 10.552526473999023, -20.040817260742188, -10.149459838867188, 13.6832275390625, 22.501251220703125, 86.65078735351562, 31.33904266357422, 36.13529968261719], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000311.npy"}
{"epoch": 0.47014361300075586, "step": 312, "batch_size": 64, "mean": 26.31338119506836, "std": 37.9128532409668, "min": -48.25700378417969, "p10": -13.360794830322265, "median": 11.319777488708496, "p90": 86.83178176879885, "max": 102.97763061523438, "pos_frac": 0.75, "sample": [33.592041015625, 19.220260620117188, -35.23893737792969, -4.679450988769531, 7.786014556884766, 40.84710693359375, 0.10766220092773438, 1.1521759033203125, -0.6720809936523438, 7.0297698974609375, 101.30701446533203, 26.722564697265625, 31.986900329589844, 8.131221771240234, 5.098606109619141, 71.02233123779297, -8.569267272949219, 18.840641021728516, 1.1343650817871094, 47.80499267578125, 94.23796081542969, 4.241260528564453, -20.306156158447266, 17.494224548339844, 3.53387451171875, 89.32537841796875, -17.969131469726562, -1.780731201171875, 36.834808349609375, 60.53254318237305, 90.83712768554688, -8.659294128417969, 65.34536743164062, 91.61740112304688, -48.25700378417969, 10.956871032714844, 70.37417602539062, -2.5749740600585938, 56.29905700683594, 74.88400268554688, 102.97763061523438, 10.813941955566406, -19.675827026367188, 1.721832275390625, 43.76722717285156, -12.2896728515625, 0.8787155151367188, 83.3414535522461, 75.515380859375, 2.1858673095703125, -31.034278869628906, 11.682683944702148, 78.95526885986328, 51.473548889160156, -6.8098907470703125, 63.119972229003906, 53.11090087890625, 21.858978271484375, -13.819847106933594, 88.32763671875, 22.071788787841797, 6.4596099853515625, 10.712772369384766, -0.880035400390625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000312.npy"}
{"epoch": 0.47165532879818595, "step": 313, "batch_size": 64, "mean": 15.82878589630127, "std": 38.55817413330078, "min": -63.02177429199219, "p10": -23.01639823913574, "median": 8.014183521270752, "p90": 77.87153930664064, "max": 108.04583740234375, "pos_frac": 0.671875, "sample": [11.558038711547852, 19.885574340820312, 23.291751861572266, -5.505641937255859, 5.4276885986328125, 71.21369934082031, 15.858695983886719, 25.47883415222168, 17.81131362915039, -57.05352020263672, 0.05335426330566406, 7.889198303222656, 15.19162368774414, 14.797904968261719, 4.954254150390625, -11.008842468261719, 3.8909225463867188, -13.529125213623047, -24.07427978515625, -4.787319183349609, 13.754592895507812, 52.53410720825195, 5.199871063232422, 99.5305404663086, -36.048038482666016, -20.54800796508789, 33.555442810058594, 44.975128173828125, 21.5914306640625, 12.269569396972656, 103.5276107788086, 88.10787963867188, -36.303619384765625, -19.609210968017578, 96.13694763183594, 25.5537109375, 8.139168739318848, -5.256168365478516, -9.011024475097656, 79.25616455078125, 8.838790893554688, 7.8173828125, 46.209564208984375, -7.889305114746094, 108.04583740234375, 1.9447174072265625, 6.184272766113281, -33.541725158691406, 9.914443969726562, 1.1114006042480469, 71.32649993896484, -1.11285400390625, -0.473175048828125, 0.31237220764160156, -12.186851501464844, -63.02177429199219, 74.6407470703125, 10.04763412475586, 70.49019622802734, -19.397323608398438, -39.74186706542969, -5.12237548828125, 18.011810302734375, 81.93357849121094], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000313.npy"}
{"epoch": 0.47316704459561604, "step": 314, "batch_size": 64, "mean": 26.446006774902344, "std": 38.89383316040039, "min": -38.373985290527344, "p10": -13.1004337310791, "median": 11.662986755371094, "p90": 88.61142883300782, "max": 104.67013549804688, "pos_frac": 0.671875, "sample": [4.690406799316406, -0.4573974609375, 5.405155181884766, 92.58894348144531, -8.79608154296875, 74.44139099121094, -1.5810394287109375, 67.43812561035156, 2.218292236328125, 25.709136962890625, 2.3980789184570312, -3.7834243774414062, 104.67013549804688, 53.23081970214844, -1.4410476684570312, 84.80148315429688, 11.517868041992188, 76.75383758544922, 91.57254791259766, 40.592613220214844, 11.03521728515625, -7.778839111328125, 73.10688018798828, 99.26876831054688, 58.01142120361328, -38.373985290527344, -2.961578369140625, 90.2442626953125, 12.77191162109375, -15.53662109375, 37.34046173095703, 35.14509582519531, 24.554645538330078, 1.1264190673828125, 1.0595359802246094, 23.47228240966797, 57.954833984375, 11.00244140625, -7.714569091796875, 34.18768310546875, -13.36074447631836, -20.712867736816406, 91.9437026977539, 75.44021606445312, 42.70947265625, -30.786376953125, 45.4737434387207, 60.7603874206543, -2.06591796875, 77.50108337402344, 9.57489013671875, 98.93183898925781, 11.57501220703125, -7.1627960205078125, 11.750961303710938, -12.4930419921875, -33.24523162841797, 28.939674377441406, -12.454765319824219, -7.531463623046875, 15.114343643188477, -18.228696823120117, -11.686355590820312, 72.67113494873047], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000314.npy"}
{"epoch": 0.47467876039304613, "step": 315, "batch_size": 64, "mean": 21.451175689697266, "std": 33.2592887878418, "min": -36.930259704589844, "p10": -15.321237564086912, "median": 13.11958122253418, "p90": 67.68661041259766, "max": 102.16983032226562, "pos_frac": 0.734375, "sample": [-11.46551513671875, 32.134796142578125, 39.11374282836914, -5.953556060791016, 0.9790878295898438, 30.984275817871094, -27.613754272460938, -5.039955139160156, 0.7948150634765625, 30.98712158203125, -33.107452392578125, 17.84673500061035, 19.463241577148438, 58.085838317871094, -3.7370529174804688, 32.576393127441406, -1.2186698913574219, -12.037101745605469, 3.5455398559570312, 15.577850341796875, 8.059803009033203, 18.922225952148438, 2.315593719482422, 11.984603881835938, -0.0700531005859375, 13.284664154052734, 73.4862060546875, -36.930259704589844, -13.713497161865234, 9.014617919921875, 12.954498291015625, 6.552787780761719, 26.073776245117188, 13.916839599609375, 12.273468017578125, 97.46248626708984, 60.230926513671875, -16.010269165039062, -19.31623077392578, 67.74209594726562, 57.81569290161133, 13.604507446289062, 61.55693054199219, 87.8062515258789, -0.7303466796875, 46.18806457519531, 11.277767181396484, 95.07360076904297, -22.918495178222656, 11.7340087890625, 60.327964782714844, 0.8614006042480469, 61.18621826171875, 0.18062782287597656, 27.391246795654297, 19.73003387451172, 67.55714416503906, 18.66461181640625, 81.23548889160156, -16.542850494384766, 54.151832580566406, -5.56257438659668, 102.16983032226562, 9.965591430664062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000315.npy"}
{"epoch": 0.47619047619047616, "step": 316, "batch_size": 64, "mean": 26.249847412109375, "std": 33.72225570678711, "min": -30.681434631347656, "p10": -10.729630279541015, "median": 23.12143325805664, "p90": 74.42426452636721, "max": 98.98816680908203, "pos_frac": 0.71875, "sample": [-17.748703002929688, 20.116592407226562, 48.31849670410156, -6.213352203369141, -2.6880645751953125, -11.131782531738281, 59.159271240234375, 13.640251159667969, 26.775657653808594, 53.570831298828125, 90.36268615722656, 96.11700439453125, 51.30004119873047, 90.75657653808594, 76.489501953125, -3.0154876708984375, -28.22088623046875, 26.12627410888672, -30.681434631347656, 53.47480773925781, 14.479782104492188, 53.00909423828125, 5.646888732910156, 3.50799560546875, -12.630966186523438, 33.018089294433594, 37.90941619873047, -18.755905151367188, 2.378387451171875, 29.627639770507812, 42.84290313720703, 57.75914001464844, -7.24468994140625, 82.40921020507812, -0.19553375244140625, -4.755130767822266, -4.71405029296875, 15.202659606933594, 1.0832366943359375, 15.407814025878906, 46.91026306152344, 39.63459014892578, 33.0043830871582, -6.691921234130859, 33.99702453613281, 5.986202239990234, -1.6403388977050781, 44.832855224609375, -28.884483337402344, 57.41706848144531, 4.36474609375, -7.430547714233398, 3.5225601196289062, -9.791275024414062, 33.3057746887207, 52.53125, 1.6903228759765625, 94.16178894042969, 42.088531494140625, 56.475006103515625, 69.60537719726562, 13.295967102050781, 98.98816680908203, 50.12279510498047], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000316.npy"}
{"epoch": 0.47770219198790626, "step": 317, "batch_size": 64, "mean": 24.650264739990234, "std": 41.813175201416016, "min": -111.52285766601562, "p10": -12.535662460327147, "median": 15.989572525024414, "p90": 81.83214645385743, "max": 112.2913589477539, "pos_frac": 0.828125, "sample": [34.14997100830078, 5.486114501953125, 15.659000396728516, 12.320144653320312, 32.19075012207031, 112.2913589477539, -15.45412826538086, 27.80572509765625, -0.2175750732421875, 4.872661590576172, -10.743282318115234, 0.3131103515625, -0.299591064453125, 12.514480590820312, 24.742855072021484, 6.394199371337891, 70.10560607910156, 10.598312377929688, 63.3116455078125, 53.76271438598633, -76.28394317626953, 0.23256683349609375, 29.04932403564453, 3.6336212158203125, 68.09761810302734, 12.389564514160156, 16.320144653320312, 30.252361297607422, 29.37712860107422, -111.52285766601562, 77.70055389404297, -20.2967529296875, 35.32350540161133, 107.6448974609375, 105.14444732666016, -51.484283447265625, 32.8128662109375, 6.2605133056640625, 38.52610778808594, 92.95758056640625, -9.123680114746094, 16.683700561523438, 101.19058227539062, -13.303825378417969, 1.3633956909179688, 10.465415954589844, 18.50627899169922, 88.05289459228516, 11.445060729980469, 4.392723083496094, 50.833946228027344, 53.19455337524414, 2.791799545288086, 72.77178192138672, -35.324405670166016, 12.01239013671875, 0.22740936279296875, 65.15408325195312, 53.2098274230957, 75.08749389648438, 0.4513397216796875, 83.60282897949219, 5.782989501953125, 22.209274291992188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000317.npy"}
{"epoch": 0.47921390778533635, "step": 318, "batch_size": 64, "mean": 21.925718307495117, "std": 43.72169876098633, "min": -58.993614196777344, "p10": -30.710823440551756, "median": 9.830099105834961, "p90": 89.28314437866212, "max": 114.57603454589844, "pos_frac": 0.671875, "sample": [41.12553405761719, 86.98040008544922, -6.2537994384765625, -48.07667541503906, 100.00273132324219, 22.784080505371094, 6.655364990234375, -47.890869140625, 62.68679428100586, 46.89651870727539, 6.414642333984375, -1.2767333984375, 15.864425659179688, 56.877479553222656, 104.93641662597656, 5.979198455810547, 25.340904235839844, -39.50236129760742, 25.8607120513916, 9.318675994873047, -22.41057014465332, 30.90187644958496, 90.27003479003906, 5.5015411376953125, 2.142333984375, 10.341522216796875, 97.17778778076172, -25.03771209716797, 74.2791748046875, 53.98334503173828, -18.531356811523438, 34.661956787109375, -19.971229553222656, 11.334480285644531, 4.5517578125, 81.232666015625, 114.57603454589844, -30.718002319335938, 92.21842956542969, 13.864082336425781, 1.2642822265625, -2.1144638061523438, 57.60340881347656, -15.872512817382812, -10.305984497070312, 14.702743530273438, 9.196563720703125, 53.287376403808594, -4.522674560546875, 67.9382095336914, -0.04151725769042969, 108.52891540527344, 27.614463806152344, 73.67939758300781, -30.694072723388672, -36.901004791259766, 9.155067443847656, -38.14978790283203, 35.81721496582031, -5.1116180419921875, 82.50411224365234, -58.993614196777344, -18.420373916625977, 7.990257263183594], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000318.npy"}
{"epoch": 0.48072562358276644, "step": 319, "batch_size": 64, "mean": 26.597312927246094, "std": 41.552955627441406, "min": -83.49974822998047, "p10": -12.925460815429688, "median": 22.061195373535156, "p90": 83.5127685546875, "max": 117.97090911865234, "pos_frac": 0.734375, "sample": [83.43069458007812, 66.3919906616211, -12.701904296875, 63.881256103515625, 39.64069366455078, 47.74658203125, 33.42151641845703, 2.9925079345703125, -8.783035278320312, 25.775737762451172, 77.53703308105469, 78.39654541015625, -0.12294578552246094, 20.152603149414062, 6.772735595703125, 31.608978271484375, -35.01850891113281, 117.97090911865234, 2.0149078369140625, 83.54794311523438, 32.647315979003906, 6.6594085693359375, 23.96978759765625, 59.38225173950195, 74.69505310058594, -8.382854461669922, 99.90046691894531, 30.495655059814453, 61.02241516113281, 77.92901611328125, -21.402671813964844, -19.43610382080078, 35.52363967895508, 109.87005615234375, -10.30864143371582, 30.888046264648438, -43.548583984375, 27.89142608642578, 3.2698516845703125, -0.058864593505859375, 12.52535629272461, 18.503925323486328, 4.406078338623047, 1.7841644287109375, 44.077919006347656, -7.1847381591796875, -13.021270751953125, -3.9394302368164062, -58.724403381347656, 50.220829010009766, -0.5978622436523438, 8.955093383789062, 17.73180389404297, 2.9158172607421875, 18.94342803955078, 85.28434753417969, 32.18634033203125, 44.90625, 99.36138916015625, 25.799076080322266, -5.5182647705078125, -83.49974822998047, 109.20418548583984, 2.2447509765625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000319.npy"}
{"epoch": 0.48223733938019653, "step": 320, "batch_size": 64, "mean": 14.949155807495117, "std": 43.94546127319336, "min": -82.6217269897461, "p10": -31.965448379516594, "median": 7.080024719238281, "p90": 73.04505157470703, "max": 118.26311492919922, "pos_frac": 0.6875, "sample": [47.969146728515625, -74.36277770996094, 2.2039623260498047, 68.19561767578125, 11.991401672363281, -7.803794860839844, 72.37539672851562, -0.9492778778076172, 15.509544372558594, 2.1859207153320312, 43.856201171875, 25.806018829345703, -3.7843971252441406, 26.242389678955078, -42.26874542236328, 89.59225463867188, 8.774581909179688, -82.6217269897461, -2.1009063720703125, 14.615516662597656, -17.019554138183594, 56.32921600341797, -80.63644409179688, 4.648859024047852, 5.385467529296875, 19.80584716796875, 118.26311492919922, 73.33204650878906, 0.7091255187988281, -80.22355651855469, 26.235069274902344, -34.18220901489258, 19.267391204833984, 4.185455322265625, 0.45301342010498047, 65.3106918334961, 3.2365570068359375, -11.509574890136719, -16.309669494628906, 19.03879165649414, 34.64314270019531, 105.62850952148438, -6.138072967529297, -23.458690643310547, 55.71666717529297, 3.5311126708984375, 88.67375183105469, 43.36472702026367, -26.793006896972656, 0.7293605804443359, 0.5929775238037109, -62.933685302734375, 98.86117553710938, -9.21923828125, 0.8156852722167969, 82.27510070800781, -3.3974151611328125, 48.92647933959961, 19.958740234375, 14.443046569824219, 49.431610107421875, -16.969139099121094, 54.95574188232422, 11.361443519592285], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000320.npy"}
{"epoch": 0.4837490551776266, "step": 321, "batch_size": 64, "mean": 24.483470916748047, "std": 46.093597412109375, "min": -64.07330322265625, "p10": -36.66706390380859, "median": 16.611507415771484, "p90": 92.0536117553711, "max": 116.42634582519531, "pos_frac": 0.78125, "sample": [31.396713256835938, 64.00526428222656, 1.3108901977539062, 79.73384094238281, 11.271942138671875, -53.77069854736328, 90.40470886230469, 80.44332122802734, 76.83798217773438, 1.697957992553711, 104.55439758300781, 4.4058990478515625, 28.749126434326172, 54.54916000366211, -64.07330322265625, 2.0441360473632812, 22.210861206054688, 19.69727325439453, 42.81684875488281, 92.76028442382812, 18.728904724121094, 10.214836120605469, -59.12729263305664, -13.365303039550781, 26.699085235595703, -62.57484436035156, -34.06475067138672, -37.78234100341797, 38.12535095214844, 4.3097686767578125, 0.43354034423828125, 16.094600677490234, 52.22752380371094, 12.861083984375, 17.983924865722656, 116.42634582519531, 16.739364624023438, -0.2865753173828125, 23.73529815673828, 83.72113037109375, -2.8093299865722656, 3.0748062133789062, -16.510845184326172, 30.286602020263672, 11.835563659667969, -43.385948181152344, 84.65323638916016, 108.28060150146484, 59.904075622558594, 77.66413116455078, 4.145160675048828, 7.43194580078125, 16.48365020751953, 113.00880432128906, 102.69950866699219, -17.964630126953125, 17.30657196044922, 15.147178649902344, 106.56238555908203, -15.495742797851562, 0.747772216796875, 6.238433837890625, 28.087127685546875, -52.59522247314453], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000321.npy"}
{"epoch": 0.4852607709750567, "step": 322, "batch_size": 64, "mean": 33.61131286621094, "std": 40.812808990478516, "min": -103.16629791259766, "p10": -6.964430236816406, "median": 26.13081169128418, "p90": 86.7677604675293, "max": 114.8828353881836, "pos_frac": 0.84375, "sample": [-7.2626953125, 69.3783950805664, 84.20952606201172, -6.2684783935546875, 23.518272399902344, 6.673730850219727, 70.92137908935547, 41.195831298828125, 85.05644226074219, 25.08742904663086, -7.572837829589844, 27.1741943359375, 74.14875030517578, 98.27838134765625, 8.922576904296875, 14.819427490234375, 14.24822998046875, 70.18870544433594, 2.1793670654296875, -0.6588134765625, 16.021068572998047, 21.043075561523438, -26.16455078125, 44.94232177734375, 54.75672912597656, 14.251277923583984, 2.8819141387939453, 30.94696807861328, 1.4010562896728516, -31.63824462890625, -2.5726470947265625, 44.41902160644531, 58.89448547363281, 60.186492919921875, 112.57472229003906, 6.954641342163086, 9.105964660644531, 114.8828353881836, 4.53155517578125, 30.768165588378906, 63.03852844238281, 79.33763885498047, 2.996458053588867, 87.50118255615234, 2.1914291381835938, 102.37519836425781, 79.6293716430664, 55.32304382324219, -29.26436996459961, -103.16629791259766, 39.759498596191406, 22.814559936523438, 31.217403411865234, 22.889545440673828, 99.4775390625, 55.743133544921875, 62.97539520263672, 11.011932373046875, 47.682403564453125, -18.981552124023438, 105.30545043945312, 9.033248901367188, 22.282394409179688, 37.5262565612793], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000322.npy"}
{"epoch": 0.48677248677248675, "step": 323, "batch_size": 64, "mean": 23.097457885742188, "std": 54.450889587402344, "min": -104.39694213867188, "p10": -38.66170806884765, "median": 15.218879699707031, "p90": 108.86482696533203, "max": 128.73538208007812, "pos_frac": 0.671875, "sample": [33.257102966308594, -37.69879150390625, 47.07325744628906, 6.73974609375, 35.102294921875, 84.68738555908203, 7.7752532958984375, 36.8586540222168, 24.4415283203125, 107.87117004394531, 13.813400268554688, -17.91222381591797, 96.01194763183594, -0.14250946044921875, 128.37062072753906, -37.36575698852539, -81.49359893798828, 16.627166748046875, 70.2728042602539, -34.69258117675781, 30.4501953125, 31.16437530517578, 59.64885711669922, 10.70068359375, 116.06610107421875, -31.658723831176758, -39.07438659667969, 43.48801803588867, 8.288211822509766, -42.123443603515625, 127.05238342285156, 19.465805053710938, -21.083263397216797, -44.301780700683594, 112.26212310791016, 2.7847900390625, -62.55743408203125, 70.76676177978516, 61.32133483886719, -0.4522819519042969, 15.1942138671875, 11.26043701171875, -5.059989929199219, 0.7522201538085938, 112.32849884033203, 128.73538208007812, 76.71586608886719, -94.55485534667969, 106.49966430664062, 21.766700744628906, -104.39694213867188, -15.152030944824219, 31.524738311767578, 9.931983947753906, -0.450714111328125, 71.02124786376953, 109.29067993164062, 14.633255004882812, 17.3310546875, -1.6342105865478516, -4.166404724121094, -12.824756622314453, 22.442550659179688, 15.243545532226562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000323.npy"}
{"epoch": 0.48828420256991684, "step": 324, "batch_size": 64, "mean": 27.562347412109375, "std": 46.602264404296875, "min": -96.79757690429688, "p10": -17.90002899169922, "median": 23.942258834838867, "p90": 90.31042861938481, "max": 123.72372436523438, "pos_frac": 0.703125, "sample": [100.6954574584961, 72.64244842529297, 10.160835266113281, -3.36517333984375, -85.88111877441406, 80.2586898803711, 15.035388946533203, 78.96258544921875, -17.23497772216797, -22.596969604492188, 41.94330596923828, 102.34626770019531, 20.27729034423828, 9.860488891601562, 59.3981819152832, 11.160285949707031, 64.29081726074219, -58.08720016479492, 13.954376220703125, -8.32171630859375, 49.144859313964844, -96.79757690429688, 2.9250106811523438, 105.21495056152344, 7.645587921142578, -5.838190078735352, -30.47270965576172, 44.9345588684082, 94.61831665039062, 17.41321563720703, 44.82070541381836, -75.82952880859375, 34.148155212402344, 70.50139617919922, -1.4860916137695312, 72.59730529785156, 59.27855682373047, 16.512523651123047, 47.33057403564453, 51.447731018066406, 34.541725158691406, 12.570465087890625, 107.45390319824219, -5.488346099853516, 123.72372436523438, 6.232763290405273, 18.570205688476562, 47.77093505859375, 52.8406982421875, 74.34370422363281, 112.6928939819336, -4.139137268066406, 29.691242218017578, -2.227367401123047, 78.35232543945312, -0.538909912109375, 32.31597137451172, -16.53105926513672, -8.458663940429688, 27.607227325439453, 39.646484375, -18.18505096435547, 28.471702575683594, -0.8757801055908203], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000324.npy"}
{"epoch": 0.4897959183673469, "step": 325, "batch_size": 64, "mean": 28.89820098876953, "std": 47.91515350341797, "min": -101.43913269042969, "p10": -28.650778198242186, "median": 25.958471298217773, "p90": 89.92961502075195, "max": 124.1451187133789, "pos_frac": 0.734375, "sample": [39.0517578125, 124.1451187133789, 18.702720642089844, 101.55142211914062, -21.788822174072266, 30.104049682617188, 83.00003051757812, -5.524848937988281, 24.4761962890625, 49.645511627197266, 47.373741149902344, 81.66637420654297, -28.7296142578125, -41.08613204956055, 51.92578125, 27.440746307373047, 73.82218933105469, 34.98133087158203, 46.806861877441406, -33.26070022583008, 122.19424438476562, -41.422210693359375, 89.94285583496094, 16.064495086669922, -7.402509689331055, 120.62914276123047, 9.138740539550781, 54.82940673828125, -29.364959716796875, 35.91191101074219, -67.76318359375, 2.8582611083984375, 62.521949768066406, 15.76220703125, -1.2900390625, 87.02677917480469, 9.351886749267578, 89.89871978759766, 41.51845169067383, 18.79302978515625, 57.06703186035156, 12.058677673339844, -1.5215301513671875, 119.02227783203125, -101.43913269042969, 6.7009735107421875, 9.969581604003906, 60.92320251464844, 86.35185241699219, 72.36322784423828, 112.98345947265625, -21.976207733154297, -26.049407958984375, -1.6985397338867188, 7.454338073730469, 0.36281585693359375, 29.427127838134766, -28.466827392578125, 38.28712463378906, 43.944175720214844, 7.5215301513671875, 37.85353088378906, 6.860107421875, -12.017417907714844], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000325.npy"}
{"epoch": 0.491307634164777, "step": 326, "batch_size": 64, "mean": 19.33672332763672, "std": 45.39617919921875, "min": -82.67153930664062, "p10": -40.367236328124996, "median": 12.307668685913086, "p90": 77.84005966186524, "max": 116.94453430175781, "pos_frac": 0.65625, "sample": [-10.992729187011719, -7.662086486816406, 14.263477325439453, 19.159400939941406, 39.415618896484375, -29.278915405273438, 116.94453430175781, 9.001983642578125, 23.2894287109375, 26.294921875, -3.0064620971679688, 77.96353912353516, 37.10454559326172, 2.5140228271484375, 1.2038040161132812, 99.6251449584961, 8.261308670043945, 34.02711486816406, -4.533016204833984, 53.050506591796875, 90.65005493164062, -5.206573486328125, 4.358489990234375, 49.17284393310547, -45.87078094482422, 23.784759521484375, -77.97784423828125, -5.065887451171875, -27.212310791015625, 10.321029663085938, 106.73031616210938, -17.758621215820312, 44.43752670288086, -1.4306259155273438, 8.874923706054688, 16.99878692626953, 7.6208038330078125, -41.16194152832031, -38.51292419433594, 36.26995849609375, 29.08660888671875, 72.46055603027344, -46.739715576171875, -3.2878036499023438, 74.79146575927734, 34.05632019042969, 57.90830993652344, -58.67766571044922, 72.88172149658203, 111.36395263671875, 4.75640869140625, -58.94830322265625, -4.0922088623046875, 85.00342559814453, -12.053680419921875, 77.55194091796875, -82.67153930664062, 33.22294616699219, 65.61573791503906, 27.603851318359375, 65.08169555664062, -19.31764030456543, 10.351860046386719, 55.933937072753906], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000326.npy"}
{"epoch": 0.4928193499622071, "step": 327, "batch_size": 64, "mean": 20.763713836669922, "std": 52.06520080566406, "min": -115.74806213378906, "p10": -37.16413497924805, "median": 17.042034149169922, "p90": 93.28424148559571, "max": 132.9388427734375, "pos_frac": 0.625, "sample": [16.97760772705078, -37.425323486328125, -11.394874572753906, 81.34321594238281, -12.958908081054688, 5.071319580078125, -37.35430145263672, 28.2593994140625, 6.356813430786133, 76.93196105957031, 93.76693725585938, -24.56292724609375, -90.65618896484375, 92.15795135498047, 63.677268981933594, 81.43110656738281, 87.79277038574219, -8.32870864868164, -14.037872314453125, 18.273895263671875, -2.6595420837402344, -1.9921150207519531, 27.981075286865234, 106.54249572753906, 111.69650268554688, 42.49949645996094, -1.2779541015625, 26.282882690429688, 32.832576751708984, 9.368034362792969, 17.281679153442383, 7.687221527099609, -76.197265625, 58.777793884277344, -11.195762634277344, -0.6622238159179688, 23.980026245117188, 17.106460571289062, -22.454158782958984, -13.8636474609375, 117.439208984375, 22.072111129760742, -115.74806213378906, 4.94403076171875, 79.07418823242188, 132.9388427734375, -36.72041320800781, 97.55683135986328, 123.5875473022461, 61.471893310546875, 58.56082534790039, 34.629188537597656, 8.972896575927734, 22.135665893554688, -49.5095329284668, -0.5312271118164062, 26.883087158203125, 37.657997131347656, -4.1030731201171875, 10.882606506347656, -22.85113525390625, -62.09342956542969, 50.005889892578125, -35.43287658691406], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000327.npy"}
{"epoch": 0.4943310657596372, "step": 328, "batch_size": 64, "mean": 25.054725646972656, "std": 46.785343170166016, "min": -113.35356140136719, "p10": -19.158366394042968, "median": 10.929274559020996, "p90": 93.44778976440429, "max": 125.00071716308594, "pos_frac": 0.671875, "sample": [80.35169982910156, -0.5065841674804688, 26.35488510131836, 40.46630859375, 64.53703308105469, 9.373046875, 4.140193939208984, 59.49475860595703, 90.81022644042969, 3.5978622436523438, 10.129837036132812, 61.618743896484375, 101.00825500488281, -1.3082046508789062, 27.45074462890625, -12.524200439453125, 94.83270263671875, 31.86644172668457, 8.4757080078125, 13.31854248046875, 9.709487915039062, -8.5745849609375, -4.802604675292969, 102.53002166748047, -2.0776519775390625, -59.31791687011719, 125.00071716308594, 2.7696151733398438, -59.13836669921875, 6.013481140136719, -113.35356140136719, 68.35831451416016, 33.52210998535156, 1.748046875, 109.59811401367188, -12.064102172851562, -19.990585327148438, 11.72871208190918, 93.39448547363281, -3.3229217529296875, 51.87117004394531, 21.444320678710938, -30.483509063720703, -49.0782470703125, -1.9380950927734375, 62.80818176269531, -17.216522216796875, -4.9879150390625, 4.3483734130859375, -1.4786720275878906, 33.78258514404297, -12.590957641601562, 66.63273620605469, 32.55035400390625, -1.9177360534667969, 24.8770694732666, 64.05918884277344, 27.842426300048828, 93.47063446044922, 4.867458343505859, 83.10997772216797, -23.966064453125, 113.50114440917969, 66.77571105957031], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000328.npy"}
{"epoch": 0.4958427815570673, "step": 329, "batch_size": 64, "mean": 30.581024169921875, "std": 46.23456954956055, "min": -52.010711669921875, "p10": -22.537074279785156, "median": 20.177570343017578, "p90": 105.3409637451172, "max": 125.8364486694336, "pos_frac": 0.6875, "sample": [49.80146026611328, 69.96652221679688, 31.721603393554688, 34.76698303222656, -31.29442596435547, 22.019752502441406, 124.54660034179688, 0.36562347412109375, 46.63458251953125, -3.4785003662109375, 6.52752685546875, 15.792366027832031, -11.961885452270508, 73.86288452148438, -20.59667205810547, -34.6300048828125, -5.429939270019531, 103.69241333007812, -5.3408660888671875, 36.612457275390625, -1.0774002075195312, 7.574642181396484, 115.42432403564453, 37.94996643066406, -23.368675231933594, 120.32908630371094, 125.8364486694336, 57.789085388183594, 41.021583557128906, 39.378623962402344, 86.21395874023438, 24.817237854003906, -39.53349304199219, 23.363189697265625, 18.33538818359375, -1.8325347900390625, 103.68803405761719, -1.7732887268066406, 106.0474853515625, 69.41988372802734, 117.31526184082031, 60.965972900390625, 12.717681884765625, -33.386070251464844, -3.4163131713867188, 66.56368255615234, 1.6835289001464844, 12.292301177978516, 38.69618225097656, 63.45151138305664, 10.615550994873047, 66.39949035644531, 2.2536544799804688, -30.224166870117188, 119.59210205078125, 25.649932861328125, -12.810935974121094, 78.51752471923828, -6.254848480224609, -52.010711669921875, 5.3614044189453125, -6.0540618896484375, 13.958492279052734, -7.8736114501953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000329.npy"}
{"epoch": 0.4973544973544973, "step": 330, "batch_size": 64, "mean": 36.05565643310547, "std": 54.9324951171875, "min": -110.32714080810547, "p10": -36.65367050170898, "median": 35.45375061035156, "p90": 104.37629241943361, "max": 131.09779357910156, "pos_frac": 0.8125, "sample": [-33.205970764160156, -42.58973693847656, 0.8856925964355469, -73.50891876220703, 42.41455841064453, -110.32714080810547, 43.188377380371094, 15.644584655761719, 5.121002197265625, 68.42142486572266, 67.1442642211914, 18.168949127197266, -38.131256103515625, 2.5478286743164062, 1.3026123046875, 114.68775939941406, 100.70744323730469, -25.749114990234375, 131.09779357910156, 43.43901062011719, 68.60169982910156, -81.17716217041016, 73.29379272460938, -43.976348876953125, 101.05182647705078, 83.9290771484375, 52.36670684814453, 7.766506195068359, -32.68041229248047, 49.43718719482422, 31.18515396118164, 105.80106353759766, 9.626426696777344, 87.77008056640625, -31.79676055908203, 67.69422912597656, 65.5001220703125, 116.87115478515625, 111.77206420898438, 1.4572887420654297, 10.092681884765625, 95.58761596679688, 40.932167053222656, 96.4449462890625, 85.11809539794922, 90.86457824707031, 5.210704803466797, 31.86541748046875, 10.850692749023438, 4.7227325439453125, 13.996246337890625, 39.042083740234375, 100.40603637695312, 67.17784881591797, 1.19598388671875, 23.15618896484375, 66.979736328125, -0.9139328002929688, 126.9166259765625, -43.21348571777344, 30.328197479248047, 81.50546264648438, 31.240283966064453, 122.30211639404297], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000330.npy"}
{"epoch": 0.4988662131519274, "step": 331, "batch_size": 64, "mean": 36.94060516357422, "std": 53.52018356323242, "min": -129.0894775390625, "p10": -4.280590820312499, "median": 17.36176300048828, "p90": 115.77055206298829, "max": 132.3403778076172, "pos_frac": 0.78125, "sample": [127.44312286376953, 6.656089782714844, 15.345863342285156, 20.68092155456543, 60.99604797363281, 104.2879867553711, 48.071380615234375, 32.24908447265625, 21.141036987304688, 8.4068603515625, 6.460172653198242, 130.31907653808594, 91.66500854492188, 122.23347473144531, 43.32154846191406, 1.922027587890625, 132.3403778076172, 18.479049682617188, 59.44617462158203, 120.16281127929688, 67.36856079101562, 14.000946044921875, 0.25888824462890625, -0.30340576171875, 14.219551086425781, 120.46762084960938, 113.5770492553711, 109.94694519042969, -4.6964874267578125, -2.1582183837890625, 77.33003234863281, 11.745956420898438, 104.64397430419922, 81.00762939453125, -1.2737503051757812, -0.943939208984375, 16.244476318359375, -71.18157196044922, 1.6221485137939453, 102.70964813232422, -20.389911651611328, 15.817581176757812, -0.1429443359375, -0.4721221923828125, 58.38017272949219, -129.0894775390625, 58.612586975097656, -3.3101654052734375, -14.91915512084961, 2.5184326171875, 1.39630126953125, 116.71062469482422, 12.33758544921875, 3.5907669067382812, 70.39717864990234, 0.38543701171875, -15.853597640991211, 73.07492065429688, 3.7243728637695312, -60.12751770019531, 101.95884704589844, 45.56475067138672, 32.5009765625, 85.31897735595703], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000331.npy"}
{"epoch": 0.5003779289493575, "step": 332, "batch_size": 64, "mean": 35.20355224609375, "std": 47.12651062011719, "min": -49.75714111328125, "p10": -11.50102367401123, "median": 25.939918518066406, "p90": 111.62744674682618, "max": 145.95289611816406, "pos_frac": 0.796875, "sample": [-43.4936408996582, 53.134735107421875, 0.90228271484375, -6.0086212158203125, 26.142868041992188, 8.948410034179688, 30.8238525390625, -10.450149536132812, 24.27691650390625, -0.231964111328125, 2.5312881469726562, 129.06771850585938, 77.93547058105469, 38.354576110839844, 31.39264678955078, 16.08087921142578, -11.612947463989258, 3.8665504455566406, 56.61541748046875, 39.03166198730469, 25.736968994140625, 12.099227905273438, -49.75714111328125, 122.93175506591797, 77.86360168457031, 63.93163299560547, 5.086887359619141, 10.995124816894531, 12.335784912109375, 3.6718215942382812, 55.02378845214844, 127.20972442626953, -11.2398681640625, 14.148681640625, -41.59861755371094, 41.47612380981445, 12.399526596069336, 108.31639099121094, -35.19991683959961, 28.849796295166016, 71.652587890625, 67.74836730957031, 70.41317749023438, 41.26931381225586, 14.347846984863281, -14.190147399902344, 113.02555084228516, 54.85758590698242, 66.28816223144531, 23.922012329101562, 145.95289611816406, 108.36520385742188, 39.337913513183594, -10.411079406738281, -7.300750732421875, 2.652496337890625, 126.07234191894531, 43.63594436645508, 14.641357421875, 93.27606201171875, 122.6849594116211, 49.889434814453125, -39.804229736328125, 3.111114501953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000332.npy"}
{"epoch": 0.5018896447467877, "step": 333, "batch_size": 64, "mean": 38.178375244140625, "std": 44.858299255371094, "min": -50.887115478515625, "p10": -12.598713493347168, "median": 25.196500778198242, "p90": 102.87680816650392, "max": 133.50682067871094, "pos_frac": 0.796875, "sample": [77.82361602783203, -12.489032745361328, 15.868606567382812, 59.465118408203125, 7.508575439453125, -21.524738311767578, 60.08635711669922, 56.43603515625, 15.983779907226562, 1.482666015625, 44.0484619140625, 8.921821594238281, 21.70867156982422, 71.9812240600586, 75.06062316894531, 19.102279663085938, 64.33514404296875, 132.64016723632812, 100.99259948730469, 66.21041107177734, 10.183456420898438, 18.71692657470703, 3.1117172241210938, 19.34722137451172, -12.645719528198242, -0.04062652587890625, -16.391395568847656, 46.63398742675781, 90.70308685302734, 22.87860107421875, 77.80940246582031, 122.93722534179688, -20.021827697753906, -23.229293823242188, 48.77794647216797, 18.647483825683594, 133.50682067871094, -2.9171371459960938, 12.14682388305664, -5.04833984375, -1.9533576965332031, -8.822799682617188, -50.887115478515625, 29.360794067382812, 35.974517822265625, 8.909812927246094, 83.44906616210938, 60.17797088623047, 27.514400482177734, 10.034751892089844, 18.24042510986328, 4.79345703125, 52.63018798828125, 129.84185791015625, 50.091800689697266, -37.42535400390625, 78.01416015625, 0.4893341064453125, 105.06752014160156, 39.997337341308594, 124.00360107421875, 81.65771484375, 87.82288360595703, 103.684326171875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000333.npy"}
{"epoch": 0.5034013605442177, "step": 334, "batch_size": 64, "mean": 25.9740047454834, "std": 59.574127197265625, "min": -123.23603057861328, "p10": -50.63372802734375, "median": 18.816144943237305, "p90": 104.01175537109377, "max": 138.8667449951172, "pos_frac": 0.671875, "sample": [-42.811378479003906, -123.23603057861328, -6.97076416015625, 75.45667266845703, 29.209335327148438, -3.9192733764648438, -51.69010925292969, 70.66361999511719, 20.282989501953125, -113.9766616821289, 2.7178268432617188, 138.8667449951172, -69.9097900390625, 8.087604522705078, 3.978870391845703, 109.02233123779297, 35.274871826171875, -5.752227783203125, -6.603363037109375, 63.9451904296875, 74.94810485839844, 130.40589904785156, -33.685306549072266, 66.09680938720703, 89.30194091796875, -85.56172180175781, 37.74041748046875, 1.164773941040039, 86.32173156738281, 2.0075912475585938, -2.829303741455078, 34.213104248046875, 112.5222396850586, -3.35980224609375, 88.32710266113281, 46.47160339355469, 98.3619384765625, -1.2961273193359375, 9.882965087890625, 0.1065826416015625, 11.289867401123047, -55.428287506103516, -32.54950714111328, -57.392677307128906, 97.36672973632812, -1.9314231872558594, 18.678329467773438, 30.070608139038086, -7.0458221435546875, 49.34291076660156, 65.37705993652344, 28.349197387695312, -48.16883850097656, 44.08371353149414, -26.077913284301758, 94.1260986328125, 3.9126815795898438, 134.67800903320312, 18.209030151367188, 18.953960418701172, 106.43310546875, 120.97244262695312, 77.74156188964844, 87.5685043334961], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000334.npy"}
{"epoch": 0.5049130763416477, "step": 335, "batch_size": 64, "mean": 25.188629150390625, "std": 64.10665130615234, "min": -118.56272888183594, "p10": -45.05284805297852, "median": 9.137935638427734, "p90": 125.32320098876953, "max": 142.232421875, "pos_frac": 0.609375, "sample": [51.22630310058594, 66.97926330566406, -50.752708435058594, 142.232421875, -4.5347137451171875, -17.88116455078125, 44.628822326660156, 0.026006698608398438, 132.52737426757812, 103.53094482421875, 135.22467041015625, 125.06702423095703, 8.9415283203125, -0.420013427734375, 119.23309326171875, 110.26032257080078, 15.155097961425781, 55.84989547729492, -28.13397216796875, 129.09671020507812, -29.371021270751953, -0.19062042236328125, 36.62760925292969, 125.43299102783203, -8.192176818847656, 7.775787353515625, 19.0982666015625, -65.88619995117188, 1.6604747772216797, 64.58482360839844, 129.36099243164062, -2.02691650390625, 14.897491455078125, 9.334342956542969, 63.618289947509766, -96.86644744873047, 129.29690551757812, -19.243858337402344, -2.912750244140625, 118.13187408447266, -15.306083679199219, -3.0395050048828125, -31.30677032470703, 13.496376037597656, 8.6263427734375, 118.71761322021484, -3.6549911499023438, -76.61932373046875, 21.012908935546875, -25.839275360107422, 39.41075134277344, -44.683509826660156, 25.05419158935547, -118.56272888183594, -113.63809204101562, -2.3537445068359375, 84.650146484375, 55.31059265136719, 6.07379150390625, 77.25943756103516, -22.557174682617188, 1.4338226318359375, -45.21113586425781, 30.411911010742188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000335.npy"}
{"epoch": 0.5064247921390779, "step": 336, "batch_size": 64, "mean": 28.666759490966797, "std": 49.9303092956543, "min": -121.03214263916016, "p10": -13.57588424682617, "median": 20.76153564453125, "p90": 108.43801269531252, "max": 138.78152465820312, "pos_frac": 0.75, "sample": [21.884140014648438, 16.44598388671875, 26.850616455078125, 13.479629516601562, 113.96185302734375, 110.9269790649414, 69.01963806152344, 1.8267669677734375, 121.6481704711914, 15.984134674072266, 42.35246276855469, 24.728031158447266, 127.1371841430664, -11.931617736816406, 70.1513900756836, -12.390022277832031, -13.944793701171875, 26.872146606445312, 128.80859375, -3.009563446044922, 15.18130111694336, 69.19117736816406, 56.57164001464844, 45.440208435058594, -24.181381225585938, 66.06697082519531, 51.47010040283203, -121.03214263916016, 12.907478332519531, -97.9931640625, 102.63042449951172, -30.221115112304688, 6.6410980224609375, 4.840789794921875, -9.814994812011719, -1.5209026336669922, 9.508499145507812, -12.715095520019531, 122.87738037109375, 21.984458923339844, 65.02020263671875, 0.11395263671875, 32.50154113769531, 52.55081558227539, 32.76427459716797, 6.09820556640625, 86.04472351074219, -2.3651885986328125, 15.884834289550781, 49.48042297363281, -6.450582504272461, -48.57987976074219, 28.40619659423828, 47.21022033691406, 138.78152465820312, 48.076385498046875, 0.701812744140625, 25.798538208007812, 3.3387908935546875, -18.334762573242188, 4.5715179443359375, 81.22112274169922, 19.638931274414062, -6.435272216796875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000336.npy"}
{"epoch": 0.5079365079365079, "step": 337, "batch_size": 64, "mean": 25.066566467285156, "std": 54.04571533203125, "min": -106.23277282714844, "p10": -28.879697418212885, "median": 12.848283767700195, "p90": 102.23646011352542, "max": 145.57115173339844, "pos_frac": 0.65625, "sample": [16.076263427734375, -46.28431701660156, -65.36116027832031, -17.984642028808594, 5.174236297607422, 109.40763854980469, -30.935997009277344, 105.22608947753906, 29.786605834960938, -9.922134399414062, 26.478530883789062, -14.121856689453125, 93.8131332397461, -106.23277282714844, -24.0816650390625, -6.0325927734375, 9.055797576904297, 47.864776611328125, 0.6791763305664062, -7.433258056640625, -1.7332839965820312, -7.070331573486328, -23.767759323120117, -22.601272583007812, 84.65202331542969, 63.13866424560547, -47.285614013671875, 94.57901763916016, 12.18802261352539, 132.23338317871094, 105.79434204101562, -81.02469635009766, 79.9699478149414, -19.820037841796875, 55.9671630859375, 39.36866760253906, 128.52059936523438, 17.415679931640625, 10.232059478759766, 55.35722351074219, 23.735328674316406, 5.902935028076172, 50.46195983886719, 66.60679626464844, 13.508544921875, 0.6010513305664062, 72.16018676757812, 3.6388397216796875, 4.8317413330078125, 85.12921905517578, 34.37251281738281, 145.57115173339844, -1.8130264282226562, -0.0029296875, -2.4333457946777344, 64.13296508789062, 3.7916603088378906, -14.886734008789062, 76.3479995727539, 95.26065826416016, -72.26431274414062, 24.212509155273438, 112.38218688964844, 21.726699829101562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000337.npy"}
{"epoch": 0.509448223733938, "step": 338, "batch_size": 64, "mean": 24.654888153076172, "std": 51.632423400878906, "min": -77.33429718017578, "p10": -30.21963424682617, "median": 7.83917236328125, "p90": 115.88485412597656, "max": 137.45651245117188, "pos_frac": 0.65625, "sample": [-11.585723876953125, 28.09600830078125, 116.33119201660156, 3.6133346557617188, 72.21641540527344, 97.33599853515625, -33.18248748779297, 98.55174255371094, 7.389801025390625, 122.32981872558594, 122.74128723144531, 3.8811721801757812, -10.650352478027344, -31.366043090820312, 5.13031005859375, -20.677406311035156, -8.78143310546875, -57.05763244628906, -39.27496337890625, -0.06818389892578125, -2.131744384765625, 114.84339904785156, 16.03577423095703, -15.971084594726562, 51.64801788330078, -28.206146240234375, -77.33429718017578, 19.136024475097656, 24.8310546875, 45.1502685546875, 3.9399871826171875, 11.851348876953125, 5.085807800292969, 9.87103271484375, 10.743690490722656, 90.0654067993164, 21.856063842773438, -0.314727783203125, 117.42445373535156, 2.328826904296875, -2.7183380126953125, 66.74208068847656, -18.78464126586914, 51.18079376220703, 6.412269592285156, 5.420402526855469, -0.9636306762695312, -13.400867462158203, 18.725990295410156, -31.082557678222656, -57.392333984375, 119.35031127929688, 14.34625244140625, -9.503097534179688, 36.117469787597656, 72.41439819335938, 8.288543701171875, 133.5025177001953, 111.96647644042969, -9.79486083984375, 5.802471160888672, 17.489490509033203, 137.45651245117188, 30.511211395263672], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000338.npy"}
{"epoch": 0.5109599395313681, "step": 339, "batch_size": 64, "mean": 38.45646667480469, "std": 50.42205047607422, "min": -78.60813903808594, "p10": -20.003403091430663, "median": 36.905052185058594, "p90": 107.43688354492187, "max": 137.8557586669922, "pos_frac": 0.78125, "sample": [8.779586791992188, 17.749549865722656, -19.565387725830078, -8.012847900390625, -14.083549499511719, 106.66522979736328, 53.514892578125, 35.863853454589844, 15.113658905029297, 6.588470458984375, 54.120697021484375, 65.87640380859375, 11.38677978515625, 8.08150863647461, 70.70169830322266, 94.65708923339844, -23.587242126464844, 112.26631164550781, 97.7366943359375, 40.01600646972656, 55.867496490478516, 46.61286163330078, 81.8338623046875, 80.89457702636719, 5.172767639160156, 109.17228698730469, 7.1651611328125, 46.53211975097656, -13.751571655273438, 37.946250915527344, 72.43647003173828, 73.42308044433594, 97.70616912841797, 134.2293243408203, 109.09788513183594, 66.28565979003906, 101.42516326904297, -13.479194641113281, 42.89923095703125, 25.625579833984375, 44.792274475097656, 5.67486572265625, -20.191123962402344, 0.9157772064208984, 4.423126220703125, -26.059829711914062, 7.774322509765625, -35.20085144042969, -78.60813903808594, -50.45082092285156, -4.15794563293457, 20.079402923583984, 8.624031066894531, 81.94851684570312, -35.55110168457031, 137.8557586669922, -13.671463012695312, 44.7524528503418, 14.214820861816406, 107.53053283691406, 107.21836853027344, 105.91661071777344, 7.408945083618164, 125.0106201171875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000339.npy"}
{"epoch": 0.5124716553287982, "step": 340, "batch_size": 64, "mean": 32.514930725097656, "std": 60.792144775390625, "min": -112.54609680175781, "p10": -29.390523529052732, "median": 15.773733139038086, "p90": 122.46867065429689, "max": 133.66616821289062, "pos_frac": 0.671875, "sample": [8.533409118652344, 120.36483764648438, 3.798093795776367, 5.3982391357421875, 3.8762054443359375, 133.66616821289062, -5.763639450073242, -0.12509918212890625, -1.3660907745361328, 64.49729919433594, 2.4717864990234375, 53.30314636230469, 16.923080444335938, 81.88276672363281, 132.81744384765625, 110.4976806640625, 41.26853561401367, -45.81526184082031, 21.410995483398438, 10.124099731445312, -10.547662734985352, -2.1847476959228516, -9.130485534667969, 46.0681037902832, -1.4835433959960938, 123.08273315429688, -4.845546722412109, 108.11750030517578, 21.725921630859375, 102.7014389038086, -39.62989044189453, 119.19371032714844, -2.291473388671875, 121.03585815429688, 99.0913314819336, 62.20904541015625, 108.21309661865234, 8.091180801391602, 19.30413818359375, 129.9274139404297, 7.326549530029297, -46.45055389404297, -68.08197021484375, 133.55838012695312, 77.51292419433594, -20.772371292114258, -19.679794311523438, -30.19536590576172, 41.69718551635742, 93.24609375, 4.2115020751953125, 23.077329635620117, -27.512557983398438, -11.745529174804688, -112.54609680175781, -103.85627746582031, 131.68258666992188, 15.403697967529297, -26.81163787841797, 130.3464813232422, 72.30023956298828, 16.143768310546875, 39.02847671508789, 6.6608428955078125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000340.npy"}
{"epoch": 0.5139833711262283, "step": 341, "batch_size": 64, "mean": 37.80937194824219, "std": 60.292999267578125, "min": -108.4425048828125, "p10": -14.042908477783202, "median": 25.59035301208496, "p90": 120.52195510864259, "max": 140.55465698242188, "pos_frac": 0.75, "sample": [-100.43517303466797, 114.66548919677734, 80.73919677734375, 89.83695220947266, 140.55465698242188, 109.04583740234375, 77.50228881835938, 131.7613983154297, 79.80467224121094, 0.743316650390625, 108.66265869140625, 31.610029220581055, -60.74748992919922, -14.77105712890625, 50.188262939453125, 4.107290267944336, 117.44861602783203, -1.1003646850585938, 44.72038269042969, 119.03274536132812, 131.87310791015625, 139.5257110595703, 128.7976837158203, 55.03797149658203, -6.6695556640625, 120.67416381835938, 20.5648193359375, -49.52302551269531, -11.465511322021484, -10.384674072265625, 4.681186676025391, 42.77146911621094, -3.6198348999023438, -9.259899139404297, 4.2305145263671875, 57.13067626953125, 58.68048095703125, 61.04420471191406, 5.7949066162109375, -108.4425048828125, -11.996047973632812, -44.32879638671875, 0.434173583984375, 77.59219360351562, 9.596101760864258, 129.75286865234375, 10.246658325195312, 9.67431640625, 80.75773620605469, 91.6497573852539, 14.589767456054688, 22.826133728027344, 120.16680145263672, 24.682373046875, 26.498332977294922, 108.697998046875, 0.9315643310546875, 30.522533416748047, 3.2115421295166016, -12.343894958496094, -12.255233764648438, 52.94929504394531, 0.438079833984375, -69.30586242675781], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000341.npy"}
{"epoch": 0.5154950869236583, "step": 342, "batch_size": 64, "mean": 34.284732818603516, "std": 57.55671310424805, "min": -88.90591430664062, "p10": -21.13600425720215, "median": 12.987146377563477, "p90": 121.76344451904298, "max": 148.71566772460938, "pos_frac": 0.6875, "sample": [9.193588256835938, 18.20947265625, -2.9421005249023438, -6.529544830322266, 4.647970199584961, 119.30429077148438, -53.91176223754883, 112.064453125, -32.08037567138672, 73.25653839111328, -17.92208480834961, 52.531768798828125, 2.352010726928711, 105.1551513671875, 11.65887451171875, -26.675918579101562, -4.64283561706543, -63.29780197143555, -4.139963150024414, -13.735580444335938, 19.46747589111328, 68.58268737792969, 40.75718688964844, 122.81736755371094, -20.909748077392578, 11.223381042480469, 13.408004760742188, 79.58929443359375, 75.04171752929688, 41.84864807128906, -8.874885559082031, 9.180255889892578, 23.36767578125, 71.07164001464844, 10.546051025390625, -21.23297119140625, 111.23385620117188, 3.233489990234375, 3.89617919921875, 105.72883605957031, 8.27020263671875, 132.74237060546875, 75.93807983398438, 12.566287994384766, -88.90591430664062, 136.77597045898438, 36.8646125793457, -1.5086822509765625, 148.71566772460938, 139.4709014892578, 21.05689239501953, -69.29158782958984, -4.3015594482421875, 141.7428741455078, -15.892616271972656, 86.45274353027344, 71.77367401123047, 0.8556137084960938, -6.07000732421875, 13.626789093017578, 106.08848571777344, 78.76929473876953, -1.1844635009765625, 127.19503784179688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000342.npy"}
{"epoch": 0.5170068027210885, "step": 343, "batch_size": 64, "mean": 21.631834030151367, "std": 55.56227111816406, "min": -126.7045669555664, "p10": -36.127074813842775, "median": 10.317619323730469, "p90": 108.90605239868168, "max": 134.00296020507812, "pos_frac": 0.65625, "sample": [-0.26511383056640625, 65.58075714111328, -12.914939880371094, -52.052581787109375, 12.414207458496094, 91.00923156738281, 5.143623352050781, -5.150279998779297, -28.97515106201172, -8.464500427246094, -126.7045669555664, -15.842254638671875, 27.923629760742188, 56.153594970703125, 134.00296020507812, -9.145790100097656, 96.13441467285156, 23.454116821289062, 33.7647590637207, 47.06227493286133, 31.190185546875, 7.715400695800781, 121.3527603149414, 35.55122375488281, 5.69683837890625, 70.80713653564453, 36.87812042236328, -14.75180435180664, 24.428970336914062, 61.57587814331055, 8.00927734375, 25.997528076171875, 84.12619018554688, 16.616836547851562, 8.221031188964844, -1.5436820983886719, 49.002479553222656, 120.04749298095703, -93.50878143310547, 116.30470275878906, -80.90814208984375, -6.340385437011719, 4.4684600830078125, 129.58340454101562, 99.25299835205078, -4.374702453613281, 27.14459228515625, 1.32513427734375, -36.035465240478516, 113.04307556152344, 123.44171905517578, 32.945770263671875, -36.16633605957031, 27.004928588867188, -0.10010528564453125, 92.56625366210938, -44.02687454223633, 5.155555725097656, 7.464599609375, 1.725006103515625, -67.08038330078125, -33.556983947753906, -35.92239761352539, 16.9814453125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000343.npy"}
{"epoch": 0.5185185185185185, "step": 344, "batch_size": 64, "mean": 33.79200744628906, "std": 55.6348876953125, "min": -109.86031341552734, "p10": -27.79608688354491, "median": 30.047847747802734, "p90": 110.0718849182129, "max": 157.8804931640625, "pos_frac": 0.703125, "sample": [-69.26277923583984, 93.50341033935547, 28.22545623779297, 59.185585021972656, 104.62403869628906, 1.4770050048828125, 118.10868835449219, 18.468788146972656, 3.3945236206054688, 31.8702392578125, -2.86517333984375, 4.589012145996094, 71.49819946289062, 42.02459716796875, 14.377090454101562, 90.57865905761719, -8.964103698730469, 157.8804931640625, 1.4345245361328125, 73.71890258789062, -74.95335388183594, 2.164764404296875, -32.75316619873047, 139.81597900390625, 53.88361358642578, -2.8480758666992188, -41.1044921875, -0.24549102783203125, 128.25926208496094, 31.910804748535156, -16.229568481445312, 111.18721008300781, -0.7667236328125, 47.34307861328125, 61.29154968261719, 107.4694595336914, 89.36129760742188, 32.11422348022461, 11.341144561767578, -64.48228454589844, -1.2505035400390625, 10.180763244628906, -41.3216552734375, -0.19115447998046875, -3.9758758544921875, 2.6381263732910156, 38.27655029296875, 38.050994873046875, 81.81259155273438, -0.5360946655273438, 105.1021957397461, 44.74951171875, 118.08497619628906, 65.95011901855469, 91.36354064941406, 20.01434326171875, -10.15692138671875, 60.277870178222656, -2.0981903076171875, 130.6647491455078, -109.86031341552734, 61.52240753173828, 33.9468994140625, 12.817413330078125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000344.npy"}
{"epoch": 0.5200302343159486, "step": 345, "batch_size": 64, "mean": 29.83697509765625, "std": 58.805747985839844, "min": -121.694091796875, "p10": -35.41397476196289, "median": 24.004108428955078, "p90": 120.8025993347168, "max": 150.78529357910156, "pos_frac": 0.6875, "sample": [-11.525848388671875, 4.3121337890625, 129.92160034179688, -94.61102294921875, -0.722869873046875, -15.486480712890625, -9.113021850585938, -5.9777679443359375, -17.331375122070312, 35.80138397216797, 54.01953125, 136.27252197265625, 57.17867660522461, 16.976049423217773, -33.8508415222168, 30.007644653320312, -76.58171844482422, 22.92223358154297, 57.29425811767578, 57.46025848388672, 150.78529357910156, 11.857797622680664, 67.54568481445312, 51.834590911865234, 35.158607482910156, 0.6547393798828125, 68.84765625, 32.44036865234375, -14.953094482421875, 1.86285400390625, 59.79444122314453, 119.4428939819336, 29.14519500732422, 66.24040222167969, 25.085983276367188, -6.541572570800781, 88.53540802001953, 141.87420654296875, 121.38533020019531, -58.561519622802734, 101.48324584960938, -6.275611877441406, -121.694091796875, 112.92153930664062, 57.22154235839844, -51.655372619628906, 6.249076843261719, -2.6354827880859375, 39.857025146484375, 32.782310485839844, 126.70259094238281, 17.28009033203125, -53.93854522705078, 7.762237548828125, -28.822662353515625, 42.24085998535156, -0.3036518096923828, -36.08388900756836, 2.953746795654297, 4.644966125488281, 86.40202331542969, 18.540924072265625, 102.16255950927734, 122.37029266357422], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000345.npy"}
{"epoch": 0.5215419501133787, "step": 346, "batch_size": 64, "mean": 33.6463737487793, "std": 64.26007080078125, "min": -119.83827209472656, "p10": -35.910682296752924, "median": 36.87029457092285, "p90": 120.88513946533203, "max": 150.62086486816406, "pos_frac": 0.65625, "sample": [70.1690673828125, 120.43423461914062, 59.2083740234375, 98.26258087158203, 47.95414733886719, 11.8145751953125, 4.096244812011719, 36.4757194519043, 134.10916137695312, 93.33731842041016, 146.5868682861328, -2.34197998046875, 7.9091796875, 72.07914733886719, -31.424232482910156, 133.34170532226562, -63.65223693847656, 74.06442260742188, -8.097705841064453, 51.430423736572266, 50.12059783935547, 77.71226501464844, -6.9431610107421875, -6.639568328857422, 119.14037322998047, 150.62086486816406, -119.22452545166016, 47.43523025512695, 8.178543090820312, -79.01018524169922, 0.97442626953125, -0.6566162109375, 44.6183967590332, -18.227487564086914, 51.033477783203125, 10.60845947265625, 37.264869689941406, -21.98493194580078, 51.501564025878906, 149.60769653320312, 100.73757934570312, 119.25238037109375, -4.71649169921875, 65.08000183105469, 16.640663146972656, -73.34063720703125, 51.45538330078125, 13.487770080566406, -21.145404815673828, -12.831207275390625, 86.1547622680664, -4.67962646484375, -5.250799179077148, -37.83344650268555, -119.83827209472656, 31.678009033203125, 38.48184585571289, 122.59341430664062, -73.71492767333984, 121.07838439941406, -20.78204345703125, -20.794361114501953, 86.20423889160156, 93.5635757446289], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000346.npy"}
{"epoch": 0.5230536659108088, "step": 347, "batch_size": 64, "mean": 30.27042007446289, "std": 58.93278503417969, "min": -87.32212829589844, "p10": -39.69871025085449, "median": 13.377861022949219, "p90": 124.15255661010742, "max": 159.64215087890625, "pos_frac": 0.640625, "sample": [-6.453865051269531, 39.8244514465332, 107.6619644165039, -1.8399658203125, -85.80877685546875, 20.821693420410156, 33.83349609375, 24.584327697753906, 48.691646575927734, 84.41024017333984, -0.17073822021484375, 111.15321350097656, -1.9451751708984375, -19.923065185546875, -41.496238708496094, 159.64215087890625, 36.82682800292969, 52.32078552246094, -36.083099365234375, 15.488510131835938, 0.5105438232421875, 3.508281707763672, 155.252197265625, -3.5552825927734375, -51.348392486572266, -51.77085876464844, 50.780853271484375, 129.29678344726562, 126.62013244628906, -6.602760314941406, 12.1341552734375, -25.900379180908203, -21.897541046142578, 87.88008117675781, -6.368885040283203, 9.564247131347656, -87.32212829589844, 77.38945007324219, -43.80841064453125, 58.861000061035156, 45.80554962158203, -2.8322296142578125, 4.330684661865234, 84.43026733398438, 31.19893455505371, 124.12346649169922, -8.81430435180664, -39.00293731689453, 79.83173370361328, 128.3634033203125, 55.4454345703125, 70.52719116210938, 4.04595947265625, 0.649688720703125, -39.99689865112305, 124.16502380371094, 14.621566772460938, 102.182373046875, -12.654594421386719, -0.6100387573242188, 1.9241867065429688, 3.1729507446289062, 134.96914672851562, 76.66899108886719], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000347.npy"}
{"epoch": 0.5245653817082389, "step": 348, "batch_size": 64, "mean": 35.64274215698242, "std": 58.457847595214844, "min": -82.35874938964844, "p10": -21.777795791625977, "median": 17.698288917541504, "p90": 127.68589935302735, "max": 152.69003295898438, "pos_frac": 0.6875, "sample": [99.922119140625, 128.5362548828125, -15.696395874023438, 15.4425048828125, 129.93185424804688, 121.0596923828125, -33.91265869140625, 6.282360076904297, -21.336463928222656, -2.518505096435547, 148.2913818359375, -17.2127685546875, 43.95085144042969, 97.43402099609375, 68.59095001220703, 125.04196166992188, 60.103614807128906, 149.66465759277344, -14.267127990722656, 146.0181427001953, 3.6463661193847656, 76.91217041015625, 128.80276489257812, 13.401420593261719, 37.183197021484375, 122.23677825927734, 3.4850082397460938, 101.69252014160156, -27.785873413085938, -12.378032684326172, 28.539466857910156, 1.8942413330078125, -2.4804229736328125, 25.913040161132812, 18.958175659179688, -56.438819885253906, 79.8668441772461, -21.966938018798828, 14.285102844238281, 67.57676696777344, 25.109153747558594, 152.69003295898438, 45.06544494628906, -15.027667999267578, 33.40377426147461, -8.128381729125977, -47.75980758666992, -21.133281707763672, -82.35874938964844, 44.21562957763672, -17.212894439697266, 125.70173645019531, 16.43840217590332, 8.379940032958984, 84.51204681396484, -0.4498291015625, 67.0426025390625, 0.5647659301757812, 1.1637382507324219, 35.61069107055664, -41.72929382324219, 36.45591735839844, 1.9431304931640625, -2.0316810607910156], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000348.npy"}
{"epoch": 0.5260770975056689, "step": 349, "batch_size": 64, "mean": 33.6384162902832, "std": 62.4005241394043, "min": -130.9694061279297, "p10": -37.798469543457024, "median": 18.059871673583984, "p90": 128.40546112060548, "max": 150.92967224121094, "pos_frac": 0.734375, "sample": [-34.16062927246094, 6.268890380859375, 130.97207641601562, -2.0113525390625, -4.452350616455078, 0.9280014038085938, 100.3771743774414, 46.9871940612793, 32.53926086425781, 34.62028884887695, -12.770584106445312, 14.858165740966797, 122.99781799316406, 45.49762725830078, 19.68138885498047, 140.76666259765625, 112.3685302734375, -58.37004852294922, -130.9694061279297, 111.57356262207031, 1.7542877197265625, 102.5453109741211, 46.81000518798828, 8.3958740234375, 81.04792785644531, 4.670417785644531, 11.327499389648438, 75.6861572265625, -3.3865890502929688, 8.208343505859375, 60.823585510253906, 22.509113311767578, -66.0733642578125, 16.45679473876953, -0.8120651245117188, 89.28426361083984, 46.7729606628418, 144.13629150390625, -67.12932586669922, 110.8720474243164, 139.91607666015625, 6.484626770019531, -3.3596420288085938, 8.348953247070312, -5.850330352783203, 114.97496032714844, 130.7230224609375, 6.771495819091797, -28.680633544921875, -18.028310775756836, 45.84761047363281, -39.3575439453125, -55.88313293457031, 4.440093994140625, 37.97700881958008, 136.40138244628906, 2.833648681640625, 19.662948608398438, -75.51608276367188, 115.46281433105469, 9.947357177734375, 34.74903106689453, 42.46195983886719, 150.92967224121094], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000349.npy"}
{"epoch": 0.527588813303099, "step": 350, "batch_size": 64, "mean": 47.97602081298828, "std": 54.81682205200195, "min": -56.25741195678711, "p10": -11.901790618896483, "median": 36.26223373413086, "p90": 135.5051193237305, "max": 147.29690551757812, "pos_frac": 0.75, "sample": [140.84622192382812, -7.368000030517578, 78.57946014404297, -18.209835052490234, 43.43412780761719, 19.886165618896484, 113.90020751953125, 19.416290283203125, 108.33354187011719, 99.45750427246094, -23.194625854492188, -6.598907470703125, 7.904571533203125, 66.25311279296875, 72.6513900756836, 68.51564025878906, 138.53594970703125, 15.443618774414062, -8.2208251953125, -36.34209442138672, 13.37005615234375, 115.09697723388672, -19.987808227539062, 81.0875244140625, 36.50556182861328, -9.39013671875, -7.161975860595703, 38.83441162109375, 48.88359832763672, -56.25741195678711, 128.4331817626953, -9.810768127441406, 115.68630981445312, 4.128517150878906, 72.9036865234375, 142.7311553955078, 75.42784118652344, 100.5810775756836, 147.29690551757812, 140.39442443847656, 1.3498001098632812, 145.42703247070312, 95.76655578613281, 36.01890563964844, -12.797943115234375, 5.567188262939453, 17.890281677246094, -3.0566349029541016, 7.491992950439453, 28.178451538085938, 70.28456115722656, 104.69647979736328, 1.3676233291625977, 27.919979095458984, 19.80333709716797, -12.806777954101562, 46.981964111328125, 33.590110778808594, -1.1608352661132812, -3.218475341796875, 139.79591369628906, 64.09149932861328, 104.11528015136719, 101.1922607421875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000350.npy"}
{"epoch": 0.5291005291005291, "step": 351, "batch_size": 64, "mean": 33.90068817138672, "std": 64.59574890136719, "min": -142.2456817626953, "p10": -39.72111587524413, "median": 28.920940399169922, "p90": 111.94438858032228, "max": 143.3680419921875, "pos_frac": 0.765625, "sample": [0.9386978149414062, 94.4582290649414, 125.8450927734375, 73.86187744140625, 13.616119384765625, 131.85494995117188, -61.35844802856445, 10.869598388671875, 73.04344177246094, 36.21330261230469, -0.9437408447265625, 143.3680419921875, 11.516483306884766, -42.563720703125, 9.253616333007812, -85.43496704101562, 9.511680603027344, 30.477886199951172, 84.76390075683594, 9.862648010253906, 14.246467590332031, -14.184921264648438, 108.54705047607422, 9.125785827636719, 131.7028045654297, 142.6925506591797, 106.05027770996094, 56.295928955078125, -117.97761535644531, 20.794479370117188, 80.95336151123047, -33.08837127685547, 29.305374145507812, 79.5011215209961, 83.80015563964844, -55.00562286376953, 10.241203308105469, 84.85415649414062, -22.930675506591797, 11.754669189453125, 100.73146057128906, 87.72409057617188, 3.5837860107421875, -8.060157775878906, 47.68501281738281, 3.7738800048828125, 89.35084533691406, 55.03016662597656, 1.7303009033203125, 40.48609924316406, -15.320320129394531, -17.960968017578125, 67.19992065429688, 83.13980102539062, 113.400390625, 28.53650665283203, 66.01985931396484, 140.75770568847656, -142.2456817626953, 21.311843872070312, -124.77143859863281, 46.092445373535156, 105.13777160644531, -19.522232055664062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000351.npy"}
{"epoch": 0.5306122448979592, "step": 352, "batch_size": 64, "mean": 30.417400360107422, "std": 60.67817687988281, "min": -96.60028839111328, "p10": -35.94844055175781, "median": 18.32604217529297, "p90": 116.54869766235352, "max": 169.47970581054688, "pos_frac": 0.703125, "sample": [102.27194213867188, 155.7237548828125, 148.72576904296875, 47.288421630859375, 50.56602478027344, -82.62806701660156, 5.9964447021484375, 97.85952758789062, 0.6000518798828125, 75.00962829589844, 8.637535095214844, 31.621726989746094, -76.74470520019531, -5.893896102905273, 56.419281005859375, 27.755508422851562, 7.0455780029296875, -35.98570251464844, -35.86149597167969, 44.45616912841797, 88.75196838378906, 34.232688903808594, -9.841690063476562, 126.2598876953125, 108.84593963623047, 33.1390380859375, 23.879383087158203, -49.15789031982422, -3.5511932373046875, -11.6300048828125, 12.539262771606445, 31.674579620361328, 3.3030242919921875, 48.015079498291016, -17.777328491210938, 9.186431884765625, -11.233936309814453, 146.56103515625, 117.24886322021484, 33.0491943359375, 47.402061462402344, -70.96746826171875, 169.47970581054688, -3.2799110412597656, 16.068233489990234, 114.68413543701172, -59.517982482910156, 19.30455780029297, 4.362939834594727, 71.39456176757812, -96.60028839111328, 30.224639892578125, 18.697372436523438, 114.91497802734375, 17.9547119140625, 3.880035400390625, -2.7900238037109375, 132.71238708496094, -21.696060180664062, -12.404632568359375, 4.99598503112793, 114.59149169921875, 5.8990631103515625, -8.954643249511719], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000352.npy"}
{"epoch": 0.5321239606953893, "step": 353, "batch_size": 64, "mean": 37.6354866027832, "std": 60.95381164550781, "min": -158.184814453125, "p10": -32.82641487121582, "median": 26.78347396850586, "p90": 128.10864562988283, "max": 147.06541442871094, "pos_frac": 0.75, "sample": [66.86743927001953, 145.87754821777344, -61.84291458129883, 31.801353454589844, -33.3878288269043, 24.896305084228516, 10.246650695800781, -46.97015380859375, 24.655258178710938, 132.616455078125, 18.895126342773438, 2.307525634765625, 65.59317016601562, -30.530029296875, 147.06541442871094, 29.70614242553711, 39.03236389160156, 93.33909606933594, 120.6265869140625, 130.729248046875, 59.087646484375, 79.51012420654297, 1.0704421997070312, 67.32537841796875, -5.813819885253906, -0.091949462890625, 111.02655029296875, 25.48560333251953, 132.94168090820312, -0.9661636352539062, 7.580879211425781, 0.844268798828125, 69.39866638183594, -38.91278076171875, -158.184814453125, 120.94943237304688, -36.12750244140625, -67.81182861328125, 129.5945281982422, 110.61727905273438, 44.90336227416992, 92.31970977783203, 143.3190155029297, 19.595314025878906, -11.688003540039062, 7.799491882324219, 64.89910888671875, -6.063899993896484, -2.458984375, 7.727745056152344, 4.186672210693359, 68.35965728759766, 24.571380615234375, 28.081344604492188, 36.762603759765625, 54.655052185058594, 39.98017120361328, 22.193634033203125, -11.248954772949219, -31.516448974609375, 100.45713806152344, 1.9815216064453125, 124.64158630371094, 66.16461944580078], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000353.npy"}
{"epoch": 0.5336356764928194, "step": 354, "batch_size": 64, "mean": 40.35693359375, "std": 63.358680725097656, "min": -110.82695007324219, "p10": -37.10819168090819, "median": 36.49871063232422, "p90": 119.03841629028322, "max": 154.3318634033203, "pos_frac": 0.75, "sample": [154.3318634033203, 75.13504791259766, 4.560890197753906, 21.537071228027344, -18.745025634765625, -12.312784194946289, 40.34990692138672, 87.3326187133789, 139.5198516845703, -2.9576416015625, 52.198204040527344, 109.7057113647461, 121.10993194580078, 30.318115234375, 33.51909255981445, 7.111265182495117, 90.19667053222656, -40.78789520263672, 125.25387573242188, 109.55148315429688, 130.05560302734375, 82.10259246826172, -22.373779296875, 79.92964172363281, -104.45025634765625, -5.2847900390625, -64.47309112548828, 97.19282531738281, 73.26644134521484, 31.23760986328125, 33.66691589355469, 72.44236755371094, 92.49636840820312, 110.51776123046875, 39.33050537109375, 25.191856384277344, -14.545578002929688, -110.82695007324219, 101.2122802734375, -28.522216796875, 101.4664535522461, 137.16885375976562, -54.19136047363281, -19.957157135009766, 121.89340209960938, -109.36576843261719, 87.79830932617188, 81.5357437133789, 11.051559448242188, -53.920654296875, 12.645904541015625, 62.53281784057617, 102.88015747070312, 81.08648681640625, 114.20487976074219, 3.705120086669922, -4.254997253417969, 50.493125915527344, 3.5620269775390625, 12.720123291015625, 17.923433303833008, 67.58206939697266, 0.24337005615234375, 8.94533920288086], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000354.npy"}
{"epoch": 0.5351473922902494, "step": 355, "batch_size": 64, "mean": 40.101593017578125, "std": 60.62971496582031, "min": -112.04150390625, "p10": -23.158495330810545, "median": 33.6904354095459, "p90": 131.2061294555664, "max": 153.0780792236328, "pos_frac": 0.796875, "sample": [41.604835510253906, 110.67550659179688, -39.58976745605469, 103.89368438720703, 69.61895751953125, 111.98373413085938, 43.12000274658203, -21.131378173828125, -2.4215240478515625, 32.39385986328125, 17.892959594726562, 15.121788024902344, 45.53466033935547, 11.266677856445312, 52.455806732177734, 8.759902954101562, -4.542510986328125, 59.169036865234375, -86.76144409179688, 38.56968688964844, -12.172012329101562, 33.82155227661133, 35.78723907470703, 92.92457580566406, 133.08692932128906, 16.71381378173828, 18.7987060546875, 85.57520294189453, 139.08912658691406, 116.51121520996094, -24.027259826660156, 90.99256134033203, 16.222896575927734, 103.74493408203125, 8.282791137695312, -104.1312484741211, 45.005950927734375, -3.2360610961914062, 6.00665283203125, 48.49198913574219, 1.0348052978515625, -0.928619384765625, 135.7574462890625, 6.0278778076171875, 6.602184295654297, 150.3721923828125, 20.0750732421875, 0.39070892333984375, 133.1944580078125, -112.04150390625, 123.29891967773438, 40.9610595703125, 126.81759643554688, 33.55931854248047, 142.3405303955078, -59.942832946777344, 7.449409484863281, 10.143630981445312, 153.0780792236328, -24.656906127929688, 78.45616149902344, 19.069839477539062, 71.13040161132812, 49.20793151855469], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000355.npy"}
{"epoch": 0.5366591080876795, "step": 356, "batch_size": 64, "mean": 38.54914474487305, "std": 72.73888397216797, "min": -140.9590301513672, "p10": -58.36884689331054, "median": 30.519437789916992, "p90": 143.5486587524414, "max": 155.87777709960938, "pos_frac": 0.734375, "sample": [100.42774963378906, 100.84456634521484, 32.95978546142578, -50.188499450683594, 120.95153045654297, 46.89537811279297, -0.019412994384765625, 143.13650512695312, 77.85722351074219, 1.9628124237060547, -87.19664001464844, -140.9590301513672, 82.6861572265625, -61.87471008300781, 145.0848846435547, 150.59988403320312, 5.9066162109375, 8.414512634277344, 83.43280029296875, -2.213520050048828, 144.59298706054688, -134.17013549804688, 46.33367919921875, 33.44914245605469, 22.79651641845703, 147.55616760253906, -6.337043762207031, -74.90025329589844, 28.079090118408203, 40.76348114013672, 58.80516815185547, 8.149223327636719, 3.578868865966797, 3.891887664794922, -76.66270446777344, 115.79733276367188, 94.32855224609375, 11.172309875488281, 88.94808197021484, 103.09003448486328, -0.6817722320556641, 17.347885131835938, 53.36753845214844, -16.04486083984375, 26.082275390625, 12.442447662353516, 129.3759765625, 23.782608032226562, -8.2352294921875, 68.58708190917969, -10.519706726074219, 22.24618148803711, 142.12635803222656, 151.45474243164062, -22.818954467773438, 127.31391906738281, 46.57619094848633, 2.1244583129882812, 143.7252960205078, -49.20103454589844, -94.85243225097656, 155.87777709960938, 43.26864242553711, 85.828857421875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000356.npy"}
{"epoch": 0.5381708238851096, "step": 357, "batch_size": 64, "mean": 33.18156051635742, "std": 60.75816345214844, "min": -115.36149597167969, "p10": -29.75768814086914, "median": 23.491241455078125, "p90": 123.29677963256839, "max": 155.2151641845703, "pos_frac": 0.6875, "sample": [131.98043823242188, -3.8453712463378906, 155.2151641845703, 36.640037536621094, -59.79404830932617, 149.09854125976562, 62.67942810058594, 40.70113754272461, 14.503463745117188, -4.258087158203125, -64.13105773925781, 33.466041564941406, -28.430923461914062, 27.680984497070312, 153.8350830078125, 52.775146484375, 126.5522232055664, 42.13812255859375, 85.52474975585938, 9.699954986572266, 47.43153381347656, 15.342422485351562, 25.479995727539062, 36.214698791503906, 18.74219512939453, 56.2019157409668, -4.607870101928711, -30.32630157470703, -71.9500732421875, 3.736339569091797, 13.694526672363281, 90.46359252929688, -16.923980712890625, 46.14655303955078, 80.04782104492188, 139.75503540039062, 91.46176147460938, -2.8310699462890625, 39.54087829589844, 144.9856719970703, -3.6648025512695312, 107.4368667602539, -14.24557113647461, 6.799957275390625, 52.23298645019531, 115.70074462890625, -1.4390544891357422, -24.366744995117188, -10.586006164550781, 12.169105529785156, 5.927135467529297, 81.69392395019531, 84.04093933105469, 2.5234146118164062, 98.63700866699219, -6.017608642578125, 10.148662567138672, 21.502487182617188, 97.8900146484375, -102.79666137695312, -41.339385986328125, -115.36149597167969, 62.89386749267578, -0.7965888977050781], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000357.npy"}
{"epoch": 0.5396825396825397, "step": 358, "batch_size": 64, "mean": 44.70811080932617, "std": 61.56004333496094, "min": -107.16242980957031, "p10": -23.26435623168945, "median": 36.374380111694336, "p90": 121.64589157104493, "max": 162.17486572265625, "pos_frac": 0.71875, "sample": [42.45155715942383, 103.80008697509766, 110.61514282226562, 107.36402893066406, -12.653207778930664, 114.70496368408203, 96.90576171875, -22.140907287597656, -4.743076324462891, -23.745834350585938, -1.7834510803222656, 120.31896209716797, 42.85966491699219, 23.087928771972656, 132.87405395507812, -20.329437255859375, -19.560577392578125, 22.587249755859375, 82.56097412109375, 92.69944763183594, 32.639373779296875, 49.298336029052734, 33.3170166015625, 83.10478210449219, 120.7823257446289, 1.4335098266601562, 8.071548461914062, -27.44976043701172, -10.000892639160156, 3.9858856201171875, -30.252578735351562, 28.732879638671875, -29.510223388671875, 6.029064178466797, -2.1414413452148438, 14.423782348632812, 110.54186248779297, 110.75460815429688, 34.391517639160156, -3.2378616333007812, 140.59136962890625, 96.92437744140625, 122.0159912109375, 3.702289581298828, 162.17486572265625, 151.21661376953125, 38.357242584228516, -40.72886657714844, 17.717811584472656, 75.10647583007812, -13.528377532958984, -107.16242980957031, -101.78517150878906, 140.74032592773438, 26.900390625, 103.54200744628906, -10.736618041992188, 88.94642639160156, 41.98682403564453, 40.64198684692383, 76.93769836425781, 94.1379165649414, 53.49815368652344, 137.33468627929688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000358.npy"}
{"epoch": 0.5411942554799698, "step": 359, "batch_size": 64, "mean": 25.97496795654297, "std": 59.16669845581055, "min": -116.45428466796875, "p10": -30.905977249145504, "median": 9.5693359375, "p90": 120.56938400268557, "max": 151.81004333496094, "pos_frac": 0.640625, "sample": [9.004150390625, 115.53482818603516, 0.4083137512207031, 16.892929077148438, 63.236419677734375, 13.331565856933594, -8.92264175415039, 45.03227233886719, -56.85408020019531, -53.33428955078125, 151.81004333496094, 1.3341712951660156, 42.017921447753906, 60.17793273925781, -0.4198036193847656, 55.93101501464844, 123.6658935546875, -31.87833023071289, 7.511222839355469, -116.45428466796875, 14.76082992553711, 67.44563293457031, 6.8185272216796875, -42.08607482910156, 8.759559631347656, 28.10765838623047, 98.6129150390625, -3.5510482788085938, 5.5584564208984375, -98.76295471191406, 129.88290405273438, 139.4878692626953, -16.724998474121094, 17.46853256225586, -60.767372131347656, 24.717025756835938, 68.55839538574219, 75.61199951171875, -12.22726821899414, 4.44525146484375, -25.372844696044922, 113.64646911621094, 150.70175170898438, 64.5932846069336, -28.63715362548828, 10.134521484375, -2.0513458251953125, -17.126983642578125, -0.6179008483886719, 24.44623565673828, 14.980533599853516, -11.453725814819336, -19.263519287109375, -22.34877586364746, 94.1137924194336, -6.110511779785156, 122.72705078125, 34.38066101074219, 71.7513427734375, 74.24148559570312, 4.600215911865234, 145.6485595703125, -17.715957641601562, -7.010322570800781], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000359.npy"}
{"epoch": 0.5427059712773998, "step": 360, "batch_size": 64, "mean": 36.662925720214844, "std": 61.432647705078125, "min": -81.24394226074219, "p10": -19.47937469482422, "median": 19.962553024291992, "p90": 140.80062866210938, "max": 166.521484375, "pos_frac": 0.6875, "sample": [41.032745361328125, 160.17633056640625, 32.381103515625, 2.221923828125, 85.08235168457031, 55.25090789794922, -14.75771713256836, 75.89363098144531, -19.880462646484375, 166.521484375, -0.6403579711914062, -18.543502807617188, -40.57933044433594, -1.55078125, 76.92294311523438, -6.891815185546875, 95.971435546875, -55.937828063964844, 20.976749420166016, 142.336669921875, 7.3883056640625, -2.1750717163085938, -31.538543701171875, 0.9291915893554688, -12.087982177734375, 100.6899185180664, 43.7015380859375, 1.5176620483398438, 110.59291076660156, -3.5559234619140625, 7.7510223388671875, 20.51398468017578, 45.63328552246094, -11.961990356445312, -81.24394226074219, 127.97109985351562, -1.97027587890625, 58.680938720703125, 12.389793395996094, 2.8024978637695312, 160.05154418945312, -2.6753692626953125, 53.63141632080078, -15.666500091552734, 90.0406494140625, 31.870285034179688, 141.03207397460938, 19.411121368408203, 25.25885009765625, 137.15318298339844, 140.26058959960938, 20.998016357421875, 146.19154357910156, -38.704132080078125, 11.444007873535156, 76.74951171875, 0.09052467346191406, -79.68063354492188, 10.696794509887695, 34.543495178222656, 4.35809326171875, -4.6649322509765625, 40.040260314941406, 151.98196411132812], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000360.npy"}
{"epoch": 0.54421768707483, "step": 361, "batch_size": 64, "mean": 46.20123291015625, "std": 61.170814514160156, "min": -120.49183654785156, "p10": -12.87688636779785, "median": 35.77136993408203, "p90": 133.87224884033205, "max": 172.5376739501953, "pos_frac": 0.71875, "sample": [2.00738525390625, 90.00361633300781, -13.373222351074219, -2.425884246826172, 102.60063934326172, -0.2282562255859375, -10.266059875488281, -33.9666633605957, 45.70600891113281, 68.86592102050781, 9.791107177734375, -9.120635986328125, -11.718769073486328, 115.16532897949219, 114.12451171875, 30.01458740234375, -30.894027709960938, -22.372161865234375, 92.60173034667969, 11.968170166015625, 60.335357666015625, 138.90240478515625, -3.251617431640625, 145.82391357421875, 92.39482879638672, 43.35987854003906, 94.43009948730469, -120.49183654785156, 147.4502410888672, -1.2005958557128906, 28.692134857177734, 24.5125732421875, 67.53399658203125, 165.394287109375, 79.03984069824219, 41.52815246582031, 134.19754028320312, 20.70568084716797, -74.04281616210938, -5.909446716308594, 8.539606094360352, -0.6293411254882812, 133.1132354736328, 105.02006530761719, 98.71460723876953, 127.80052947998047, 72.37003326416016, -2.4336013793945312, 91.73872375488281, 18.444618225097656, 86.95172882080078, 15.251136779785156, 70.64439392089844, 5.1313323974609375, 172.5376739501953, 19.045608520507812, 138.00621032714844, 10.046180725097656, -46.338134765625, 19.33557891845703, 46.74727249145508, 83.22525787353516, 61.48783874511719, -5.7597808837890625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000361.npy"}
{"epoch": 0.54572940287226, "step": 362, "batch_size": 64, "mean": 32.0594482421875, "std": 65.18132019042969, "min": -140.8440399169922, "p10": -32.475660705566405, "median": 22.250091552734375, "p90": 124.47196655273437, "max": 220.0008544921875, "pos_frac": 0.75, "sample": [-18.482337951660156, 22.9505615234375, 0.214111328125, -11.777931213378906, 25.26394271850586, -140.8440399169922, 48.72666931152344, 76.68556213378906, 220.0008544921875, 59.31660461425781, 5.1199493408203125, 0.35825347900390625, 24.202239990234375, 61.66980743408203, -13.62601089477539, 46.800392150878906, 0.127410888671875, 21.872100830078125, 2.317352294921875, 135.7329559326172, -16.01880645751953, 7.2526397705078125, 8.936920166015625, -32.534515380859375, 124.771728515625, 31.73974609375, 123.77252197265625, -75.99053955078125, -50.219303131103516, 52.71566390991211, 127.721435546875, 7.315187454223633, -6.562652587890625, 77.2484130859375, 3.6329345703125, 159.24014282226562, 47.156131744384766, -33.419586181640625, 72.40545654296875, -13.43603515625, 22.628082275390625, 104.55445861816406, 102.25447082519531, 75.47820281982422, 7.736915588378906, 99.53903198242188, 67.16289520263672, 112.54257202148438, 72.5195083618164, -22.3155517578125, -27.08251953125, 34.03131103515625, 17.538612365722656, -120.93400573730469, 15.647239685058594, 44.897918701171875, 129.31275939941406, -83.61459350585938, -32.33833312988281, 16.5850830078125, 145.46450805664062, 13.617294311523438, 18.540023803710938, 55.682647705078125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000362.npy"}
{"epoch": 0.54724111866969, "step": 363, "batch_size": 64, "mean": 44.12169647216797, "std": 56.699703216552734, "min": -87.96994018554688, "p10": -1.4033084869384755, "median": 34.11709403991699, "p90": 126.09786529541017, "max": 170.93243408203125, "pos_frac": 0.875, "sample": [51.660186767578125, -87.96994018554688, 134.577880859375, 37.58893585205078, 40.34820556640625, 64.78901672363281, 30.645252227783203, 170.93243408203125, 10.888969421386719, 17.965232849121094, 26.476043701171875, 64.44168090820312, 141.1278076171875, 114.81956481933594, 37.792537689208984, 152.46592712402344, -1.8528518676757812, 59.946876525878906, 15.497421264648438, 127.47880554199219, 45.644039154052734, 69.7719955444336, 18.70559310913086, 78.70474243164062, 114.58331298828125, 44.40379333496094, 4.074462890625, 76.68682861328125, -49.91381072998047, 4.033180236816406, -65.66310119628906, -84.25170135498047, 52.552276611328125, -75.80733489990234, 78.91681671142578, 60.70862579345703, 26.738216400146484, 30.58646011352539, 28.318513870239258, -0.3543739318847656, 13.921737670898438, 18.143695831298828, 76.28150939941406, 147.67364501953125, 5.787803649902344, 149.7051239013672, 96.89212036132812, 52.1370849609375, 111.40580749511719, 17.457908630371094, 102.85137176513672, 24.35770034790039, 122.87567138671875, 0.6245765686035156, -8.95870590209961, 3.9610595703125, 68.93980407714844, 23.510778427124023, 23.384918212890625, 60.61397933959961, 1.3282394409179688, 16.6646728515625, 18.13634490966797, 8.033126831054688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000363.npy"}
{"epoch": 0.5487528344671202, "step": 364, "batch_size": 64, "mean": 42.31687545776367, "std": 73.32832336425781, "min": -119.17048645019531, "p10": -47.62224655151367, "median": 23.679218292236328, "p90": 153.54473419189455, "max": 170.01119995117188, "pos_frac": 0.734375, "sample": [155.80020141601562, -0.8095054626464844, 19.523937225341797, 159.33673095703125, 157.24179077148438, 1.8921279907226562, 23.283004760742188, 101.49803161621094, -13.285026550292969, 109.30166625976562, 159.43115234375, 30.403976440429688, -48.44611358642578, -5.041099548339844, 44.47920227050781, 154.99671936035156, 1.3497886657714844, 16.76156997680664, 170.01119995117188, 4.8968353271484375, 94.56513977050781, 14.38698959350586, 12.537330627441406, 150.15676879882812, -109.1268310546875, -39.25602722167969, 115.24359130859375, -0.429779052734375, 3.01715087890625, 49.623992919921875, 122.65069580078125, 116.98532104492188, 15.87088394165039, 38.429931640625, 9.395072937011719, 24.07543182373047, 123.15536499023438, 121.0201416015625, -45.69989013671875, -49.98233413696289, 43.78042984008789, 26.098770141601562, -19.29901123046875, -48.932586669921875, 112.70407104492188, 4.644287109375, -76.45886993408203, 101.71504211425781, 13.088874816894531, 158.27639770507812, -5.612754821777344, 103.9074935913086, -119.17048645019531, 39.95428466796875, -105.43917083740234, 72.760498046875, 49.522254943847656, 139.29702758789062, 5.058170318603516, 5.783653259277344, -1.2868423461914062, -25.879043579101562, 121.0604019165039, 103.46195983886719], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000364.npy"}
{"epoch": 0.5502645502645502, "step": 365, "batch_size": 64, "mean": 40.340911865234375, "std": 53.938438415527344, "min": -108.64702606201172, "p10": -14.162694168090818, "median": 20.77633285522461, "p90": 114.12872772216798, "max": 152.86721801757812, "pos_frac": 0.765625, "sample": [30.938732147216797, 36.780113220214844, 87.79906463623047, 12.20663833618164, 39.96681213378906, 12.38247299194336, 108.40786743164062, 55.09673309326172, 109.81495666503906, 13.989456176757812, 111.6358642578125, 9.967643737792969, -108.64702606201172, -0.017917633056640625, 10.806846618652344, 135.4535369873047, -1.6356678009033203, 46.46875762939453, 2.61383056640625, 56.60646057128906, 23.987091064453125, 8.785064697265625, 110.68379974365234, -16.841001510620117, -7.519935607910156, -17.655014038085938, 36.18761444091797, 146.04286193847656, -0.491668701171875, 115.19709777832031, -12.138137817382812, 6.209726333618164, 80.45530700683594, 9.647563934326172, 95.2984619140625, 100.09318542480469, 99.81849670410156, 0.7707290649414062, -3.4419479370117188, 76.75536346435547, 84.27442932128906, 9.630828857421875, 90.95865631103516, 3.6943359375, 42.2066650390625, 140.13946533203125, 4.548980712890625, -15.03036117553711, 119.748779296875, 6.382099151611328, 51.474517822265625, 1.0401115417480469, 64.43283081054688, -27.53082275390625, 17.565574645996094, 8.748992919921875, -19.951831817626953, -42.10836410522461, 79.30686950683594, -6.123363494873047, 152.86721801757812, 115.53874969482422, -7.642612457275391, 85.16690063476562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000365.npy"}
{"epoch": 0.5517762660619804, "step": 366, "batch_size": 64, "mean": 35.035865783691406, "std": 69.108154296875, "min": -121.13526916503906, "p10": -54.97842025756834, "median": 15.215842247009277, "p90": 138.61712341308598, "max": 163.3368377685547, "pos_frac": 0.671875, "sample": [15.52235221862793, 148.73495483398438, 93.97093200683594, 75.29293823242188, -0.0143585205078125, 127.2895278930664, -16.7286376953125, -68.88401794433594, -96.72756958007812, 4.221038818359375, -1.8699417114257812, 98.57032775878906, 17.969715118408203, 73.32054138183594, 34.75011444091797, -34.416961669921875, -63.79047393798828, 26.951629638671875, 7.191551208496094, 120.7648696899414, 14.909332275390625, -6.075462341308594, 41.169219970703125, -0.10567283630371094, -2.7852554321289062, 72.02340698242188, 5.161323547363281, 97.77488708496094, -65.63673400878906, 11.525543212890625, 163.3368377685547, 9.554634094238281, -12.84033203125, 153.67361450195312, -121.13526916503906, 8.302375793457031, 3.6545944213867188, -28.92552947998047, 27.895397186279297, 8.19482421875, -0.00402069091796875, -19.65245819091797, 1.8568744659423828, 4.9299774169921875, -83.57331848144531, 17.655479431152344, -25.416770935058594, 84.75173950195312, 148.84971618652344, 82.0589599609375, 40.43547821044922, 145.72244262695312, 92.58331298828125, -63.88085174560547, 117.50548553466797, -20.88141632080078, 154.98004150390625, 97.51626586914062, 129.95883178710938, 142.32781982421875, 127.33631896972656, 92.16166687011719, 41.31098175048828, -8.027433395385742], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000366.npy"}
{"epoch": 0.5532879818594104, "step": 367, "batch_size": 64, "mean": 31.818260192871094, "std": 59.64484786987305, "min": -139.290283203125, "p10": -25.471389579772943, "median": 31.539657592773438, "p90": 107.97613754272462, "max": 139.56028747558594, "pos_frac": 0.703125, "sample": [6.961101531982422, 101.14254760742188, 106.4392318725586, -64.09308624267578, 52.782344818115234, -1.2143898010253906, 74.7659912109375, 131.65634155273438, -12.065773010253906, 14.515789031982422, 2.6463623046875, 37.50315856933594, 24.15686798095703, 2.217041015625, 95.82502746582031, 37.9817008972168, 47.939483642578125, -14.322891235351562, -15.468429565429688, 122.62030792236328, -12.404876708984375, 72.17143249511719, -5.127452850341797, -55.8316764831543, 16.71771240234375, -125.74082946777344, 10.78701400756836, 38.34889221191406, 20.741649627685547, 108.63481140136719, 75.67637634277344, 78.88883209228516, 83.12667083740234, 25.576156616210938, 38.53668975830078, 1.0203361511230469, 48.569580078125, -37.75773620605469, 51.59516906738281, 82.10472869873047, -13.988117218017578, -139.290283203125, 2.20660400390625, 139.56028747558594, 67.82188415527344, -10.7476806640625, 17.252464294433594, 39.2282600402832, 50.83197021484375, -14.982589721679688, 14.038436889648438, 41.91172790527344, -85.28804016113281, 127.43719482421875, 97.92805480957031, 101.1672592163086, -13.880138397216797, -28.74688148498535, 118.70258331298828, 50.378089904785156, -17.828575134277344, 94.55854797363281, -0.07696533203125, 130.5524139404297], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000367.npy"}
{"epoch": 0.5547996976568406, "step": 368, "batch_size": 64, "mean": 39.47632598876953, "std": 65.3609390258789, "min": -109.0461196899414, "p10": -30.304658889770504, "median": 33.70068550109863, "p90": 137.5193313598633, "max": 165.00747680664062, "pos_frac": 0.6875, "sample": [-88.42214965820312, 83.34649658203125, -4.584644317626953, 88.54124450683594, -7.5352783203125, 75.23490905761719, 137.8445587158203, 41.46675109863281, 0.6396713256835938, -3.399749755859375, 54.525115966796875, -5.767463684082031, -13.124832153320312, 70.12699890136719, 154.53866577148438, -47.99285125732422, -2.5677642822265625, -74.16539001464844, -19.907928466796875, 136.76046752929688, 7.739067077636719, 50.267303466796875, 35.10565948486328, -14.191429138183594, 143.9365997314453, 20.85001564025879, -17.761856079101562, 131.84327697753906, 159.3430938720703, 136.10836791992188, -3.08172607421875, 153.33859252929688, 135.3458709716797, 32.295711517333984, 73.065185546875, 83.6746597290039, 49.284584045410156, 89.36488342285156, 6.3653717041015625, 20.116546630859375, 35.959686279296875, -23.427810668945312, 150.60476684570312, 39.5341796875, -25.395267486572266, -6.809043884277344, 44.93852996826172, -32.40868377685547, 90.38225555419922, -54.02978515625, 72.53996276855469, 25.284225463867188, 165.00747680664062, 73.40148162841797, 19.636295318603516, 24.90972900390625, 37.49668884277344, -68.56785583496094, 8.60189437866211, 83.84404754638672, 80.17955017089844, 11.309757232666016, -109.0461196899414, 13.972503662109375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000368.npy"}
{"epoch": 0.5563114134542706, "step": 369, "batch_size": 64, "mean": 26.496131896972656, "std": 63.249977111816406, "min": -109.54093933105469, "p10": -57.944987487792964, "median": 15.543636322021484, "p90": 124.16342697143556, "max": 162.5362091064453, "pos_frac": 0.71875, "sample": [131.2581787109375, -85.2211685180664, 11.312557220458984, 46.289154052734375, -96.0968017578125, 34.23063659667969, 99.42217254638672, 43.683555603027344, 122.2622299194336, 5.9812469482421875, 5.0107574462890625, 67.7206039428711, 91.58457946777344, 104.15106201171875, 143.95388793945312, 21.929302215576172, 1.4621248245239258, 13.509048461914062, -13.021072387695312, 6.0637054443359375, 53.63744354248047, -109.54093933105469, 58.99848937988281, -44.626922607421875, 56.739967346191406, 29.721542358398438, -62.375877380371094, 99.83177947998047, 17.578224182128906, -63.040496826171875, 125.19369506835938, -2.1290206909179688, 17.580204010009766, 143.9212646484375, 48.72554016113281, 124.97822570800781, 156.7965545654297, 5.340484619140625, 18.325393676757812, -92.78990173339844, -52.96339416503906, 28.840736389160156, 5.414482116699219, 28.78978729248047, 6.785663604736328, -7.5117950439453125, 20.39410400390625, -0.6198654174804688, 29.874053955078125, -17.927513122558594, 0.7612762451171875, 11.001708984375, 1.2029685974121094, 78.16304016113281, -12.706024169921875, -8.806339263916016, 2.5857620239257812, 162.5362091064453, 10.368789672851562, 78.60337829589844, -60.0799560546875, -14.2720947265625, -34.59762954711914, 101.56359100341797], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000369.npy"}
{"epoch": 0.5578231292517006, "step": 370, "batch_size": 64, "mean": 51.689735412597656, "std": 72.73224639892578, "min": -144.62799072265625, "p10": -23.610411453247067, "median": 46.466997146606445, "p90": 150.48071136474613, "max": 164.98963928222656, "pos_frac": 0.734375, "sample": [40.64203643798828, 155.39926147460938, 119.80950927734375, 72.35372924804688, 163.1219482421875, 1.2218666076660156, 164.98963928222656, -14.995529174804688, 152.82220458984375, 79.05522155761719, 34.71392059326172, -7.860748291015625, 4.712860107421875, 117.89730072021484, 145.01722717285156, 76.10179901123047, 16.681655883789062, -61.41564178466797, -22.533241271972656, 77.86181640625, -29.2611083984375, 123.25662231445312, 144.95071411132812, 32.1221923828125, 126.60102844238281, 67.80616760253906, 95.369140625, -42.11973571777344, 24.240280151367188, -2.5756607055664062, 1.1580543518066406, -22.36553955078125, 19.0924072265625, -10.259284973144531, -0.851043701171875, 73.81869506835938, 155.45333862304688, 84.20853424072266, 154.13082885742188, 128.9991455078125, 141.23281860351562, -120.1298828125, 13.501544952392578, 11.52703857421875, 104.67599487304688, 153.72964477539062, 92.49652099609375, 118.92415618896484, 100.58935546875, 127.41084289550781, -144.62799072265625, 43.68290710449219, 88.68141174316406, -1.7646751403808594, 11.623119354248047, 0.8416366577148438, 52.0926513671875, -3.4975509643554688, -24.07205581665039, -2.3488197326660156, 11.436904907226562, 130.48326110839844, -86.96844482421875, 49.2510871887207], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000370.npy"}
{"epoch": 0.5593348450491308, "step": 371, "batch_size": 64, "mean": 42.74407958984375, "std": 70.41353607177734, "min": -119.2830810546875, "p10": -46.2608097076416, "median": 43.49521827697754, "p90": 135.56940612792968, "max": 167.1683807373047, "pos_frac": 0.765625, "sample": [146.618408203125, -119.2830810546875, 167.1683807373047, 63.37986755371094, 134.48817443847656, 1.87548828125, 109.1629867553711, -0.4469795227050781, 40.053802490234375, -0.020235061645507812, 11.9207763671875, 107.16117095947266, -92.03634643554688, 155.39981079101562, 53.221771240234375, 102.5299072265625, 1.7226181030273438, 61.642513275146484, 9.706579208374023, 150.27593994140625, 18.113845825195312, -117.3978271484375, 14.898845672607422, 31.274005889892578, 34.9921875, 133.31314086914062, 44.99272918701172, 164.07513427734375, 111.75240325927734, 46.8217887878418, 29.34771728515625, 134.01947021484375, -34.37970733642578, -27.161598205566406, 67.85594177246094, 5.865272521972656, 45.09944152832031, -67.99504852294922, -10.397720336914062, 45.51811218261719, -42.36661911010742, 55.10480499267578, -47.92974853515625, 49.262367248535156, 53.36323547363281, -74.19038391113281, 75.04996490478516, 37.872894287109375, -17.812278747558594, 118.16812133789062, 41.99770736694336, -10.452268600463867, 15.771600723266602, 111.84597778320312, 115.05950927734375, 5.198575973510742, 151.49781799316406, 65.41358947753906, 2.784393310546875, 81.06302642822266, 104.03775024414062, 136.0327911376953, 22.698997497558594, -89.00051879882812], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000371.npy"}
{"epoch": 0.5608465608465608, "step": 372, "batch_size": 64, "mean": 34.201377868652344, "std": 76.28560638427734, "min": -135.74160766601562, "p10": -57.01190414428711, "median": 16.975546836853027, "p90": 142.52586364746094, "max": 163.76560974121094, "pos_frac": 0.640625, "sample": [-101.12464904785156, 65.48731231689453, 130.1575164794922, -24.33556365966797, 78.04115295410156, 30.877098083496094, 145.35720825195312, -35.11637878417969, 130.70230102539062, 42.76575469970703, 143.1016845703125, 47.825843811035156, 71.87844848632812, 163.76560974121094, 34.136199951171875, 35.87535095214844, 75.28089904785156, -135.74160766601562, -1.8350296020507812, 58.4813232421875, 13.853248596191406, -4.5258026123046875, 14.051506042480469, 141.18228149414062, -9.354120254516602, 1.8312225341796875, -11.0347900390625, 127.3642578125, 161.01177978515625, 17.69340705871582, 144.68246459960938, 92.81346130371094, -56.47579574584961, -13.99819564819336, -31.26593017578125, 90.7323989868164, -126.16796112060547, 39.96495819091797, 130.80201721191406, 12.862159729003906, 128.10711669921875, 15.203330993652344, 98.60272216796875, 144.77938842773438, 29.41845703125, 94.91233825683594, -57.24166488647461, 5.45806884765625, 16.257686614990234, -62.1558837890625, 11.97176742553711, -108.5121078491211, 107.19340515136719, -1.3357505798339844, -27.84838104248047, -19.755329132080078, -83.56525421142578, -41.139366149902344, 139.10092163085938, 154.93116760253906, -24.695579528808594, -3.559478759765625, -20.35748291015625, 1.5151710510253906], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000372.npy"}
{"epoch": 0.562358276643991, "step": 373, "batch_size": 64, "mean": 38.09349822998047, "std": 74.06066131591797, "min": -137.27841186523438, "p10": -37.86333999633789, "median": 24.6796817779541, "p90": 144.250439453125, "max": 175.32725524902344, "pos_frac": 0.671875, "sample": [-28.788814544677734, 35.238182067871094, 125.6817855834961, 50.964324951171875, 106.18099212646484, -120.04815673828125, 59.542022705078125, 144.75885009765625, 105.88027954101562, 117.9858169555664, 102.57472229003906, -82.57056427001953, 2.9945526123046875, -23.556182861328125, -137.27841186523438, 2.2769813537597656, 25.712947845458984, -19.410400390625, -37.85948944091797, 20.86343002319336, 7.506278991699219, -12.233713150024414, 70.7005615234375, -42.26415252685547, 75.39491271972656, 136.41552734375, -12.126556396484375, 105.95436096191406, 35.92616271972656, 39.95165252685547, 161.81480407714844, -4.543281555175781, 11.883056640625, 156.13385009765625, -37.864990234375, -3.6493492126464844, -4.91387939453125, 12.750732421875, -5.8775787353515625, 143.06414794921875, 125.53792572021484, -110.2327880859375, -58.825416564941406, -5.377510070800781, -0.19489288330078125, 70.76905822753906, -34.8540153503418, 154.17034912109375, -24.75725555419922, 30.011734008789062, 51.78247833251953, 167.05111694335938, 3.3740501403808594, 175.32725524902344, 64.92906188964844, 35.772003173828125, 50.10075378417969, 97.60955810546875, 23.64641571044922, 165.38336181640625, 9.413837432861328, 1.3564910888671875, 19.776763916015625, 141.04830932617188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000373.npy"}
{"epoch": 0.563869992441421, "step": 374, "batch_size": 64, "mean": 59.59770965576172, "std": 68.91622161865234, "min": -130.4908447265625, "p10": -10.008390808105467, "median": 59.1057014465332, "p90": 151.55732727050784, "max": 179.42660522460938, "pos_frac": 0.78125, "sample": [174.95611572265625, 48.22016906738281, 14.320480346679688, -20.324859619140625, 2.2476634979248047, 140.25738525390625, 142.2918701171875, 134.65435791015625, 155.70794677734375, 114.91352844238281, 89.15544128417969, -10.87493896484375, 58.848114013671875, 179.42660522460938, 10.125564575195312, -33.837005615234375, 156.83380126953125, 8.058464050292969, 6.852874755859375, 7.1395111083984375, -9.286911010742188, 123.14308166503906, 61.19736862182617, -10.317596435546875, 104.82615661621094, 29.01816177368164, 109.8817138671875, 113.0057144165039, -6.451324462890625, 55.863059997558594, 59.36328887939453, 109.19029998779297, 9.143657684326172, 39.012786865234375, 127.6597671508789, 6.82818603515625, -1.7035140991210938, -130.4908447265625, 90.9803466796875, 88.54823303222656, 61.462493896484375, -122.3808364868164, 2.4752979278564453, 163.7010498046875, 164.24070739746094, 58.50944519042969, -11.124740600585938, -7.000419616699219, 93.68228149414062, -2.842041015625, 129.2093048095703, 108.176025390625, 155.14175415039062, 97.16495513916016, 67.1211166381836, 3.932819366455078, 138.19305419921875, 82.62326049804688, 58.61389923095703, 143.19366455078125, 13.900802612304688, 74.75274658203125, -3.6845321655273438, -3.1934356689453125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000374.npy"}
{"epoch": 0.5653817082388511, "step": 375, "batch_size": 64, "mean": 40.572608947753906, "std": 68.56319427490234, "min": -122.81816101074219, "p10": -29.66341972351074, "median": 30.243148803710938, "p90": 129.45578002929688, "max": 166.319091796875, "pos_frac": 0.75, "sample": [56.666656494140625, 28.688735961914062, -17.71337890625, 121.37115478515625, 77.18612670898438, 19.15953254699707, 155.8505096435547, -15.17239761352539, 6.6488494873046875, 44.85371780395508, 31.797561645507812, 85.79559326171875, -1.2402191162109375, 145.77984619140625, 84.15052795410156, -9.134445190429688, 105.37834930419922, -76.171142578125, 78.88814544677734, -18.523954391479492, 39.091094970703125, 53.23625946044922, 105.66605377197266, -5.0289306640625, 79.97724914550781, 127.98453521728516, 18.526397705078125, -77.87332153320312, 72.11009216308594, 129.3624267578125, 15.159297943115234, -2.2019615173339844, 4.392467498779297, -30.699630737304688, 103.03565979003906, 43.892906188964844, 78.14926147460938, 5.8110198974609375, 166.319091796875, 0.7882804870605469, 104.05517578125, 22.734207153320312, 145.30032348632812, -122.81816101074219, 22.052642822265625, 12.53570556640625, 156.10580444335938, 5.032936096191406, 94.4298095703125, 59.141754150390625, 69.32221984863281, 12.373489379882812, -93.9196548461914, 21.94955825805664, 21.636768341064453, 139.04168701171875, 113.20513916015625, -120.9974365234375, -27.245594024658203, 129.49578857421875, -10.413108825683594, 92.08869934082031, 4.0015716552734375, -84.42041015625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000375.npy"}
{"epoch": 0.5668934240362812, "step": 376, "batch_size": 64, "mean": 51.301605224609375, "std": 77.24413299560547, "min": -119.00216674804688, "p10": -37.85110397338867, "median": 48.698333740234375, "p90": 159.86123962402345, "max": 176.95455932617188, "pos_frac": 0.75, "sample": [14.200927734375, 76.0145492553711, -4.401512145996094, 14.006324768066406, 97.43132781982422, -36.134498596191406, -14.263603210449219, 144.57688903808594, 90.63267517089844, 9.623371124267578, 78.460205078125, 86.21116638183594, -102.73991394042969, -7.488492965698242, -31.216140747070312, -61.62831115722656, 161.25302124023438, 16.803741455078125, 76.7146987915039, 176.95455932617188, 9.3267822265625, 169.60198974609375, -3.5816879272460938, 82.7801284790039, 159.02627563476562, -1.253936767578125, 159.03872680664062, 152.953857421875, 82.60531616210938, 114.55250549316406, 140.96388244628906, 107.5867691040039, 154.95858764648438, 10.697128295898438, 26.45612335205078, 162.39927673339844, 7.89434814453125, -96.01264953613281, 160.2137451171875, -77.65669250488281, 50.894134521484375, -34.21735382080078, 79.5064468383789, 66.95963287353516, 18.45532989501953, 175.53762817382812, -61.170570373535156, 80.10884857177734, -38.5867919921875, 114.31001281738281, -30.919342041015625, 5.092227935791016, 33.62843322753906, 55.686302185058594, 162.01368713378906, 7.015155792236328, 72.1116943359375, 46.502532958984375, 145.07882690429688, 115.18122863769531, 4.393211364746094, 12.718917846679688, 14.44314956665039, -119.00216674804688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000376.npy"}
{"epoch": 0.5684051398337112, "step": 377, "batch_size": 64, "mean": 28.120010375976562, "std": 65.12940216064453, "min": -119.14077758789062, "p10": -62.68948173522948, "median": 18.187946319580078, "p90": 126.21025314331057, "max": 163.59335327148438, "pos_frac": 0.65625, "sample": [128.4371795654297, -67.98550415039062, 121.01409149169922, 142.91046142578125, 90.44721221923828, 8.869232177734375, -89.01072692871094, -119.14077758789062, 19.680496215820312, 53.25469207763672, 33.45256042480469, -19.778274536132812, 63.74967956542969, 6.125205993652344, 49.66558074951172, 113.05610656738281, -78.91693115234375, 41.97551727294922, -21.59762191772461, -9.177841186523438, 143.144287109375, -3.4259281158447266, 6.977928161621094, 32.041961669921875, 114.69535827636719, -85.90179443359375, 0.12290573120117188, 135.9888458251953, -8.600044250488281, -27.056774139404297, 65.44529724121094, -10.256263732910156, 112.38789367675781, -28.37435531616211, 21.905776977539062, -71.6335220336914, 14.509002685546875, 26.502410888671875, 88.36068725585938, 29.153335571289062, -16.53691864013672, 37.681861877441406, 22.955097198486328, 8.903209686279297, 39.145782470703125, 151.6625213623047, 65.94412231445312, -50.332096099853516, -6.389656066894531, 31.783802032470703, 74.54510498046875, 4.5936279296875, -9.4151611328125, 163.59335327148438, -0.4048919677734375, -3.3873519897460938, 152.0593719482422, 101.217041015625, -73.90913391113281, 10.59500503540039, 43.259891510009766, -1.039398193359375, 16.695396423339844, 13.442626953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000377.npy"}
{"epoch": 0.5699168556311414, "step": 378, "batch_size": 64, "mean": 29.75958251953125, "std": 66.46939849853516, "min": -155.72470092773438, "p10": -29.937714004516593, "median": 20.154754638671875, "p90": 120.64833221435548, "max": 177.98419189453125, "pos_frac": 0.765625, "sample": [42.1842041015625, 10.296180725097656, 29.611669540405273, 14.026241302490234, 11.614990234375, 141.99647521972656, -6.5757904052734375, 17.395660400390625, 7.5505523681640625, 44.28155517578125, -155.72470092773438, 1.2814788818359375, 167.64431762695312, 3.8931884765625, 147.21359252929688, 118.12242126464844, 29.74109649658203, 32.8803596496582, 141.23619079589844, -6.304828643798828, 104.3619155883789, 29.345535278320312, 50.490020751953125, 12.368804931640625, 87.77893829345703, 66.876220703125, 44.659095764160156, 21.345733642578125, -33.58445739746094, 7.401157379150391, 121.73086547851562, 85.42803955078125, -65.61212921142578, -1.9643898010253906, 114.78579711914062, 13.929317474365234, 28.78044891357422, 19.02507781982422, 115.46349334716797, -9.154170989990234, 177.98419189453125, -34.33546447753906, 11.672653198242188, 33.10533905029297, -16.153396606445312, 114.00709533691406, 17.76932144165039, 143.93788146972656, -18.590068817138672, 46.339210510253906, 25.713233947753906, 24.972999572753906, 8.131099700927734, -149.65310668945312, -41.72007751464844, 13.075225830078125, -21.428646087646484, 33.11297607421875, 3.912364959716797, 21.28443145751953, -6.611564636230469, -135.80722045898438, 1.5906562805175781, 46.483924865722656], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000378.npy"}
{"epoch": 0.5714285714285714, "step": 379, "batch_size": 64, "mean": 46.461585998535156, "std": 69.07056427001953, "min": -127.2374496459961, "p10": -39.16366271972655, "median": 38.2607536315918, "p90": 137.47486419677736, "max": 172.28738403320312, "pos_frac": 0.796875, "sample": [57.54833984375, 43.086822509765625, 93.85874938964844, 3.1437759399414062, 4.6864471435546875, 126.13958740234375, 33.43468475341797, -127.2374496459961, 10.409149169921875, 4.865100860595703, 2.564117431640625, 27.10235595703125, -42.65657043457031, 138.8565673828125, -2.1780872344970703, 154.27635192871094, 2.372467041015625, -46.39292907714844, 150.94192504882812, 134.03970336914062, 172.28738403320312, 8.46868896484375, 119.24933624267578, 128.02328491210938, 10.7821044921875, -17.9801025390625, 112.98222351074219, -96.60517883300781, 71.6488265991211, -31.526275634765625, 129.31605529785156, 50.62443161010742, 43.110389709472656, 47.95655059814453, -48.117889404296875, 138.77615356445312, 129.5807647705078, 22.641429901123047, 6.956733703613281, -5.169731140136719, -83.689208984375, 15.866165161132812, 71.35029602050781, 152.89588928222656, 85.1768569946289, 0.9345169067382812, 167.56475830078125, 116.35137939453125, 4.576087951660156, -18.154098510742188, 57.81995391845703, -2.8435440063476562, 1.4615020751953125, -42.43682861328125, 133.4949188232422, 0.5069446563720703, 134.4385223388672, 57.33860778808594, 70.10549926757812, 30.73407745361328, 60.50906753540039, 58.484825134277344, 115.58450317382812, 23.60455322265625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000379.npy"}
{"epoch": 0.5729402872260015, "step": 380, "batch_size": 64, "mean": 53.23714828491211, "std": 78.42414855957031, "min": -146.3389434814453, "p10": -52.965354156494136, "median": 51.69475555419922, "p90": 149.76389770507814, "max": 171.8060760498047, "pos_frac": 0.828125, "sample": [35.57213592529297, 128.79736328125, 68.68283081054688, 26.396743774414062, -137.93264770507812, -133.12367248535156, 155.02662658691406, 87.67491912841797, -1.1130294799804688, 145.67770385742188, 29.72390365600586, -79.34156799316406, -8.982320785522461, 134.6924591064453, 45.652732849121094, 1.0876846313476562, -2.7312774658203125, 57.736778259277344, 115.49285125732422, 146.90194702148438, 18.295570373535156, 97.55720520019531, 77.84020233154297, -146.3389434814453, 150.99044799804688, 5.077793121337891, 100.88398742675781, 21.253623962402344, -56.47401428222656, 8.934272766113281, 87.43990325927734, 31.889450073242188, 98.03276062011719, 110.56843566894531, 154.16964721679688, 3.985905647277832, 99.2027587890625, 134.2837371826172, 3.81170654296875, 29.423126220703125, 78.95775604248047, -119.85200500488281, 151.74334716796875, 171.8060760498047, 37.55403137207031, 144.41766357421875, -55.361167907714844, 67.29354858398438, 27.849700927734375, 142.36183166503906, 5.409881591796875, 10.310901641845703, -47.3751220703125, 152.52175903320312, 112.851318359375, 151.84674072265625, 21.99443817138672, 111.79098510742188, 133.4310302734375, 114.73814392089844, 17.758407592773438, 91.77979278564453, 4.52088737487793, 32.107574462890625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000380.npy"}
{"epoch": 0.5744520030234316, "step": 381, "batch_size": 64, "mean": 24.543724060058594, "std": 71.78839874267578, "min": -162.33448791503906, "p10": -52.96679000854492, "median": 12.4955415725708, "p90": 108.67081146240237, "max": 172.25021362304688, "pos_frac": 0.671875, "sample": [-11.111557006835938, 4.92059326171875, -3.064208984375, -54.215667724609375, -162.33448791503906, 13.399454116821289, 9.052078247070312, 25.112375259399414, 160.34835815429688, 45.077049255371094, 11.591629028320312, 1.4635086059570312, 29.597675323486328, 68.91957092285156, 89.02786254882812, 172.25021362304688, 151.07940673828125, -17.96463394165039, -7.3678436279296875, 0.5916519165039062, 26.380821228027344, -83.41206359863281, 8.14925765991211, 110.90231323242188, -50.05274200439453, 19.37384796142578, -13.388336181640625, -130.04327392578125, 93.5984878540039, 0.07986831665039062, 88.49776458740234, 102.69945526123047, 101.74870300292969, -13.02777099609375, -25.996826171875, 103.46397399902344, 80.06574249267578, -29.305503845214844, -116.57890319824219, 98.90705871582031, 122.24821472167969, -70.16180419921875, 71.11073303222656, 2.7482452392578125, 36.37919616699219, 162.23629760742188, 50.7354850769043, 65.02153778076172, -1.56475830078125, 18.330177307128906, 5.033363342285156, 48.455997467041016, 58.71983337402344, 3.97210693359375, -7.618297576904297, 33.81004333496094, 96.9012451171875, -142.86544799804688, -11.817665100097656, 26.33647918701172, 6.6865692138671875, -48.73820495605469, -9.064651489257812, 155.46868896484375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000381.npy"}
{"epoch": 0.5759637188208617, "step": 382, "batch_size": 64, "mean": 33.828857421875, "std": 63.21122741699219, "min": -88.41815948486328, "p10": -46.032841491699216, "median": 21.836727142333984, "p90": 130.48335342407228, "max": 174.75497436523438, "pos_frac": 0.703125, "sample": [-9.3582763671875, 13.294885635375977, 174.75497436523438, 43.380287170410156, 142.7661590576172, 0.26464080810546875, 25.6019287109375, -2.5309906005859375, 93.85377502441406, 122.99604797363281, 54.64436340332031, 28.64874267578125, -23.894615173339844, 8.247100830078125, 148.95297241210938, -1.215911865234375, -36.16161346435547, -50.09550476074219, -88.41815948486328, 29.192337036132812, -1.5923194885253906, 66.13623046875, 3.806549072265625, 1.362762451171875, -4.811683654785156, 5.4155120849609375, 23.379976272583008, 90.16606903076172, -60.923828125, 171.78282165527344, 90.7422866821289, 20.397064208984375, -46.664894104003906, -84.82443237304688, 145.40985107421875, -31.638484954833984, -50.90411376953125, 49.934940338134766, 33.27272033691406, 21.23206329345703, -12.259464263916016, 137.11477661132812, 7.7665557861328125, 66.83292388916016, 2.5733413696289062, 86.791259765625, -44.55805206298828, -47.93340301513672, 1.363983154296875, 7.16217041015625, 66.02691650390625, 22.441390991210938, -26.591800689697266, 38.499568939208984, 132.97723388671875, -0.4591541290283203, 106.48239135742188, 78.42240905761719, 95.94317626953125, 113.1084213256836, 124.66429901123047, 56.86962890625, 22.651412963867188, 12.55466079711914], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000382.npy"}
{"epoch": 0.5774754346182918, "step": 383, "batch_size": 64, "mean": 41.718406677246094, "std": 74.59596252441406, "min": -151.29782104492188, "p10": -39.59741058349609, "median": 20.63580322265625, "p90": 150.19256439208985, "max": 168.5812530517578, "pos_frac": 0.703125, "sample": [-0.9917221069335938, 160.4830322265625, 94.25773620605469, 7.949516296386719, -83.5400161743164, -25.07000732421875, 15.005970001220703, 19.986438751220703, 48.72845458984375, -8.754756927490234, 13.471672058105469, 28.175758361816406, 6.613945007324219, 52.61812210083008, 117.7016830444336, -27.461669921875, 27.318077087402344, -12.251178741455078, 166.43093872070312, 123.55717468261719, 9.080245971679688, -44.86982727050781, 146.8690643310547, 5.252227783203125, 145.38211059570312, -5.3023223876953125, 14.985198974609375, -13.75836181640625, 21.285167694091797, 158.49093627929688, 106.2292251586914, -39.884071350097656, -40.54176330566406, 159.74078369140625, -26.462196350097656, 60.37517166137695, 127.92266082763672, -151.29782104492188, 4.7622833251953125, 74.1141128540039, 47.391326904296875, 27.3941650390625, 15.733184814453125, -16.921470642089844, 139.2755126953125, 5.673864364624023, 67.39057922363281, 34.54447937011719, -26.665512084960938, 165.81362915039062, 56.779747009277344, 168.5812530517578, 132.07586669921875, 1.8774185180664062, -5.48419189453125, 102.11632537841797, 122.68670654296875, -54.6085205078125, 8.820846557617188, -106.78973388671875, 147.979248046875, -38.92853546142578, 151.14112854003906, 87.49879455566406], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000383.npy"}
{"epoch": 0.5789871504157218, "step": 384, "batch_size": 64, "mean": 40.22323989868164, "std": 68.19451904296875, "min": -147.60736083984375, "p10": -23.427957534790036, "median": 24.773768424987793, "p90": 132.04878692626954, "max": 170.2766571044922, "pos_frac": 0.765625, "sample": [-25.131668090820312, 169.19338989257812, 10.047523498535156, 62.412696838378906, -2.284423828125, 97.71640014648438, 57.731201171875, 91.8736343383789, 115.28404998779297, -18.155555725097656, 2.4470367431640625, 4.5214996337890625, 51.97657775878906, -5.118324279785156, 1.1494178771972656, 51.819000244140625, 27.78253936767578, 157.51156616210938, -3.939361572265625, 113.55189514160156, 53.94093322753906, 39.888031005859375, 13.13031005859375, -104.51927185058594, -35.83020782470703, 89.38749694824219, 170.2766571044922, 9.887435913085938, -117.7296371459961, 89.17654418945312, 145.17274475097656, 127.19059753417969, -0.40636253356933594, 15.454408645629883, 111.42123413085938, -19.452632904052734, 166.42291259765625, 0.5934638977050781, -147.60736083984375, 75.49607849121094, -16.456954956054688, 18.33959197998047, 117.01535034179688, -60.641361236572266, 107.21338653564453, 129.30438232421875, 18.05352783203125, 21.764997482299805, 54.33592987060547, 141.99447631835938, 3.651458740234375, -5.372035980224609, 4.955360412597656, 133.22496032714844, 5.41802978515625, 0.676605224609375, -48.19837951660156, 56.26727294921875, 5.8192138671875, 15.447097778320312, 36.31416320800781, 59.461936950683594, 53.83258819580078, 79.583251953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000384.npy"}
{"epoch": 0.5804988662131519, "step": 385, "batch_size": 64, "mean": 43.44647979736328, "std": 69.39318084716797, "min": -131.99974060058594, "p10": -36.160119628906244, "median": 26.201587677001953, "p90": 146.64451904296874, "max": 160.74212646484375, "pos_frac": 0.71875, "sample": [-131.99974060058594, -28.187515258789062, 149.13641357421875, 77.89205932617188, 39.38196563720703, 1.8472747802734375, 5.0044708251953125, 93.1352767944336, 15.907764434814453, 93.14503479003906, -74.72490692138672, 108.61070251464844, 94.69110870361328, 120.75666046142578, 21.569595336914062, 156.3509521484375, 146.54052734375, 131.81692504882812, 16.03677749633789, 52.58702087402344, -22.884971618652344, -14.85296630859375, 8.953842163085938, 89.20089721679688, 138.57180786132812, 159.94894409179688, 5.395713806152344, 11.78387451171875, -65.16062927246094, 0.7636871337890625, 26.398223876953125, -43.818782806396484, 47.18665313720703, 42.001953125, 26.00495147705078, 160.5704345703125, 71.5576171875, 68.30894470214844, 66.468994140625, 65.69896697998047, -3.040843963623047, -70.71062469482422, 61.40788269042969, 12.459732055664062, -39.57695007324219, 152.93109130859375, 88.980224609375, -1.55645751953125, -16.97735595703125, 11.584197998046875, -3.23614501953125, 39.15468215942383, -65.26300048828125, -12.28936767578125, 133.37286376953125, 11.879440307617188, 5.765625, 146.6890869140625, -4.6761627197265625, 135.31924438476562, -6.524620056152344, -6.1314239501953125, 160.74212646484375, 118.67485046386719], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000385.npy"}
{"epoch": 0.582010582010582, "step": 386, "batch_size": 64, "mean": 21.092330932617188, "std": 68.57764434814453, "min": -166.15887451171875, "p10": -55.47704238891601, "median": 11.160747528076172, "p90": 106.72933578491214, "max": 170.21690368652344, "pos_frac": 0.71875, "sample": [80.20700073242188, -28.174911499023438, -17.065540313720703, 81.74285125732422, 57.37682342529297, -31.426651000976562, 0.7105026245117188, 11.084487915039062, 8.398124694824219, 82.87811279296875, 11.237007141113281, 31.816314697265625, -131.80393981933594, -5.236351013183594, 7.376014709472656, 110.80860137939453, -0.53265380859375, 118.21070861816406, 97.43431854248047, 109.29275512695312, 7.530414581298828, 58.286956787109375, 26.16815185546875, 2.53558349609375, -78.12203979492188, 98.61015319824219, 77.06047058105469, 53.276695251464844, 15.418685913085938, 7.5558624267578125, 100.7480239868164, 17.153366088867188, -4.216407775878906, -9.573272705078125, -56.894432067871094, 17.284454345703125, -81.51319885253906, 50.77178192138672, 66.71259307861328, 167.7747802734375, 6.4465179443359375, -132.19960021972656, 8.470817565917969, 10.8380126953125, 77.43067932128906, -43.56015396118164, 4.775665283203125, 13.879127502441406, 2.068574905395508, 170.21690368652344, -7.65374755859375, -52.1697998046875, -137.20101928710938, 44.92289733886719, 151.5137939453125, 5.44989013671875, 53.27827453613281, 28.73785400390625, 40.097015380859375, 134.46768188476562, 11.93048095703125, -11.894401550292969, 5.32037353515625, -166.15887451171875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000386.npy"}
{"epoch": 0.5835222978080121, "step": 387, "batch_size": 64, "mean": 46.60486602783203, "std": 71.6932144165039, "min": -111.30252075195312, "p10": -37.82851638793945, "median": 26.693586349487305, "p90": 153.70264434814456, "max": 177.944580078125, "pos_frac": 0.703125, "sample": [15.939407348632812, -0.20957565307617188, 142.57022094726562, -46.06318664550781, 11.588951110839844, 172.89776611328125, -62.075592041015625, 11.209922790527344, 8.21575927734375, 177.944580078125, 26.19301986694336, -0.5799484252929688, 81.3930892944336, 14.642723083496094, 135.86264038085938, 5.2905731201171875, 2.8989105224609375, 149.93748474121094, 111.77006530761719, -12.38763427734375, 66.94027709960938, 160.65850830078125, 52.735008239746094, -31.09064483642578, -5.672374725341797, 115.81187438964844, -39.287132263183594, 64.38997650146484, 128.26097106933594, -11.817588806152344, 3.145263671875, 41.47631072998047, 18.210670471191406, 90.31364440917969, -34.425079345703125, 109.11212921142578, 114.70353698730469, -0.4002342224121094, 69.5338134765625, -111.30252075195312, -2.5342483520507812, 122.77549743652344, 85.97148895263672, -104.07234191894531, -40.68968963623047, 34.85792922973633, 162.40484619140625, 155.52801513671875, -4.286350250244141, 103.24637603759766, -75.82783508300781, 48.10945129394531, 18.258995056152344, 9.056472778320312, -5.857337951660156, 68.23112487792969, -1.7412948608398438, 27.19415283203125, 59.21418762207031, 171.68295288085938, 155.3162841796875, 99.54214477539062, 20.945999145507812, 127.04884338378906], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000387.npy"}
{"epoch": 0.5850340136054422, "step": 388, "batch_size": 64, "mean": 26.10688591003418, "std": 77.25288391113281, "min": -140.64707946777344, "p10": -87.9730499267578, "median": 15.120244026184082, "p90": 137.97976379394532, "max": 155.5747833251953, "pos_frac": 0.6875, "sample": [73.4273681640625, 92.10359191894531, 130.87332153320312, 139.63912963867188, -9.812034606933594, 24.876983642578125, 44.110008239746094, 10.22381591796875, 7.93267822265625, -107.55422973632812, 64.12117767333984, 86.49583435058594, -47.40987777709961, 55.33394241333008, 0.56951904296875, 139.93797302246094, -3.93505859375, 149.0526885986328, 7.834695816040039, 108.39541625976562, 132.6346435546875, 10.844100952148438, 151.8262481689453, 7.02056884765625, 134.10791015625, -95.64207458496094, 19.454666137695312, 43.70326232910156, 72.52020263671875, 7.4564361572265625, -1.3108596801757812, -106.7135009765625, 14.118768692016602, 147.0159912109375, -70.07865905761719, 5.3658294677734375, 4.597053527832031, 29.902755737304688, 16.230005264282227, -138.72340393066406, 45.144691467285156, 112.35560607910156, -28.35565948486328, 129.16001892089844, 129.8008270263672, 36.21350860595703, -57.817604064941406, 5.936103820800781, -42.96507263183594, 153.24688720703125, -4.066761016845703, 155.5747833251953, -22.787078857421875, 3.11724853515625, -129.47955322265625, -101.8929672241211, -8.32305908203125, 53.5743408203125, -30.80450439453125, 16.121719360351562, 29.207382202148438, -0.4517669677734375, -140.64707946777344, 18.431907653808594], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000388.npy"}
{"epoch": 0.5865457294028723, "step": 389, "batch_size": 64, "mean": 44.669090270996094, "std": 69.976806640625, "min": -123.76606750488281, "p10": -38.712569808959955, "median": 34.39033889770508, "p90": 142.01041412353516, "max": 160.76458740234375, "pos_frac": 0.78125, "sample": [101.04254150390625, -123.76606750488281, 5.738136291503906, 36.272705078125, 88.77322387695312, 35.30577087402344, 52.700286865234375, 141.12969970703125, 142.3878631591797, 108.94003295898438, -55.62762451171875, -110.13088989257812, 33.47490692138672, 87.93243408203125, 14.083663940429688, 107.50814819335938, 29.627716064453125, 154.46649169921875, 18.769378662109375, 37.824623107910156, 120.58488464355469, 105.90350341796875, 0.20339202880859375, 9.100608825683594, 136.41986083984375, 13.555526733398438, 132.34759521484375, -3.726520538330078, 53.992889404296875, 49.05815124511719, -53.46516418457031, -16.713226318359375, -1.44256591796875, 159.7908935546875, 160.76458740234375, 5.73931884765625, 83.46226501464844, 4.667778015136719, 1.4725532531738281, -101.2751693725586, 78.6591796875, -13.18624496459961, 7.133087158203125, -29.641983032226562, 17.131267547607422, 27.322647094726562, 129.63543701171875, -42.5999641418457, -21.02458953857422, 79.38502502441406, 4.857269287109375, 7.260524749755859, 86.62762451171875, -7.018847465515137, -54.45466232299805, 156.26075744628906, 159.67152404785156, 106.02584838867188, 12.583526611328125, 143.20578002929688, 0.3520393371582031, 123.6805419921875, 58.963775634765625, 61.09794998168945], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000389.npy"}
{"epoch": 0.5880574452003023, "step": 390, "batch_size": 64, "mean": 42.60907745361328, "std": 65.4664077758789, "min": -126.37174224853516, "p10": -24.08834686279296, "median": 24.702659606933594, "p90": 136.03032531738285, "max": 181.17831420898438, "pos_frac": 0.75, "sample": [-27.905303955078125, 30.480695724487305, 105.25652313232422, -10.902999877929688, 112.91923522949219, 8.2884521484375, 68.6618423461914, 110.68309020996094, -13.443046569824219, 66.7049331665039, 3.658203125, -36.39019775390625, 65.4921646118164, 9.431941986083984, 15.171791076660156, 181.17831420898438, 108.11001586914062, -46.77643585205078, 110.15872192382812, -39.08184814453125, 129.08135986328125, 84.73690795898438, 6.355762481689453, 141.87850952148438, 66.65484619140625, 124.75192260742188, -3.63818359375, -126.37174224853516, 23.229393005371094, 159.24208068847656, -15.182113647460938, 58.567779541015625, 17.47962188720703, 12.365127563476562, -5.92170524597168, 81.29998016357422, -123.47016906738281, 103.91104125976562, 149.25363159179688, 0.0066623687744140625, 155.33924865722656, -2.6966896057128906, -0.20618438720703125, 33.28129196166992, 0.3444404602050781, 64.61869812011719, 17.73605728149414, 1.0810012817382812, 165.60665893554688, 26.175926208496094, 49.19367218017578, 68.47460174560547, -0.5615272521972656, 63.271697998046875, 6.680450439453125, 17.959617614746094, 4.111566543579102, -39.32754898071289, -0.13677215576171875, 6.9646759033203125, 107.9510726928711, 86.63490295410156, 49.548927307128906, 139.00845336914062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000390.npy"}
{"epoch": 0.5895691609977324, "step": 391, "batch_size": 64, "mean": 57.90020751953125, "std": 80.36710357666016, "min": -153.61778259277344, "p10": -28.846365737915036, "median": 57.767377853393555, "p90": 152.2497329711914, "max": 195.86831665039062, "pos_frac": 0.765625, "sample": [38.583984375, -153.61778259277344, 152.77728271484375, 32.81610107421875, 91.29070281982422, 146.0406036376953, 5.897514343261719, 165.54058837890625, 113.48499298095703, 185.27676391601562, 45.913238525390625, -127.01924133300781, 111.96980285644531, -0.8762931823730469, 94.41751098632812, 156.68316650390625, -52.36461639404297, 118.04505157470703, -24.7618408203125, 5.839691162109375, -27.38365936279297, 124.68247985839844, 11.46978759765625, 0.5232772827148438, 63.446868896484375, -22.05782699584961, -123.919677734375, -8.594781875610352, 50.97438049316406, 1.7318458557128906, 195.86831665039062, 57.58995056152344, 72.05007934570312, -5.895729064941406, 14.091033935546875, -6.887947082519531, 13.40336799621582, 116.696044921875, 142.40798950195312, 100.04451751708984, 86.44757080078125, 99.8357925415039, -29.47323989868164, -50.54133605957031, 22.074886322021484, 149.0390625, 166.30999755859375, 44.57914733886719, 145.95025634765625, 138.1709747314453, 7.1707000732421875, -20.275184631347656, 164.64910888671875, 151.01878356933594, 151.01124572753906, 137.56800842285156, 129.5704345703125, 57.94480514526367, -54.96507263183594, 2.0171451568603516, 32.43225860595703, 136.99758911132812, 97.16015625, 64.74282836914062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000391.npy"}
{"epoch": 0.5910808767951625, "step": 392, "batch_size": 64, "mean": 56.08826446533203, "std": 75.27732849121094, "min": -147.0717010498047, "p10": -16.130133056640624, "median": 46.314964294433594, "p90": 159.91156158447265, "max": 184.85903930664062, "pos_frac": 0.78125, "sample": [-120.68806457519531, 20.272756576538086, 116.64693450927734, 35.3314208984375, 14.907569885253906, -15.768951416015625, 159.6414031982422, -13.710853576660156, 94.77513122558594, 96.00577545166016, 99.62940979003906, -147.0717010498047, 96.5168228149414, 115.16629791259766, 161.29678344726562, 52.34449005126953, -2.2639083862304688, 72.6348876953125, 160.02734375, 30.440467834472656, 35.440643310546875, -15.785697937011719, 125.27960968017578, -45.87133026123047, 105.12870788574219, 167.32423400878906, 140.1181182861328, 24.361419677734375, 36.747528076171875, -43.44810485839844, -16.277748107910156, 182.80589294433594, 152.09982299804688, 9.43450927734375, -97.6478271484375, 59.454322814941406, 6.005279541015625, 58.90305709838867, 85.98277282714844, 143.26498413085938, 174.8339385986328, 163.16049194335938, -13.276721954345703, 62.32550048828125, 1.4662818908691406, 0.5531044006347656, 1.9685287475585938, 20.280975341796875, 145.68234252929688, 44.657196044921875, -1.2022209167480469, 184.85903930664062, 0.14084434509277344, 38.35498046875, -23.350128173828125, -8.252721786499023, 108.28094482421875, 47.97273254394531, 109.6865234375, 148.65821838378906, 2.965099334716797, 87.66592407226562, 36.12543487548828, 116.63815307617188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000392.npy"}
{"epoch": 0.5925925925925926, "step": 393, "batch_size": 64, "mean": 41.60166931152344, "std": 68.5902328491211, "min": -115.64421081542969, "p10": -27.048696517944332, "median": 22.03780746459961, "p90": 145.74575347900392, "max": 180.71853637695312, "pos_frac": 0.734375, "sample": [30.62665557861328, -5.435695648193359, 146.4035186767578, 52.262603759765625, 159.9892120361328, 82.86003112792969, -0.0511474609375, 54.054954528808594, 13.580482482910156, 72.26095581054688, -22.854061126708984, 2.2076492309570312, 144.21096801757812, 40.0035514831543, 4.970638275146484, -109.88481140136719, 58.014373779296875, 134.62689208984375, -28.846397399902344, 23.80176544189453, 0.32501220703125, 18.651241302490234, -58.291046142578125, 162.0567169189453, 20.273849487304688, 137.94931030273438, -3.17047119140625, 128.68511962890625, 100.18431091308594, 140.215087890625, 13.694602966308594, -61.934017181396484, 65.39198303222656, 154.40725708007812, 4.669528961181641, 60.29307556152344, -73.7340087890625, 153.97869873046875, -7.282875061035156, -5.9477386474609375, 59.74755859375, 42.930694580078125, 32.63424301147461, 6.419197082519531, -32.7275276184082, 10.814498901367188, 132.93714904785156, -2.8808212280273438, 14.19211196899414, 5.634918212890625, 164.44247436523438, 180.71853637695312, 74.57565307617188, -7.399066925048828, 15.355670928955078, 39.78177261352539, 6.436317443847656, -3.79132080078125, -115.64421081542969, -7.7718963623046875, 9.127605438232422, 120.57722473144531, 57.85356140136719, 55.324729919433594], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000393.npy"}
{"epoch": 0.5941043083900227, "step": 394, "batch_size": 64, "mean": 46.6732292175293, "std": 63.610782623291016, "min": -116.8082275390625, "p10": -16.370226287841795, "median": 35.37763214111328, "p90": 143.1432998657227, "max": 161.10635375976562, "pos_frac": 0.75, "sample": [120.96464538574219, 157.68246459960938, 159.23583984375, -14.912467956542969, 5.85467529296875, -1.155181884765625, -116.8082275390625, 62.35607147216797, 148.62796020507812, 37.575523376464844, 161.10635375976562, 60.234928131103516, 2.447509765625, -9.151420593261719, -77.71712493896484, 64.62574005126953, -10.438301086425781, -7.305063247680664, -7.458221435546875, 21.387062072753906, 151.87890625, -0.7730598449707031, 3.655773162841797, 27.126449584960938, 69.29640197753906, 134.0511932373047, 91.51702880859375, 147.0399169921875, 21.719581604003906, 44.04405212402344, -43.58358383178711, 23.390541076660156, 151.3706512451172, 78.24357604980469, 4.9355316162109375, 126.14170837402344, 7.262237548828125, 91.42132568359375, 14.20574951171875, 0.4268798828125, 76.0865478515625, -40.94223403930664, -34.29571533203125, 84.35859680175781, 78.2823486328125, 117.34043884277344, -44.010887145996094, 80.55732727050781, -3.721729278564453, 33.17974090576172, -16.994979858398438, 121.59687805175781, 56.14631652832031, 58.61425018310547, -11.990325927734375, 97.49357604980469, 27.90127182006836, 76.52102661132812, 24.19125747680664, 0.8284759521484375, 62.17695617675781, 124.5158462524414, 2.356719970703125, 116.37135314941406], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000394.npy"}
{"epoch": 0.5956160241874527, "step": 395, "batch_size": 64, "mean": 32.703468322753906, "std": 69.50038146972656, "min": -120.54385375976562, "p10": -52.13269996643066, "median": 14.836095809936523, "p90": 142.5036560058594, "max": 183.70506286621094, "pos_frac": 0.734375, "sample": [105.61196899414062, 146.3441925048828, 100.48634338378906, 1.7378425598144531, 135.2286376953125, 23.232635498046875, -0.3064727783203125, 35.30706787109375, 146.94944763183594, -32.758602142333984, 0.3630256652832031, 2.724395751953125, 122.29328918457031, 70.10833740234375, 42.14076614379883, 151.1444091796875, -18.423995971679688, 19.28990936279297, -106.54843139648438, 6.373359680175781, -8.432106018066406, 3.2074050903320312, 1.3064727783203125, 15.961433410644531, 129.1411590576172, -42.4135856628418, 13.710758209228516, 28.36994171142578, 1.02642822265625, 20.214569091796875, 19.753936767578125, 183.70506286621094, 6.4449920654296875, 4.3569488525390625, -88.30319213867188, 47.548179626464844, -120.54385375976562, 7.421875, 83.42829132080078, -12.383729934692383, 70.44927215576172, 0.24539566040039062, 162.38504028320312, 23.206829071044922, 114.80459594726562, 149.49505615234375, -30.90026092529297, 105.00152587890625, -65.81753540039062, -18.243934631347656, -0.528900146484375, 115.07449340820312, -3.2682342529296875, 0.9375076293945312, 40.35626220703125, 113.92388916015625, -56.29803466796875, 41.70363235473633, 58.92082214355469, -56.38932800292969, 4.744361877441406, -68.59083557128906, 1.3698348999023438, 145.62152099609375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000395.npy"}
{"epoch": 0.5971277399848829, "step": 396, "batch_size": 64, "mean": 36.61708068847656, "std": 59.06542205810547, "min": -156.6551971435547, "p10": -2.432638549804685, "median": 19.130146026611328, "p90": 112.67770538330079, "max": 176.0130615234375, "pos_frac": 0.84375, "sample": [111.81549072265625, 176.0130615234375, 143.1125946044922, 163.69842529296875, 99.54666900634766, 28.685714721679688, 13.401737213134766, 61.703514099121094, 17.161094665527344, 100.73597717285156, 101.85083770751953, 7.462921142578125, 76.04483795166016, -42.68898391723633, 30.770244598388672, 2.6620121002197266, 33.77593994140625, 45.69782257080078, 12.888175964355469, 3.394145965576172, 18.65203857421875, -0.12905502319335938, 173.08331298828125, -25.57968521118164, 4.174568176269531, 10.99957275390625, 13.491249084472656, -156.6551971435547, 126.4719467163086, 19.608253479003906, 17.23373794555664, -113.70375061035156, -0.026815414428710938, 111.56465911865234, 4.360809326171875, 9.813514709472656, 70.49310302734375, 9.128143310546875, 113.04722595214844, 20.838287353515625, 11.491889953613281, 57.500648498535156, 1.4268150329589844, 102.36219787597656, 8.599868774414062, 61.463165283203125, 30.860973358154297, 27.333648681640625, 39.91288757324219, -0.2277069091796875, 24.673797607421875, -3.3776092529296875, 11.105697631835938, -7.11962890625, 47.74269485473633, 78.595458984375, 11.653419494628906, 48.875858306884766, 2.2527427673339844, 26.883766174316406, 8.920761108398438, 0.21172332763671875, -6.567192077636719, 144.3192138671875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000396.npy"}
{"epoch": 0.5986394557823129, "step": 397, "batch_size": 64, "mean": 24.13546371459961, "std": 64.92161560058594, "min": -177.0604248046875, "p10": -27.555171966552727, "median": 26.488292694091797, "p90": 89.22455139160158, "max": 157.88714599609375, "pos_frac": 0.734375, "sample": [0.7111434936523438, 40.85231018066406, -94.30621337890625, 136.53659057617188, 0.8123512268066406, 76.52720642089844, -0.35578155517578125, 18.796241760253906, 73.50192260742188, 36.321197509765625, 0.7905921936035156, 53.850685119628906, 54.306488037109375, 49.493431091308594, 80.20405578613281, 86.04887390136719, -155.1669921875, -29.897125244140625, 45.901065826416016, -0.7886390686035156, 80.27134704589844, 30.25238037109375, 17.998674392700195, 137.28746032714844, -1.4733352661132812, -143.2886199951172, -15.349609375, 18.896461486816406, 56.86712646484375, 52.635276794433594, 85.06765747070312, 7.950527191162109, 32.16642761230469, 4.395378112792969, 54.785728454589844, 23.7362060546875, -5.927436828613281, 99.30986022949219, 53.25962448120117, 34.32640075683594, 64.46043395996094, 90.58555603027344, 85.3546371459961, 44.731048583984375, -177.0604248046875, 10.737892150878906, 157.88714599609375, -116.16558837890625, -8.604286193847656, 18.184776306152344, 8.310440063476562, 35.37384796142578, 29.240379333496094, 3.3276233673095703, -2.8123779296875, -2.6938095092773438, 1.2461738586425781, 2.249357223510742, -3.4788589477539062, 96.44960021972656, -72.08512878417969, -22.090614318847656, 46.49441909790039, 157.72064208984375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000397.npy"}
{"epoch": 0.600151171579743, "step": 398, "batch_size": 64, "mean": 47.874114990234375, "std": 68.07633209228516, "min": -116.15703582763672, "p10": -18.495949554443357, "median": 24.2098388671875, "p90": 150.5515289306641, "max": 188.90628051757812, "pos_frac": 0.765625, "sample": [-16.749725341796875, 43.69511413574219, 23.18505859375, 60.42558288574219, 52.65724563598633, 5.24786376953125, 171.99363708496094, 43.829986572265625, -4.636817932128906, 135.39791870117188, 95.32120513916016, 76.4252700805664, -19.24433135986328, -5.134544372558594, 141.30581665039062, -5.0067138671875, 122.89750671386719, -6.7075347900390625, 29.691452026367188, -6.399208068847656, 40.032196044921875, 39.75961685180664, 165.13839721679688, 172.4775390625, 2.1791839599609375, 113.78425598144531, 155.44149780273438, 112.11720275878906, 5.690454483032227, 16.094558715820312, 2.4723434448242188, 111.50296020507812, 23.846389770507812, 18.21765899658203, -52.03520965576172, -22.835166931152344, 1.9296417236328125, 146.93040466308594, 22.969619750976562, 9.311065673828125, -72.2898178100586, -20.000411987304688, 141.88987731933594, 79.23155975341797, 17.01324462890625, 152.1034393310547, 7.8128814697265625, 0.8378219604492188, -25.159122467041016, 75.14855194091797, 2.933624267578125, 188.90628051757812, 26.180557250976562, 5.0522918701171875, 80.90048217773438, -10.778221130371094, 140.07406616210938, -116.15703582763672, 8.948501586914062, -0.0502471923828125, 38.72125244140625, 135.59686279296875, 159.23419189453125, 24.573287963867188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000398.npy"}
{"epoch": 0.6016628873771731, "step": 399, "batch_size": 64, "mean": 49.865936279296875, "std": 72.3907470703125, "min": -123.33493041992188, "p10": -30.201172637939447, "median": 43.380897521972656, "p90": 152.67122192382814, "max": 197.1058349609375, "pos_frac": 0.78125, "sample": [134.05654907226562, 35.71409606933594, 62.441497802734375, 75.65603637695312, -123.33493041992188, 44.956329345703125, 127.27235412597656, 30.87723159790039, -68.89898681640625, -36.24017333984375, -92.67959594726562, 40.552398681640625, 19.934314727783203, -33.464027404785156, 63.643096923828125, 49.325408935546875, 62.88127899169922, 30.417236328125, 95.02682495117188, 180.8375701904297, -20.927398681640625, 3.6137847900390625, -98.96868896484375, 124.40372467041016, 163.17840576171875, 3.12109375, 18.340110778808594, 197.1058349609375, 47.808265686035156, 6.611751556396484, 1.930908203125, -6.759426116943359, 13.52813720703125, 12.65670394897461, 5.05426025390625, -12.387763977050781, 25.918930053710938, 174.7562255859375, 104.96688842773438, 113.78946685791016, 41.80546569824219, 4.233951568603516, 173.6109161376953, -12.739974975585938, 101.94194793701172, 82.7126235961914, 21.188316345214844, 63.395469665527344, -56.67230224609375, -22.587844848632812, 138.0535888671875, 146.01498413085938, 131.32907104492188, 70.96270751953125, 114.04155731201172, 155.52389526367188, 127.64599609375, 170.15232849121094, 27.272918701171875, 48.17051315307617, 47.621543884277344, -1.5512962341308594, 52.202659606933594, -9.624954223632812], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000399.npy"}
{"epoch": 0.6031746031746031, "step": 400, "batch_size": 64, "mean": 27.590740203857422, "std": 67.0315170288086, "min": -126.37939453125, "p10": -54.867639923095695, "median": 12.332754135131836, "p90": 111.53307113647462, "max": 172.49078369140625, "pos_frac": 0.640625, "sample": [1.630126953125, 69.58027648925781, -31.705387115478516, 23.903167724609375, 34.35420227050781, 71.20176696777344, 6.0446319580078125, 101.83663177490234, 108.28439331054688, 106.09175109863281, -6.757942199707031, -11.520431518554688, 26.200584411621094, 0.7269783020019531, 24.239381790161133, 13.546600341796875, -66.16183471679688, 159.77114868164062, 9.948455810546875, 87.59473419189453, 11.08489990234375, -31.407699584960938, 99.24113464355469, -71.08589935302734, -62.74761962890625, 65.41260528564453, -34.399986267089844, 20.319744110107422, 6.5610809326171875, -0.0019378662109375, -40.14311218261719, -2.487150192260742, 81.88787841796875, -1.5640106201171875, 168.68206787109375, -83.67041015625, -48.243247985839844, -9.971481323242188, 133.34011840820312, 22.116867065429688, 56.811683654785156, 81.18428039550781, -126.37939453125, 87.00201416015625, -10.264961242675781, 102.22615814208984, 8.052993774414062, -15.379135131835938, 0.8046722412109375, 167.12562561035156, 40.885040283203125, 88.95405578613281, -5.248531341552734, 35.58880615234375, -10.62765884399414, 123.881591796875, 172.49078369140625, 48.66975402832031, 11.118907928466797, -57.7066650390625, -114.488525390625, -7.83740234375, 112.92536163330078, 24.284881591796875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000400.npy"}
{"epoch": 0.6046863189720333, "step": 401, "batch_size": 64, "mean": 53.11298370361328, "std": 70.47089385986328, "min": -119.34324645996094, "p10": -10.628409194946288, "median": 37.30547332763672, "p90": 152.37789611816407, "max": 195.80889892578125, "pos_frac": 0.859375, "sample": [-11.27703857421875, 120.34688568115234, 163.13400268554688, 72.94813537597656, 1.2299919128417969, 1.8848457336425781, 17.250038146972656, 23.568096160888672, 89.14359283447266, -119.34324645996094, 133.0428009033203, 103.57591247558594, -23.392425537109375, -7.4749908447265625, 110.98914337158203, 7.183570861816406, 33.2864990234375, 18.369384765625, 136.62747192382812, 148.7901611328125, 37.12615966796875, -9.114940643310547, 36.401641845703125, 9.325187683105469, -20.18780517578125, -116.85639190673828, 44.19346618652344, 62.792457580566406, 60.98097229003906, 6.704071044921875, 89.74986267089844, 41.793212890625, 27.766508102416992, 130.73190307617188, 1.7423934936523438, 161.44638061523438, 0.299041748046875, 118.59606170654297, 10.19828987121582, 43.00483703613281, 19.33654022216797, 153.91549682617188, 19.280975341796875, -114.42286682128906, 146.50369262695312, 25.652820587158203, 42.63677978515625, 169.02142333984375, 73.10255432128906, 90.03064727783203, 37.48478698730469, 3.6036338806152344, 84.09417724609375, 6.836090087890625, -48.997222900390625, 32.334007263183594, 113.2734375, 86.75912475585938, 159.291015625, 168.11752319335938, 140.08090209960938, 2.109363555908203, 36.800872802734375, 195.80889892578125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000401.npy"}
{"epoch": 0.6061980347694633, "step": 402, "batch_size": 64, "mean": 44.44822692871094, "std": 59.822471618652344, "min": -71.16886901855469, "p10": -23.161166381835933, "median": 34.77192306518555, "p90": 128.02275695800785, "max": 181.095947265625, "pos_frac": 0.75, "sample": [69.73085021972656, 119.95263671875, 18.039085388183594, 118.18537139892578, 49.420806884765625, 93.7263412475586, 178.55291748046875, 70.90646362304688, 62.81102752685547, 70.99979400634766, -2.477428436279297, 166.21856689453125, 33.8272590637207, -6.3148651123046875, 41.574432373046875, 35.71658706665039, 53.38593292236328, 27.20897674560547, -9.090484619140625, 12.819419860839844, 150.06475830078125, 2.607898712158203, -13.555805206298828, -71.16886901855469, 106.25279235839844, -2.8121604919433594, 87.97540283203125, 115.75106811523438, 2.4442825317382812, 11.177497863769531, 181.095947265625, 156.04588317871094, -7.186553955078125, 43.93424987792969, 11.536605834960938, 84.62457275390625, 52.45854949951172, 135.83297729492188, 18.19137191772461, 60.60334014892578, -38.358272552490234, -57.0965576171875, -47.391510009765625, 101.35240173339844, 16.045135498046875, 50.088600158691406, 10.674041748046875, -18.44451904296875, 55.96321105957031, -11.524703979492188, 130.99822998046875, 81.40986633300781, 53.50977325439453, 121.07998657226562, -33.062408447265625, 5.855278015136719, -4.67315673828125, -25.182586669921875, -42.62721252441406, 79.83837890625, 33.31047058105469, 23.57657241821289, 21.092613220214844, 7.185302734375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000402.npy"}
{"epoch": 0.6077097505668935, "step": 403, "batch_size": 64, "mean": 50.972496032714844, "std": 76.6651840209961, "min": -152.58181762695312, "p10": -39.21223831176756, "median": 31.790346145629883, "p90": 163.32159576416018, "max": 202.95794677734375, "pos_frac": 0.78125, "sample": [-1.9273605346679688, 16.052886962890625, 157.4320526123047, 8.986976623535156, 59.00678253173828, 46.12406921386719, 180.47726440429688, 104.62422943115234, 15.947235107421875, 140.40988159179688, 89.90130615234375, -12.436004638671875, 111.53211975097656, 10.336311340332031, 121.34600830078125, 55.061912536621094, 24.993762969970703, 9.822084426879883, 11.625747680664062, 21.170433044433594, -1.8750724792480469, -69.32562255859375, 202.95794677734375, 150.0435333251953, 58.1575927734375, 57.13988494873047, 178.4791717529297, -1.1788597106933594, 171.47299194335938, 6.564674377441406, 13.329170227050781, 10.313667297363281, 29.485721588134766, 7.720142364501953, 76.23339080810547, -47.325286865234375, 34.302635192871094, 171.59666442871094, 1.9577274322509766, 34.094970703125, 7.115577697753906, -1.8914604187011719, 166.05014038085938, -20.28179168701172, 113.74078369140625, -63.649993896484375, 52.14244842529297, 110.01807403564453, 36.577110290527344, -61.66416931152344, 100.83705139160156, 151.31637573242188, -152.58181762695312, -91.01734924316406, 137.24295043945312, -54.38578796386719, 161.15980529785156, 87.49665832519531, 13.265922546386719, 164.24807739257812, -2.4554672241210938, 6.7790985107421875, 11.597763061523438, 135.9470672607422], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000403.npy"}
{"epoch": 0.6092214663643235, "step": 404, "batch_size": 64, "mean": 51.221717834472656, "std": 75.42829895019531, "min": -127.4302978515625, "p10": -24.2944278717041, "median": 38.67545127868652, "p90": 154.11395416259768, "max": 189.1104736328125, "pos_frac": 0.75, "sample": [99.77101135253906, 47.12244415283203, 147.32989501953125, 16.305694580078125, 31.49483871459961, 49.0067138671875, 103.59561157226562, 24.61863899230957, -24.111392974853516, 11.484710693359375, 130.39340209960938, -13.833526611328125, -27.281673431396484, 60.37083053588867, -0.2794322967529297, 179.660400390625, -44.282691955566406, -4.611476898193359, 138.03675842285156, 146.84844970703125, 144.95989990234375, 59.59022521972656, 21.70257568359375, 167.94813537597656, -98.29171752929688, 43.7680549621582, -63.35760498046875, 109.01315307617188, 107.77874755859375, 115.74496459960938, 11.490806579589844, 50.93623352050781, -3.064014434814453, 121.15426635742188, 10.1075439453125, 6.098655700683594, -20.442832946777344, 130.17935180664062, 75.0379638671875, -2.6565933227539062, -12.9661865234375, 10.131294250488281, 139.66136169433594, 5.328582763671875, 97.72584533691406, 33.450645446777344, 96.01345825195312, 145.8934326171875, 172.99359130859375, 37.4078483581543, 4.851982116699219, 161.85733032226562, 189.1104736328125, 14.935943603515625, 160.7257080078125, 46.54150390625, 157.0214080810547, -116.40802764892578, -22.223190307617188, -127.4302978515625, -24.37287139892578, 0.915496826171875, 39.94305419921875, 7.744270324707031], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000404.npy"}
{"epoch": 0.6107331821617535, "step": 405, "batch_size": 64, "mean": 34.009098052978516, "std": 65.94513702392578, "min": -61.139892578125, "p10": -37.99359283447266, "median": 16.25048542022705, "p90": 149.30383453369143, "max": 172.97235107421875, "pos_frac": 0.640625, "sample": [158.37887573242188, 15.585182189941406, -36.69612121582031, 144.72105407714844, 40.384002685546875, -23.831832885742188, 122.64065551757812, -38.549652099609375, -48.623138427734375, -60.32167053222656, -35.931884765625, 5.122156143188477, 3.7846336364746094, -7.922554016113281, 61.411827087402344, -8.841472625732422, 170.31077575683594, -51.861724853515625, 31.496299743652344, 56.727996826171875, 14.002155303955078, 99.23397827148438, 4.977973937988281, 51.26811981201172, -22.60930633544922, -6.767589569091797, 40.84503173828125, 151.26788330078125, -8.52707290649414, 12.799356460571289, 20.36577606201172, -45.628990173339844, 15.969444274902344, 97.17301940917969, -23.884368896484375, 88.78962707519531, 21.915916442871094, -57.435367584228516, 132.3020477294922, 118.16873931884766, 0.9482460021972656, -35.918399810791016, 30.770050048828125, 75.65242004394531, 41.712242126464844, 16.531526565551758, -21.687744140625, 41.67491149902344, 171.54295349121094, 49.36417007446289, 0.1469268798828125, -3.381927490234375, 67.7969970703125, 56.83354568481445, 159.64483642578125, 172.97235107421875, -61.139892578125, 64.11336517333984, -6.8624420166015625, -13.351268768310547, -5.7139434814453125, 165.21627807617188, 36.86436462402344, -29.357154846191406], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000405.npy"}
{"epoch": 0.6122448979591837, "step": 406, "batch_size": 64, "mean": 51.688690185546875, "std": 72.72677612304688, "min": -153.5135498046875, "p10": -27.411210250854484, "median": 38.336137771606445, "p90": 154.15925598144534, "max": 171.84295654296875, "pos_frac": 0.8125, "sample": [120.21634674072266, 3.08880615234375, -8.928340911865234, -35.455322265625, 107.96170043945312, 92.86897277832031, 136.46925354003906, 148.37744140625, -45.187400817871094, 166.32012939453125, 125.30933380126953, 106.35743713378906, -5.597991943359375, -20.651432037353516, 166.27407836914062, 94.61520385742188, 168.35183715820312, 0.973602294921875, 21.164936065673828, 163.3655548095703, 162.428955078125, 47.14657211303711, 18.652488708496094, 12.419418334960938, 15.290580749511719, 40.62129211425781, 8.614151000976562, 94.14490509033203, 3.372896194458008, 29.277873992919922, -35.16424560546875, -6.312839508056641, -153.5135498046875, 4.954578399658203, 131.57762145996094, 121.27323913574219, 3.271991729736328, 146.1613311767578, 23.796398162841797, 171.84295654296875, 54.56205749511719, -30.308258056640625, 47.80897521972656, 120.11070251464844, 3.880584716796875, 108.02452087402344, -0.8533782958984375, 11.487701416015625, 34.91196823120117, 156.63717651367188, 25.01311492919922, 4.4351348876953125, 57.802490234375, -55.94767761230469, 38.74637222290039, -133.11338806152344, 37.9259033203125, 23.5372314453125, 8.328239440917969, 43.638938903808594, 121.33981323242188, 147.55169677734375, 64.26890563964844, 72.53663635253906], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000406.npy"}
{"epoch": 0.6137566137566137, "step": 407, "batch_size": 64, "mean": 39.68103790283203, "std": 63.19548416137695, "min": -98.56729125976562, "p10": -19.02438812255859, "median": 22.9208984375, "p90": 144.69481658935547, "max": 182.1561279296875, "pos_frac": 0.6875, "sample": [-0.18780136108398438, -1.629119873046875, -51.245452880859375, 77.94967651367188, -17.174209594726562, 117.99693298339844, 182.1561279296875, -9.264457702636719, -49.304710388183594, 91.2636489868164, 2.4024658203125, 40.68769836425781, 5.897918701171875, -11.044410705566406, -46.73976135253906, -4.622711181640625, 79.70709228515625, 9.768539428710938, 162.2802734375, 154.21224975585938, 35.37443542480469, -13.521257400512695, 2.6836280822753906, 48.216224670410156, -6.24395751953125, 141.58580017089844, -0.7080612182617188, 129.3284912109375, 81.13778686523438, 177.19961547851562, 80.70782470703125, -0.8550491333007812, 148.3504638671875, 96.0723876953125, 0.05810546875, 17.266563415527344, -7.702850341796875, 36.745635986328125, 67.04963684082031, 2.2952747344970703, -29.9022216796875, 31.536422729492188, 129.36331176757812, 26.982276916503906, 162.4678192138672, 74.75994873046875, 74.88465881347656, 18.859519958496094, 32.12257385253906, 1.0983352661132812, 41.110145568847656, -19.81732177734375, -98.56729125976562, -5.829010009765625, -44.37708282470703, 3.091033935546875, 63.8364372253418, 146.02725219726562, 39.61863327026367, 1.289947509765625, -2.5147628784179688, 43.861854553222656, 71.9037094116211, 9.62957763671875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000407.npy"}
{"epoch": 0.6152683295540439, "step": 408, "batch_size": 64, "mean": 53.51191329956055, "std": 65.25215911865234, "min": -104.60901641845703, "p10": -12.644281768798827, "median": 42.88787651062012, "p90": 157.61892395019532, "max": 178.04754638671875, "pos_frac": 0.78125, "sample": [-11.639724731445312, 80.77439880371094, 62.182464599609375, 79.52265167236328, 0.82568359375, -34.813194274902344, -104.60901641845703, -11.433311462402344, 141.65943908691406, 44.146080017089844, 26.04449462890625, 168.574951171875, 3.0575103759765625, 31.640716552734375, 66.83137512207031, 166.78985595703125, -13.074806213378906, -28.14466094970703, 48.84101104736328, 42.4981803894043, 33.97993469238281, -0.4724578857421875, -0.44469261169433594, 4.755302429199219, 167.04803466796875, -39.166259765625, 14.055137634277344, 172.97607421875, 106.74829864501953, 125.51592254638672, 159.7022247314453, 4.324640274047852, 80.25566101074219, 95.3387451171875, 61.4810791015625, 148.95474243164062, 26.901329040527344, -7.1910400390625, 151.2801971435547, 24.535064697265625, 13.618278503417969, 147.15928649902344, 152.7578887939453, 164.30303955078125, 0.76776123046875, 24.488609313964844, 57.18352508544922, -6.895225524902344, 46.98124313354492, 15.902236938476562, 114.40251159667969, 59.58367156982422, 178.04754638671875, -25.503114700317383, 70.33409118652344, 110.259765625, 19.91907501220703, -0.5639457702636719, 10.500129699707031, -17.447250366210938, 44.61004638671875, 43.27757263183594, 20.480148315429688, 90.34346008300781], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000408.npy"}
{"epoch": 0.6167800453514739, "step": 409, "batch_size": 64, "mean": 28.853759765625, "std": 72.2148666381836, "min": -131.33775329589844, "p10": -50.104804229736324, "median": 10.807855606079102, "p90": 146.8946304321289, "max": 175.87820434570312, "pos_frac": 0.625, "sample": [49.564910888671875, 38.260616302490234, -78.09420776367188, -23.49148178100586, -5.960025787353516, 13.520034790039062, 36.850318908691406, -5.688232421875, 53.67051696777344, -56.61503601074219, -12.878641128540039, 160.6396484375, 4.530055999755859, 159.8439483642578, 99.45927429199219, -15.608184814453125, 7.518379211425781, 10.83475112915039, -99.32023620605469, 158.7231903076172, 104.28653717041016, 10.780960083007812, 60.52851867675781, 161.87147521972656, 5.532161712646484, -51.06298828125, 0.065521240234375, 136.28900146484375, 20.9169921875, 12.649002075195312, -9.817047119140625, -25.38988494873047, 116.11640930175781, -19.404743194580078, 50.18711853027344, -9.877517700195312, 133.8535614013672, 147.1653289794922, -64.65447998046875, 10.92081069946289, -41.88222885131836, 18.330116271972656, 7.2662200927734375, -47.869041442871094, 69.6418685913086, -3.6157684326171875, -26.715866088867188, 11.04568099975586, -72.9468765258789, 169.07296752929688, -7.905422210693359, 146.26300048828125, -15.29783821105957, 132.9929656982422, -131.33775329589844, 22.10663604736328, 7.196197509765625, -35.99272537231445, 4.171241760253906, 83.65794372558594, 175.87820434570312, -6.796192169189453, 17.116668701171875, 85.54437255859375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000409.npy"}
{"epoch": 0.618291761148904, "step": 410, "batch_size": 64, "mean": 54.710350036621094, "std": 68.4581069946289, "min": -94.4690933227539, "p10": -32.928542327880855, "median": 42.543338775634766, "p90": 157.04564056396487, "max": 173.6099853515625, "pos_frac": 0.75, "sample": [81.10417175292969, 43.606266021728516, 28.634235382080078, 8.03415298461914, 164.6459503173828, 11.435821533203125, 46.45204162597656, 50.884849548339844, -18.579864501953125, -94.4690933227539, 96.22566223144531, 127.59563446044922, -1.1403427124023438, -49.08097839355469, 150.42874145507812, 100.41082000732422, 159.88145446777344, -26.672073364257812, 105.00691986083984, 149.46481323242188, 148.14083862304688, 140.64268493652344, -42.630523681640625, 91.60237121582031, 17.336105346679688, 139.32699584960938, -3.5264434814453125, -35.609886169433594, 30.4359130859375, -53.243980407714844, 160.1466064453125, -15.65570068359375, 14.815242767333984, 125.53477478027344, 11.341835021972656, 138.20025634765625, 52.169151306152344, -41.753726959228516, -0.6926765441894531, 121.15663146972656, -5.148277282714844, -47.36266326904297, 99.91584777832031, 46.534027099609375, 163.095947265625, 61.99549102783203, 161.17356872558594, 7.942192077636719, 43.0297966003418, 56.84783935546875, 12.512451171875, 39.58795166015625, 173.6099853515625, 41.93376159667969, -3.7240066528320312, 99.18849182128906, 111.73786163330078, 21.288597106933594, 41.009918212890625, 42.056880950927734, 162.72482299804688, 6.447120666503906, -2.2541122436523438, 35.713409423828125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000410.npy"}
{"epoch": 0.6198034769463341, "step": 411, "batch_size": 64, "mean": 56.17717361450195, "std": 68.1298599243164, "min": -54.850059509277344, "p10": -21.390286254882803, "median": 33.366416931152344, "p90": 157.80155334472656, "max": 174.60003662109375, "pos_frac": 0.734375, "sample": [-54.850059509277344, 134.39419555664062, 171.7139434814453, 116.8584213256836, 93.69249725341797, 21.43396759033203, -4.521270751953125, -24.84783935546875, 36.306549072265625, 3.3676376342773438, 157.99005126953125, 172.43826293945312, 82.42082214355469, 118.39832305908203, 113.98423767089844, 48.300537109375, 61.134735107421875, -44.75239562988281, 104.46192169189453, 5.974842071533203, -3.77496337890625, 93.31997680664062, 97.22299194335938, 10.220354080200195, -0.8932113647460938, 156.1742401123047, -0.24392318725585938, 78.09625244140625, -7.454154968261719, -33.11228942871094, -13.322662353515625, -5.546619415283203, -39.08244323730469, 23.893417358398438, 148.91000366210938, 13.829360961914062, 10.358051300048828, 163.8983154296875, -39.731990814208984, 163.75997924804688, 136.45361328125, 82.24787902832031, -10.011360168457031, 157.36172485351562, 96.4395980834961, 170.6888885498047, -11.853874206542969, 30.426284790039062, 4.444938659667969, 7.3237152099609375, 69.25792694091797, 41.174354553222656, 146.70965576171875, 137.95742797851562, -28.723979949951172, 109.22459411621094, 20.644893646240234, 66.60127258300781, 174.60003662109375, 14.476181030273438, 23.75627899169922, 17.447500228881836, -1.3849029541015625, 9.656417846679688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000411.npy"}
{"epoch": 0.6213151927437641, "step": 412, "batch_size": 64, "mean": 46.5667724609375, "std": 66.5781021118164, "min": -101.8561782836914, "p10": -23.397585105895992, "median": 33.2205696105957, "p90": 134.52478485107423, "max": 177.80699157714844, "pos_frac": 0.71875, "sample": [-55.68217468261719, 168.15487670898438, 70.87966918945312, 169.86257934570312, 13.34942626953125, 102.06900024414062, -1.91180419921875, 112.83668518066406, -75.51787567138672, 28.12238121032715, 26.178409576416016, 14.173381805419922, 93.64906311035156, 173.99176025390625, -0.3817424774169922, 14.56005859375, 125.85958862304688, -10.892370223999023, 24.01549530029297, -0.3855018615722656, 60.185726165771484, 134.9044952392578, 39.060638427734375, 116.5721435546875, 91.42520141601562, 1.533447265625, -4.2768707275390625, 2.109699249267578, 42.22315979003906, 45.983978271484375, -2.5582504272460938, 123.52804565429688, 39.63465118408203, 177.80699157714844, -28.066837310791016, 72.83038330078125, -9.245010375976562, 21.75098419189453, 133.6387939453125, -0.5412445068359375, -75.08917999267578, -1.97344970703125, 59.73181915283203, -20.32604217529297, 103.86175537109375, 97.813720703125, 136.55267333984375, 111.7802505493164, 171.66798400878906, -46.727638244628906, 111.90055847167969, 70.8568344116211, 38.130577087402344, -2.492870330810547, 28.310562133789062, 4.578315734863281, 0.21559524536132812, 9.06472396850586, -101.8561782836914, 61.96722412109375, 128.78392028808594, -24.713960647583008, 66.6561508178711, 0.1490020751953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000412.npy"}
{"epoch": 0.6228269085411943, "step": 413, "batch_size": 64, "mean": 36.05672073364258, "std": 76.83451843261719, "min": -143.79656982421875, "p10": -59.60893402099608, "median": 24.024526596069336, "p90": 158.31755218505864, "max": 179.93592834472656, "pos_frac": 0.765625, "sample": [1.0184402465820312, -130.1304931640625, 43.036293029785156, 85.24653625488281, -21.56640625, 19.872241973876953, 27.011390686035156, 179.93592834472656, -24.44763946533203, 177.18829345703125, -143.79656982421875, 109.46578979492188, 17.52978515625, -8.551445007324219, -2.3555755615234375, 7.700260162353516, 50.660438537597656, 164.2459259033203, 4.12322998046875, 1.666696548461914, 172.0911407470703, 75.5145034790039, 8.051422119140625, 6.864776611328125, 16.076156616210938, -8.041107177734375, 168.4637908935547, 77.14350891113281, 61.567909240722656, 25.216163635253906, 91.06060791015625, 63.911128997802734, 3.0022315979003906, 51.96562194824219, 176.005615234375, -85.33036804199219, 32.47095489501953, 6.0687408447265625, 109.28057861328125, 18.66179656982422, -30.43011474609375, 135.9275360107422, 120.80870056152344, -64.2420425415039, 23.28443145751953, -80.92237091064453, -139.35992431640625, 11.473335266113281, 39.54719543457031, 30.518531799316406, -73.61609649658203, -25.5570068359375, 132.31495666503906, 0.9522590637207031, 10.421890258789062, 53.873809814453125, 13.438966751098633, 74.4279556274414, 24.76462173461914, 96.40745544433594, -48.79834747314453, 60.598445892333984, 169.41293334960938, 144.48468017578125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000413.npy"}
{"epoch": 0.6243386243386243, "step": 414, "batch_size": 64, "mean": 45.430633544921875, "std": 71.14637756347656, "min": -162.8055419921875, "p10": -33.75173873901366, "median": 32.332916259765625, "p90": 153.31213378906253, "max": 179.53465270996094, "pos_frac": 0.765625, "sample": [120.5694351196289, -5.3910675048828125, 86.71481323242188, 12.143875122070312, 52.0811767578125, 42.401817321777344, -8.237136840820312, 61.161834716796875, 50.02045440673828, 3.7876434326171875, 6.498374938964844, 166.34896850585938, 128.275634765625, -37.993736267089844, 146.191650390625, 113.822265625, -75.66075897216797, -38.88783264160156, 35.62272644042969, -23.853744506835938, 20.032154083251953, 42.192718505859375, 179.53465270996094, 2.8852386474609375, 156.38461303710938, 16.149444580078125, 13.03582763671875, 14.628585815429688, 29.043106079101562, 91.05572509765625, 92.41925048828125, -0.9971694946289062, 122.04060363769531, 20.512218475341797, 172.27340698242188, 100.52295684814453, 36.297611236572266, 15.600555419921875, 39.30645751953125, 141.8537139892578, 19.454055786132812, -61.819671630859375, 26.305330276489258, -1.5543975830078125, 138.77394104003906, 4.107505798339844, 39.09181213378906, 61.92551040649414, -162.8055419921875, 6.9206390380859375, 82.40869140625, 39.63178253173828, -78.12745666503906, -3.698728561401367, 168.32411193847656, 174.71051025390625, 20.828205108642578, -11.830581665039062, 77.71504211425781, 115.94032287597656, -60.820220947265625, -10.970619201660156, 26.29833221435547, 156.36376953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000414.npy"}
{"epoch": 0.6258503401360545, "step": 415, "batch_size": 64, "mean": 40.086055755615234, "std": 76.4102554321289, "min": -178.5013885498047, "p10": -38.38841381072997, "median": 32.51173400878906, "p90": 151.31535186767582, "max": 178.6443634033203, "pos_frac": 0.703125, "sample": [10.575920104980469, 76.97172546386719, 1.4644966125488281, 47.83454132080078, -4.39129638671875, 9.376113891601562, 155.50584411621094, 174.19541931152344, -21.32067108154297, 103.17414855957031, -18.38863754272461, -42.034576416015625, 137.78805541992188, 164.3626251220703, -47.89997863769531, 137.48721313476562, 178.6443634033203, 57.22066116333008, 101.21829223632812, -2.1149559020996094, 32.420074462890625, 64.98222351074219, 11.193283081054688, -6.986332893371582, 32.6033935546875, 141.53753662109375, -115.76473999023438, 69.20199584960938, 52.667701721191406, 62.12005615234375, -16.160003662109375, 101.309326171875, 15.963722229003906, -8.841442108154297, -140.70822143554688, 119.23870849609375, 68.84798431396484, 2.5056381225585938, 157.61981201171875, -63.20304870605469, 57.08100128173828, -10.456146240234375, -92.473388671875, -178.5013885498047, 75.97320556640625, 19.60022735595703, 23.441871643066406, -1.8822250366210938, 12.238670349121094, 0.18148040771484375, -29.880701065063477, 63.34265899658203, 42.044677734375, 81.22236633300781, -3.2693939208984375, 124.91741180419922, 0.4263648986816406, 110.00850677490234, 62.624420166015625, -4.5565032958984375, 20.038772583007812, 47.948463439941406, 175.867919921875, 169.352294921875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000415.npy"}
{"epoch": 0.6273620559334845, "step": 416, "batch_size": 64, "mean": 32.419273376464844, "std": 76.12226104736328, "min": -159.0714874267578, "p10": -57.54793663024901, "median": 13.026079177856445, "p90": 147.1485290527344, "max": 189.17037963867188, "pos_frac": 0.6875, "sample": [-2.0207366943359375, -159.0714874267578, -49.520118713378906, 9.612861633300781, 7.912792205810547, 165.81817626953125, -13.4583740234375, 40.83976745605469, 15.476608276367188, 177.22402954101562, 12.37396240234375, 116.04080200195312, 189.17037963867188, 6.765451431274414, -6.592853546142578, 19.251121520996094, -0.014225006103515625, 87.12488555908203, -153.48638916015625, 16.420455932617188, 65.87096405029297, 27.54375457763672, 5.969276428222656, 108.33981323242188, -0.8177337646484375, 12.834686279296875, 64.5177001953125, 167.71926879882812, 119.3214340209961, -40.904151916503906, 79.16622924804688, 89.28820037841797, -8.101280212402344, -60.98843002319336, 8.35699462890625, -23.758434295654297, 1.7520713806152344, -68.66085052490234, 1.6100387573242188, 13.217472076416016, 0.597900390625, 94.12345123291016, -86.57826232910156, -24.530031204223633, 19.34540557861328, -2.07476806640625, 127.43426513671875, -14.385650634765625, -6.294313430786133, 149.98573303222656, 44.35658264160156, 4.3970794677734375, 1.9356193542480469, 149.9535675048828, -73.47900390625, 89.11213684082031, -97.69920349121094, 50.486602783203125, 73.2223129272461, 61.455230712890625, 140.6034393310547, 136.57965087890625, 154.08197021484375, 40.059539794921875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000416.npy"}
{"epoch": 0.6288737717309146, "step": 417, "batch_size": 64, "mean": 52.40943145751953, "std": 79.5389633178711, "min": -181.60382080078125, "p10": -34.82677764892577, "median": 30.380229949951172, "p90": 157.44363403320312, "max": 193.07962036132812, "pos_frac": 0.75, "sample": [62.12052917480469, -0.7186813354492188, -0.06192779541015625, -59.9534912109375, 130.8166961669922, 115.16519165039062, -181.60382080078125, 169.65130615234375, 98.30709075927734, 50.56977462768555, 193.07962036132812, 12.523818969726562, 25.36398696899414, 28.30315399169922, 116.42620086669922, -131.50323486328125, 168.85458374023438, 11.944190979003906, 18.582725524902344, 17.45549774169922, 84.75103759765625, 152.6325225830078, 22.984020233154297, 22.128982543945312, 1.6600570678710938, -45.73548126220703, 174.31935119628906, 63.326324462890625, 152.8458251953125, 156.98861694335938, -24.210342407226562, 151.102294921875, 173.10943603515625, -3.7834930419921875, 101.69573974609375, -39.376678466796875, 44.132415771484375, 25.260448455810547, 162.68167114257812, 114.38186645507812, 157.63864135742188, 2.778186798095703, 112.86089324951172, 28.259971618652344, -21.048912048339844, 24.72618865966797, -8.537496566772461, 49.44603729248047, -55.49112319946289, 96.44247436523438, 11.184906005859375, 105.12164306640625, -4.0229644775390625, -2.5736541748046875, 111.755859375, -112.16710662841797, 106.45439147949219, 32.457305908203125, 16.01811981201172, -1.14178466796875, 146.74563598632812, 16.35539436340332, 116.465087890625, 88.25808715820312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000417.npy"}
{"epoch": 0.6303854875283447, "step": 418, "batch_size": 64, "mean": 26.66490936279297, "std": 69.86579132080078, "min": -130.7006072998047, "p10": -58.70176544189452, "median": 17.295734405517578, "p90": 130.2530960083008, "max": 173.93995666503906, "pos_frac": 0.65625, "sample": [71.09310150146484, 139.75125122070312, 66.6135482788086, 63.27423095703125, -38.1930046081543, 102.05525970458984, 22.464218139648438, 90.04051208496094, 173.93995666503906, -48.564697265625, 128.40203857421875, 95.77920532226562, -117.15216064453125, 23.867767333984375, 51.756126403808594, -84.18238830566406, 116.57405090332031, -14.769603729248047, 130.86758422851562, -63.19146728515625, 14.411308288574219, 170.95901489257812, -3.46002197265625, 8.938949584960938, 132.68896484375, 42.41849136352539, -27.535003662109375, 32.83275604248047, 8.188308715820312, -52.35975646972656, -61.419769287109375, -10.379776000976562, 14.906120300292969, 19.57358741760254, 166.81114196777344, -36.55889892578125, 29.684486389160156, -86.72616577148438, 4.972940444946289, -27.624008178710938, -22.264381408691406, -23.817031860351562, -14.23990249633789, -50.514678955078125, 70.78461456298828, 18.611122131347656, -5.0430145263671875, 40.27997970581055, -72.71104431152344, 2.83892822265625, 52.222259521484375, 31.769927978515625, 92.5706558227539, 130.34861755371094, 14.507057189941406, 5.232391357421875, 3.6520156860351562, 130.03021240234375, 83.77664184570312, 31.44283676147461, 52.861854553222656, -1.8129768371582031, 15.9803466796875, -130.7006072998047], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000418.npy"}
{"epoch": 0.6318972033257747, "step": 419, "batch_size": 64, "mean": 67.94660949707031, "std": 73.0134506225586, "min": -104.09735107421875, "p10": -21.39630737304687, "median": 60.7135009765625, "p90": 168.47268218994142, "max": 180.44219970703125, "pos_frac": 0.828125, "sample": [19.441452026367188, 165.1163787841797, 16.218015670776367, 143.67591857910156, -26.592788696289062, 0.39131927490234375, 156.6946258544922, 51.39278793334961, -22.858184814453125, 102.60612487792969, -104.09735107421875, 139.44566345214844, 163.26499938964844, 118.56491088867188, 26.478038787841797, 17.348281860351562, 143.83615112304688, 49.777191162109375, 166.24557495117188, 9.434226989746094, 45.663597106933594, 75.93081665039062, -11.902366638183594, 86.38362884521484, 35.86566162109375, -17.985260009765625, 172.3214569091797, 17.169219970703125, 149.9649200439453, 80.38912200927734, 14.097373962402344, -43.49028778076172, -25.68172836303711, 166.7470245361328, 143.00567626953125, 57.08734893798828, 128.61489868164062, 9.605674743652344, 0.006801605224609375, 70.58551025390625, 2.6468582153320312, 141.302978515625, 8.341133117675781, -12.353416442871094, 169.21224975585938, -24.412384033203125, 176.4862823486328, 64.29552459716797, 13.135124206542969, 169.514892578125, 174.0399627685547, 13.184455871582031, 80.67884063720703, 75.85092163085938, 113.26884460449219, 57.13147735595703, 149.97613525390625, 82.968505859375, 180.44219970703125, 8.085678100585938, 91.19103240966797, -3.7976455688476562, -45.301395416259766, 171.9322052001953], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000419.npy"}
{"epoch": 0.6334089191232048, "step": 420, "batch_size": 64, "mean": 30.088388442993164, "std": 84.7667007446289, "min": -167.29150390625, "p10": -77.03257904052734, "median": 17.657325744628906, "p90": 158.13467559814458, "max": 179.25137329101562, "pos_frac": 0.6875, "sample": [-21.46691131591797, 9.530487060546875, 163.58619689941406, -0.6367721557617188, 30.47821807861328, -135.5994110107422, -0.384185791015625, 23.85631561279297, 6.56988525390625, -167.29150390625, -18.006629943847656, 20.36640167236328, 2.7246170043945312, 145.41445922851562, -8.504501342773438, 142.14382934570312, 179.25137329101562, 16.0235595703125, 176.97364807128906, 171.2401123046875, 71.63536071777344, -36.503414154052734, 81.29383087158203, 78.60086059570312, 112.36795043945312, 133.61477661132812, -64.1139907836914, -123.85105895996094, 86.35333251953125, 26.886428833007812, -125.54088592529297, 108.95878601074219, -13.77093505859375, 29.053306579589844, -21.23456573486328, -72.94572448730469, 1.68743896484375, -7.863704681396484, 65.35672760009766, 12.6754150390625, 18.56189727783203, -122.70758056640625, -78.78408813476562, 2.565521240234375, 140.90838623046875, 48.307918548583984, 39.978675842285156, 169.162353515625, 0.5687217712402344, 15.038772583007812, 172.21218872070312, 8.010299682617188, 18.429473876953125, -36.98914337158203, 16.885177612304688, 139.4259490966797, 116.85458374023438, 88.13737487792969, 20.69195556640625, 29.86114501953125, -28.767776489257812, 174.2188720703125, -108.4205093383789, 2.577545166015625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000420.npy"}
{"epoch": 0.6349206349206349, "step": 421, "batch_size": 64, "mean": 37.125335693359375, "std": 85.96984100341797, "min": -165.208251953125, "p10": -90.6450698852539, "median": 28.373828887939453, "p90": 145.7645980834961, "max": 197.93505859375, "pos_frac": 0.640625, "sample": [-154.1243896484375, 170.01947021484375, -56.78241729736328, 128.2652130126953, 142.91390991210938, 119.41529846191406, 62.46430206298828, 14.474555969238281, 25.08800506591797, 197.93505859375, 159.099853515625, -32.14939880371094, 170.14845275878906, 11.24948501586914, -25.122264862060547, 101.89549255371094, 30.738418579101562, -12.760032653808594, 122.85053253173828, 76.44684600830078, 120.82428741455078, 143.46664428710938, 20.468795776367188, -0.2984504699707031, 26.17407989501953, -28.910987854003906, 69.5577163696289, -165.208251953125, -86.91554260253906, -104.67991638183594, 54.08765411376953, 160.97601318359375, 113.2701416015625, -1.3798580169677734, -23.447364807128906, 152.056640625, 146.7494354248047, 79.89939880371094, -23.089759826660156, 9.633037567138672, 90.4061279296875, -22.14520263671875, 113.20431518554688, -8.181854248046875, -101.29371643066406, 129.81556701660156, -0.476593017578125, 86.55213928222656, -46.29508972167969, -46.60303497314453, 87.63751220703125, 135.49951171875, -94.65364837646484, 9.71527099609375, 30.573577880859375, -92.24343872070312, 5.86419677734375, 65.93608093261719, 102.0811767578125, -107.58718872070312, 52.99047088623047, 76.5897216796875, 2.6699180603027344, -9.334278106689453], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000421.npy"}
{"epoch": 0.636432350718065, "step": 422, "batch_size": 64, "mean": 38.49900436401367, "std": 66.01041412353516, "min": -111.90613555908203, "p10": -40.07189102172851, "median": 33.021894454956055, "p90": 127.41625366210937, "max": 185.77081298828125, "pos_frac": 0.71875, "sample": [98.00906372070312, 163.22535705566406, 18.164100646972656, 32.571659088134766, 138.463134765625, 1.3767166137695312, 11.086990356445312, -15.742919921875, 12.620220184326172, 185.77081298828125, 19.23192596435547, 23.60009765625, -3.6245574951171875, 18.95001220703125, 61.00889587402344, -11.374687194824219, 72.1965103149414, 127.46220397949219, -79.32009887695312, 75.298583984375, 111.23433685302734, -15.93115234375, 3.585723876953125, -111.90613555908203, 36.00909423828125, -23.455764770507812, 54.029815673828125, 90.47305297851562, 182.63119506835938, 33.56358337402344, 127.30903625488281, 81.48402404785156, 94.109619140625, -75.09458923339844, 68.81747436523438, -48.010528564453125, 89.8310546875, -30.220413208007812, 163.3312530517578, 12.595977783203125, -7.0917816162109375, -36.481117248535156, 53.53968048095703, 55.31407928466797, 22.410491943359375, -43.850555419921875, 5.05120849609375, 67.44407653808594, -10.977493286132812, 75.23355102539062, -41.61079406738281, 180.39346313476562, 6.897279739379883, -62.899444580078125, 57.07615661621094, 64.28434753417969, 33.472129821777344, -31.1639404296875, 80.1443862915039, 52.579322814941406, 8.791522979736328, 88.84285736083984, -18.637466430664062, 71.81367492675781], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000422.npy"}
{"epoch": 0.6379440665154951, "step": 423, "batch_size": 64, "mean": 52.01233673095703, "std": 72.15631866455078, "min": -142.60430908203125, "p10": -23.68606033325195, "median": 33.71035194396973, "p90": 162.86148986816406, "max": 181.33392333984375, "pos_frac": 0.765625, "sample": [46.85022735595703, -24.20557403564453, 174.89520263671875, -6.15900993347168, 173.7061309814453, 37.1461067199707, 23.23830795288086, -53.47407531738281, 162.87017822265625, 17.211769104003906, -24.663345336914062, 162.84121704101562, -4.8334808349609375, 16.639678955078125, -45.466033935546875, 45.191993713378906, 33.20857620239258, 21.557388305664062, 178.9317169189453, -65.80722045898438, 126.34980010986328, 67.7738037109375, -7.999717712402344, 27.34416961669922, 11.808454513549805, 131.6231689453125, 168.3251953125, -3.145244598388672, -142.60430908203125, 70.839599609375, 47.244354248046875, 159.29844665527344, 29.517913818359375, 74.60935974121094, 93.58529663085938, 12.252052307128906, 14.818649291992188, 35.031097412109375, 155.42955017089844, 181.33392333984375, 9.368522644042969, 24.90969467163086, -12.851333618164062, 132.22601318359375, 78.16203308105469, 127.66880798339844, 111.89006042480469, -5.624784469604492, 9.859210968017578, 40.26887893676758, 178.1763153076172, 55.23181915283203, -42.007442474365234, 22.116392135620117, 69.75630950927734, 139.67770385742188, 1.6151351928710938, 24.366065979003906, -22.473861694335938, 73.38299560546875, 152.83859252929688, -10.117599487304688, 34.212127685546875, 13.022686004638672], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000423.npy"}
{"epoch": 0.6394557823129252, "step": 424, "batch_size": 64, "mean": 51.92509460449219, "std": 70.55184936523438, "min": -113.634033203125, "p10": -33.35658111572265, "median": 37.15378761291504, "p90": 150.82348327636723, "max": 182.22979736328125, "pos_frac": 0.828125, "sample": [-23.416839599609375, 172.93527221679688, 15.912757873535156, 155.9327392578125, 4.005836486816406, 138.6063232421875, 112.81503295898438, 28.551719665527344, -56.08478546142578, -112.55619812011719, 11.0380859375, 12.013263702392578, 24.70104217529297, -63.57972717285156, 176.50393676757812, 22.638538360595703, 94.17015075683594, -113.634033203125, -5.168342590332031, 0.266357421875, 12.430404663085938, 27.891403198242188, 55.21220397949219, -2.9156646728515625, 31.603878021240234, 138.05877685546875, 106.5631103515625, 7.180747985839844, 175.9137725830078, 42.703697204589844, 77.19818115234375, 76.55097198486328, 20.174514770507812, 44.13683319091797, 138.90188598632812, 75.715087890625, 123.04085540771484, 94.34510803222656, 131.76373291015625, 83.93405151367188, 55.40972900390625, 30.847476959228516, 2.8861331939697266, 182.22979736328125, 8.594322204589844, 168.82177734375, 56.56682586669922, 28.956424713134766, -19.469528198242188, 11.221633911132812, 135.6744842529297, 80.53117370605469, 69.01834106445312, 130.82896423339844, 10.70172119140625, -37.760780334472656, 17.60394287109375, -62.414039611816406, 99.31121826171875, 164.27294921875, 61.713096618652344, 88.80389404296875, -37.61647033691406, 20.418296813964844], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000424.npy"}
{"epoch": 0.6409674981103552, "step": 425, "batch_size": 64, "mean": 50.87408447265625, "std": 68.49068450927734, "min": -65.87556457519531, "p10": -25.518383407592772, "median": 38.99004364013672, "p90": 146.28894500732423, "max": 185.13641357421875, "pos_frac": 0.734375, "sample": [93.86791229248047, 18.180099487304688, -1.0784912109375, 12.591781616210938, 104.46815490722656, 26.036823272705078, 185.13641357421875, -45.12636184692383, 98.70570373535156, 116.35902404785156, 114.82867431640625, -17.057456970214844, 45.99835968017578, -21.818389892578125, -55.67005920410156, 94.82359313964844, 14.683364868164062, 140.050048828125, 1.1714859008789062, 44.19175720214844, -42.4947509765625, 118.52568054199219, -6.969696044921875, 28.818872451782227, 146.5401611328125, 35.58061218261719, 7.428253173828125, 170.4734344482422, 22.626850128173828, 160.21473693847656, 58.36347961425781, -25.528011322021484, 25.33672332763672, 46.636722564697266, 42.39947509765625, 13.249755859375, 4.350151062011719, 145.55262756347656, 11.629024505615234, 15.331703186035156, 96.58394622802734, -20.77025604248047, 69.13150024414062, 139.26644897460938, 97.62773895263672, 74.6503677368164, 52.95165252685547, 126.71429443359375, 71.03135681152344, 2.5752487182617188, 173.16171264648438, -2.94403076171875, -25.49591827392578, 141.79808044433594, -3.4237518310546875, -7.461208343505859, 168.94668579101562, 50.91888427734375, -65.87556457519531, 148.4502716064453, -65.03316497802734, -1.8216323852539062, 145.70277404785156, -59.15214920043945], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000425.npy"}
{"epoch": 0.6424792139077853, "step": 426, "batch_size": 64, "mean": 26.16117286682129, "std": 78.22969818115234, "min": -164.47023010253906, "p10": -46.0675495147705, "median": 12.945327758789062, "p90": 121.2317581176758, "max": 226.24172973632812, "pos_frac": 0.625, "sample": [7.4297943115234375, -2.100574493408203, 122.22837829589844, 89.34822082519531, 63.017417907714844, 151.39895629882812, -28.893783569335938, 166.03456115722656, 12.696891784667969, 97.58135986328125, -18.96527099609375, -33.603363037109375, -4.808345794677734, -71.67257690429688, 22.475616455078125, -11.009986877441406, 41.08504867553711, 5.643211364746094, 13.639228820800781, 2.8418922424316406, -156.21990966796875, 17.71484375, -50.077239990234375, 63.15375518798828, -30.09351348876953, 1.775421142578125, 0.6076126098632812, -150.4605255126953, 85.0031509399414, 24.34100341796875, 87.99794006347656, 118.90631103515625, 85.994140625, 55.85188293457031, -164.47023010253906, 25.41131591796875, -124.42259979248047, 13.193763732910156, -0.8491058349609375, 226.24172973632812, 172.93814086914062, 176.9612274169922, -36.711605072021484, 6.934906005859375, -8.39260482788086, 177.03033447265625, -9.508903503417969, -0.8179740905761719, 18.630908966064453, -31.863677978515625, -6.627178192138672, 74.20188903808594, 17.065269470214844, 62.054290771484375, 52.91612243652344, -7.80316162109375, 5.739166259765625, 35.250572204589844, -0.3694000244140625, -79.5296630859375, 110.37702941894531, 85.75334167480469, 116.33320617675781, -10.213518142700195], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000426.npy"}
{"epoch": 0.6439909297052154, "step": 427, "batch_size": 64, "mean": 42.38200378417969, "std": 72.14573669433594, "min": -151.82302856445312, "p10": -37.055070495605456, "median": 25.38286590576172, "p90": 153.0136703491211, "max": 196.3370361328125, "pos_frac": 0.71875, "sample": [17.062362670898438, -1.3015289306640625, 0.5758476257324219, -5.743217468261719, 1.0062255859375, 54.71095657348633, 173.25042724609375, -2.3984375, 65.2886962890625, -61.47599792480469, 135.14212036132812, -104.25653839111328, -51.17792510986328, -53.3359375, 61.32776641845703, 26.586273193359375, 24.179458618164062, 6.00323486328125, -0.40459442138671875, -5.270927429199219, -1.363250732421875, 82.4325942993164, 151.42718505859375, 61.12205123901367, 19.7890625, 167.56686401367188, 41.08131790161133, 83.99166870117188, 30.230918884277344, 57.023780822753906, 151.8666534423828, 31.443302154541016, 37.468963623046875, 64.57583618164062, 6.812019348144531, 102.47863006591797, -6.464143753051758, -151.82302856445312, 153.5052490234375, 169.65834045410156, 74.47663116455078, 22.459392547607422, -5.30865478515625, 3.7090721130371094, -44.097206115722656, -13.431686401367188, 139.49734497070312, 40.95558166503906, 6.3477630615234375, 0.02417755126953125, 172.56752014160156, 102.89120483398438, -0.6635971069335938, 12.899444580078125, 62.44306945800781, 12.146377563476562, 85.72207641601562, 137.94786071777344, 8.763435363769531, -22.817596435546875, 49.1818962097168, -43.15684509277344, 196.3370361328125, 180.96148681640625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000427.npy"}
{"epoch": 0.6455026455026455, "step": 428, "batch_size": 64, "mean": 51.373046875, "std": 74.03571319580078, "min": -132.8284912109375, "p10": -21.404028701782224, "median": 27.247509002685547, "p90": 171.73414611816406, "max": 182.83714294433594, "pos_frac": 0.75, "sample": [-67.44963836669922, 14.090080261230469, 174.36964416503906, -5.010066986083984, -0.6981849670410156, 10.520454406738281, 2.3810958862304688, -132.8284912109375, -22.504619598388672, -2.3259124755859375, 180.16880798339844, -51.75091552734375, 3.0302734375, -3.95916748046875, 148.93992614746094, 14.489227294921875, 182.83714294433594, 25.4221248626709, 1.6431884765625, 19.047775268554688, 116.33960723876953, 79.2275161743164, 5.4921112060546875, 6.1066436767578125, 171.87014770507812, 28.588272094726562, 32.09919357299805, -31.757400512695312, 25.3006591796875, -33.73919677734375, 23.58966064453125, -23.588539123535156, 37.95679473876953, 37.14727783203125, 44.90687561035156, 181.03643798828125, 124.59077453613281, 36.33992004394531, 9.832664489746094, -13.24319076538086, 106.57919311523438, 164.19271850585938, 136.32925415039062, 171.41680908203125, -17.08489990234375, -6.330162048339844, 78.4593505859375, 48.4139404296875, 0.7882041931152344, 33.306854248046875, 0.2667350769042969, 69.70799255371094, 111.02318572998047, -3.03326416015625, 173.67164611816406, 177.65682983398438, 170.25341796875, 97.62118530273438, 48.335113525390625, 104.67707824707031, 155.3382568359375, 110.70565795898438, 25.90674591064453, -18.835983276367188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000428.npy"}
{"epoch": 0.6470143613000756, "step": 429, "batch_size": 64, "mean": 44.453758239746094, "std": 90.14635467529297, "min": -180.37828063964844, "p10": -78.91212844848633, "median": 44.92348098754883, "p90": 150.57518005371094, "max": 191.3180694580078, "pos_frac": 0.75, "sample": [41.40534973144531, -138.9226837158203, 0.4215431213378906, 0.46087646484375, 10.657073974609375, 137.2200927734375, 75.50968170166016, 75.32471466064453, -0.7143669128417969, 57.19679260253906, -1.0998649597167969, -80.70199584960938, 66.01032257080078, 26.081790924072266, 64.61959838867188, 138.84727478027344, -135.1808624267578, 15.61077880859375, -29.092376708984375, 150.83441162109375, 116.2081298828125, -9.386550903320312, 127.03360748291016, 117.20474243164062, 171.8214111328125, 165.56295776367188, 177.17959594726562, 104.80789947509766, 43.89502716064453, -57.11644744873047, -143.83941650390625, -180.37828063964844, -5.320976257324219, 12.139945983886719, 90.26039123535156, 16.191856384277344, -151.87655639648438, 58.098350524902344, -2.9588470458984375, 55.796302795410156, 146.1797332763672, 180.47470092773438, 40.23542785644531, 149.97030639648438, 139.83802795410156, -140.84262084960938, -74.73577117919922, 66.53591918945312, 2.3619766235351562, 73.63200378417969, 42.40123748779297, 141.9066162109375, 8.862335205078125, 45.951934814453125, 89.12748718261719, 33.266395568847656, 166.30313110351562, 129.97789001464844, 191.3180694580078, 148.9421844482422, 6.811702728271484, 7.865617752075195, -8.126289367675781, 76.9711685180664], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000429.npy"}
{"epoch": 0.6485260770975056, "step": 430, "batch_size": 64, "mean": 57.99867248535156, "std": 81.04403686523438, "min": -165.86302185058594, "p10": -16.032289123535154, "median": 48.26424789428711, "p90": 173.0375732421875, "max": 182.98365783691406, "pos_frac": 0.78125, "sample": [173.7364959716797, 6.056262969970703, -16.902374267578125, 34.10350799560547, -12.136581420898438, 121.11622619628906, 13.737953186035156, 182.98365783691406, 111.90079498291016, 71.94457244873047, -96.95867919921875, 126.17667388916016, 178.8120880126953, 38.766990661621094, 173.36447143554688, 110.40469360351562, 82.63517761230469, 25.207931518554688, 179.4871368408203, 39.08099365234375, 9.439125061035156, 181.36431884765625, 115.6807632446289, 57.11363983154297, -5.578939437866211, 164.52593994140625, -12.508167266845703, 165.14266967773438, 63.58436584472656, 21.741920471191406, 58.847015380859375, -64.81059265136719, 39.41485595703125, 115.27743530273438, 2.478954315185547, 173.71202087402344, 172.27481079101562, 79.23357391357422, -95.72639465332031, 131.01007080078125, 0.1477813720703125, 117.91275787353516, 10.362205505371094, 9.465438842773438, 13.606281280517578, 73.83601379394531, 18.68505096435547, -2.8909759521484375, -165.86302185058594, -117.76095581054688, -0.8613128662109375, 133.64097595214844, -3.2087860107421875, 170.34161376953125, 9.585174560546875, 137.66773986816406, 82.1484375, 130.5790252685547, 124.25196838378906, -24.50054931640625, 1.5428848266601562, -14.002090454101562, 57.43113708496094, 34.06273651123047], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000430.npy"}
{"epoch": 0.6500377928949358, "step": 431, "batch_size": 64, "mean": 45.52803039550781, "std": 80.23342895507812, "min": -154.55642700195312, "p10": -29.77200012207031, "median": 16.93834114074707, "p90": 162.02478637695313, "max": 182.4385986328125, "pos_frac": 0.71875, "sample": [-133.19467163085938, 39.49666213989258, 58.54231262207031, 16.947025299072266, -13.034526824951172, 109.12552642822266, 49.01731872558594, 2.4531707763671875, 116.53733825683594, -27.332061767578125, 3.331615447998047, 180.1822967529297, 115.43976593017578, -21.549179077148438, 9.207717895507812, 2.6589698791503906, 161.1043701171875, -30.81768798828125, 139.31198120117188, 12.190933227539062, 13.484901428222656, -154.55642700195312, -39.13579559326172, 142.3321075439453, -120.78555297851562, 162.41925048828125, 163.31771850585938, -14.777795791625977, 114.3484878540039, -10.599296569824219, -3.5581417083740234, -1.1153755187988281, 116.97244262695312, 148.2470703125, -2.855377197265625, -12.506683349609375, 172.41563415527344, 33.91356658935547, -90.9481201171875, 74.51211547851562, -16.655014038085938, 14.163124084472656, -30.928329467773438, 142.0460968017578, 8.338584899902344, 13.559593200683594, 172.10943603515625, 71.49908447265625, 14.514228820800781, 182.4385986328125, 2.5696640014648438, 156.90887451171875, -0.581512451171875, 16.929656982421875, 30.158897399902344, 115.52896118164062, 43.682273864746094, 28.19007110595703, 3.58758544921875, 165.43136596679688, 146.73129272460938, 100.62039947509766, 35.89103698730469, 16.31639862060547], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000431.npy"}
{"epoch": 0.6515495086923658, "step": 432, "batch_size": 64, "mean": 46.849727630615234, "std": 87.57807159423828, "min": -171.3951416015625, "p10": -33.05723648071289, "median": 13.492026329040527, "p90": 167.88143768310547, "max": 186.85800170898438, "pos_frac": 0.703125, "sample": [-10.10693359375, -13.028297424316406, -61.05754089355469, -166.73751831054688, 3.4053211212158203, 39.689048767089844, 4.544918060302734, 26.662151336669922, 15.185028076171875, 71.14177703857422, 18.03508186340332, -123.58739471435547, 9.67999267578125, 9.612491607666016, 163.1196746826172, -171.3951416015625, 9.321174621582031, 77.1466064453125, 182.9688720703125, 101.357177734375, -20.242332458496094, 129.67977905273438, 168.836181640625, -6.148529052734375, 11.79902458190918, 177.86138916015625, 52.993560791015625, 174.22206115722656, 148.02911376953125, -32.709869384765625, -43.27983856201172, 1.7293739318847656, 4.093696594238281, -11.983200073242188, -3.6999053955078125, -33.20610809326172, 177.19444274902344, 155.40122985839844, 79.48199462890625, 118.7945327758789, -5.74664306640625, 158.39706420898438, 161.2022247314453, 3.0431251525878906, 165.65370178222656, 111.250244140625, 2.730722427368164, 172.36276245117188, -1.0548248291015625, 186.85800170898438, 52.48225402832031, 150.7750244140625, 0.8587150573730469, -12.157241821289062, 10.1778564453125, 148.3343048095703, -104.67762756347656, 81.82623291015625, 78.39457702636719, 6.132030487060547, -15.59075927734375, 134.60342407226562, -28.33221435546875, 106.05640411376953], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000432.npy"}
{"epoch": 0.6530612244897959, "step": 433, "batch_size": 64, "mean": 63.8476448059082, "std": 91.57515716552734, "min": -167.31268310546875, "p10": -61.65592346191403, "median": 55.16021537780762, "p90": 179.83594360351563, "max": 205.18453979492188, "pos_frac": 0.78125, "sample": [3.8646163940429688, 75.54779052734375, -135.15267944335938, 105.30786895751953, 78.97173309326172, 9.211555480957031, 2.707489013671875, 205.18453979492188, -13.62693977355957, 34.2855224609375, 5.30145263671875, 133.952392578125, 2.8402862548828125, -0.6082077026367188, 180.36654663085938, -82.83213806152344, 30.0583553314209, 170.2747802734375, 152.81124877929688, -87.07755279541016, 116.52902221679688, -0.095733642578125, 10.702194213867188, 161.2373504638672, -167.31268310546875, 94.38075256347656, 134.2632293701172, 51.3847770690918, 176.2329559326172, 117.12359619140625, 3.0193634033203125, 167.14508056640625, -105.60734558105469, 9.297378540039062, 89.84735870361328, -7.227100372314453, -79.04669952392578, 50.4375, -16.032379150390625, 191.90130615234375, 183.94674682617188, 183.64862060546875, 180.32901000976562, 50.32868957519531, 58.93565368652344, -30.360809326171875, 8.053539276123047, 137.08889770507812, -2.5149078369140625, 178.68545532226562, 7.653617858886719, 28.08972930908203, 139.92747497558594, -75.068115234375, 58.955047607421875, 164.80477905273438, 147.68154907226562, 18.98931884765625, 168.93858337402344, 119.15365600585938, 184.7010498046875, 7.328041076660156, 176.1738739013672, 121.21131134033203], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000433.npy"}
{"epoch": 0.654572940287226, "step": 434, "batch_size": 64, "mean": 59.861175537109375, "std": 86.95237731933594, "min": -162.83029174804688, "p10": -24.636490249633788, "median": 42.08165740966797, "p90": 171.43673706054688, "max": 186.65919494628906, "pos_frac": 0.71875, "sample": [-162.83029174804688, 171.96051025390625, 140.53164672851562, -1.3334503173828125, -25.569473266601562, 176.88967895507812, -93.45870971679688, 159.79037475585938, -106.90472412109375, 108.75881958007812, -10.340507507324219, -3.4550819396972656, -144.41607666015625, 169.59764099121094, 132.64328002929688, -26.228736877441406, 125.34455108642578, 109.42280578613281, 1.430633544921875, 174.3406982421875, 160.14715576171875, -1.200063705444336, 186.65919494628906, -15.433273315429688, -46.410884857177734, 21.76836395263672, 176.83941650390625, 11.363273620605469, 0.8528709411621094, 153.03280639648438, -8.321273803710938, 30.18399429321289, 74.09689331054688, 183.07293701171875, 151.03768920898438, 116.66365814208984, -4.695518493652344, 42.83721923828125, 146.33775329589844, -1.8809127807617188, 157.30116271972656, -10.051597595214844, 4.0084991455078125, 150.8262176513672, 5.5782623291015625, 15.479606628417969, 5.172784805297852, 87.75302124023438, 11.930351257324219, 121.86958312988281, 14.29214859008789, 186.31436157226562, -22.459529876708984, 29.50904083251953, 60.97086715698242, 79.13811492919922, 97.44477844238281, 170.214599609375, 129.39686584472656, -11.549503326416016, 5.448566436767578, 69.39005279541016, 158.68612670898438, 41.32609558105469], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000434.npy"}
{"epoch": 0.656084656084656, "step": 435, "batch_size": 64, "mean": 41.774715423583984, "std": 80.83851623535156, "min": -150.8821563720703, "p10": -53.35716209411621, "median": 32.26008892059326, "p90": 159.27359924316406, "max": 180.8035888671875, "pos_frac": 0.640625, "sample": [165.73678588867188, 126.650390625, -76.78041076660156, 29.301130294799805, -66.85218048095703, 85.05107879638672, -9.662750244140625, 6.974994659423828, 128.9019775390625, -33.064369201660156, 115.30577850341797, 121.09271240234375, -37.31829833984375, 50.183387756347656, 170.82000732421875, 119.99329376220703, 20.664947509765625, -18.826828002929688, -19.046249389648438, 157.69046020507812, 54.62294006347656, -5.063350677490234, 92.08773803710938, 171.76449584960938, 24.57794189453125, 1.9434852600097656, -37.93334197998047, 17.33795166015625, 166.93359375, 168.85198974609375, -3.0671768188476562, 77.60028076171875, -144.52903747558594, 127.68553924560547, 48.17167282104492, 62.018798828125, -23.693572998046875, -23.519256591796875, 73.26788330078125, 144.76478576660156, 130.2186279296875, 159.95208740234375, 180.8035888671875, 35.21904754638672, -2.276052474975586, 78.76753234863281, -54.21883010864258, -20.208723068237305, -51.34660339355469, 89.39724731445312, -67.026123046875, 9.843025207519531, -81.68598937988281, 112.26142120361328, 61.774925231933594, -11.07244873046875, 95.71359252929688, 19.39374542236328, -150.8821563720703, 39.47996520996094, 3.9001846313476562, -27.19281005859375, 112.3062744140625, -20.17877197265625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000435.npy"}
{"epoch": 0.6575963718820862, "step": 436, "batch_size": 64, "mean": 46.05265808105469, "std": 83.71385192871094, "min": -169.50624084472656, "p10": -50.79247665405273, "median": 31.187292098999023, "p90": 158.49481048583988, "max": 184.41770935058594, "pos_frac": 0.703125, "sample": [4.0605621337890625, -58.34778594970703, 161.11691284179688, 2.7386207580566406, 65.46098327636719, 0.5139827728271484, 59.27392578125, 145.93209838867188, -60.609344482421875, 31.107852935791016, 111.41664123535156, 168.14892578125, 0.34534454345703125, 75.11297607421875, 9.314460754394531, 178.368896484375, 184.41770935058594, 143.2357635498047, -129.68548583984375, 110.55339050292969, 36.53026580810547, 10.202110290527344, -4.315071105957031, -44.453399658203125, -42.24903106689453, -0.793487548828125, 144.68624877929688, 4.173088073730469, 88.62833404541016, 97.19972229003906, 63.607872009277344, -5.0735015869140625, -169.50624084472656, 61.136444091796875, -3.5539169311523438, 31.26673126220703, 115.15959167480469, -76.80241394042969, -53.50922393798828, -21.397315979003906, 152.37657165527344, 152.3724365234375, 8.709381103515625, 11.053825378417969, -7.260791778564453, 11.211454391479492, -21.12583351135254, 174.1107177734375, 78.00717163085938, 138.21774291992188, 55.72210693359375, 2.1346168518066406, 151.8808135986328, 129.71981811523438, -5.822349548339844, 33.01808166503906, 177.76451110839844, -0.6030731201171875, 13.276535034179688, 103.07682800292969, -4.3017578125, 179.43597412109375, -140.1777801513672, 121.15997314453125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000436.npy"}
{"epoch": 0.6591080876795162, "step": 437, "batch_size": 64, "mean": 33.156768798828125, "std": 84.39581298828125, "min": -158.06517028808594, "p10": -65.73488235473633, "median": 21.808414459228516, "p90": 161.49048156738286, "max": 210.37481689453125, "pos_frac": 0.609375, "sample": [210.37481689453125, 52.62092590332031, 55.13473129272461, 13.038963317871094, 70.63311767578125, -39.567169189453125, 52.660186767578125, 22.98797607421875, 184.469970703125, -4.474212646484375, 175.56527709960938, 3.2166595458984375, -153.0131378173828, 17.80974578857422, -99.54914855957031, -22.71021270751953, 15.378829956054688, 101.15914916992188, 98.31358337402344, -158.06517028808594, 121.67442321777344, 148.63458251953125, 175.73764038085938, -77.11061096191406, -9.354774475097656, 77.93468475341797, 110.201416015625, -58.87199020385742, -125.61581420898438, -61.6456298828125, -10.170951843261719, -32.69670867919922, -41.18605041503906, -67.48741912841797, 4.975376129150391, 182.67893981933594, 95.98692321777344, -32.98786163330078, 64.00651550292969, 27.030906677246094, 103.39147186279297, -1.5669898986816406, -27.640029907226562, 125.29557037353516, -10.719757080078125, 167.00015258789062, 23.4305419921875, 4.402637481689453, 91.00218200683594, -82.38493347167969, 100.89237976074219, 81.38864135742188, 50.729862213134766, 198.08364868164062, 35.765167236328125, -3.6482162475585938, 42.373512268066406, 69.50370788574219, -10.151443481445312, 20.62885284423828, -1.2259483337402344, -45.56409454345703, -1.5259552001953125, 104.85360717773438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000437.npy"}
{"epoch": 0.6606198034769464, "step": 438, "batch_size": 64, "mean": 58.298133850097656, "std": 81.83240509033203, "min": -154.30374145507812, "p10": -11.261088943481445, "median": 42.624589920043945, "p90": 168.9696304321289, "max": 187.54502868652344, "pos_frac": 0.734375, "sample": [178.24435424804688, -7.999114990234375, 42.407440185546875, -6.746192932128906, -0.044586181640625, -10.996444702148438, 29.736557006835938, 70.90184020996094, 8.471710205078125, 100.56610107421875, -4.5774078369140625, 187.54502868652344, -31.818634033203125, -3.0586471557617188, 121.05967712402344, 1.6516609191894531, 168.29574584960938, -123.50791931152344, 25.062686920166016, 126.14961242675781, 134.4456787109375, 150.3515167236328, -4.984657287597656, 3.9161834716796875, 63.63829040527344, 39.37820053100586, 75.08303833007812, 67.43951416015625, 169.25843811035156, 163.5205078125, 58.79301452636719, -3.1674957275390625, 171.30352783203125, -11.374507904052734, 11.42367172241211, 175.05429077148438, 141.0203399658203, -7.675769805908203, 42.841739654541016, 154.0808868408203, 165.2008514404297, 41.37980651855469, 73.35486602783203, 186.5252685546875, 123.19097900390625, 40.36143493652344, 117.1239013671875, 79.53447723388672, -112.91670227050781, 18.860748291015625, -30.50897216796875, -89.6387710571289, 18.574182510375977, 142.66798400878906, 131.5712127685547, -154.30374145507812, 178.07972717285156, 157.66371154785156, -2.7100448608398438, 122.57798767089844, 1.0729179382324219, 3.6967735290527344, 3.8072433471679688, 50.22489547729492], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000438.npy"}
{"epoch": 0.6621315192743764, "step": 439, "batch_size": 64, "mean": 52.50692367553711, "std": 77.48680877685547, "min": -90.875732421875, "p10": -30.34862518310547, "median": 21.362411499023438, "p90": 172.35491790771485, "max": 192.2895965576172, "pos_frac": 0.765625, "sample": [158.9676513671875, 4.028980255126953, -10.056955337524414, -39.02141571044922, 35.295310974121094, 163.9443359375, 39.64537811279297, -11.990020751953125, 7.465263366699219, 168.53526306152344, 181.61709594726562, 29.281997680664062, 155.18321228027344, 5.459503173828125, 45.027191162109375, 192.2895965576172, 173.99191284179688, 13.594696044921875, 51.5377197265625, 15.94512939453125, 4.804714202880859, 160.32896423339844, 60.089698791503906, -49.12474822998047, 110.0357894897461, 10.747222900390625, 25.527069091796875, 4.405982971191406, 11.16183090209961, 181.84494018554688, -30.652923583984375, 179.409912109375, -29.638595581054688, -84.61296081542969, 156.19264221191406, 17.19775390625, -16.90496063232422, 2.6018142700195312, -90.875732421875, -41.422767639160156, 6.957794189453125, 0.413848876953125, 106.0738296508789, 14.27410888671875, 140.15127563476562, 159.4705352783203, 107.93263244628906, 163.63931274414062, 71.88634490966797, -2.8713150024414062, 32.074798583984375, 56.139801025390625, -1.3458404541015625, 178.286865234375, 51.58363342285156, 186.04322814941406, 3.8790283203125, -15.803695678710938, 1.2389068603515625, -39.60303497314453, 12.548027038574219, 53.14753723144531, 113.91764068603516, -1.4496002197265625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000439.npy"}
{"epoch": 0.6636432350718064, "step": 440, "batch_size": 64, "mean": 64.52052307128906, "std": 76.98979949951172, "min": -106.57707214355469, "p10": -18.304828834533687, "median": 46.00212478637695, "p90": 172.57842102050782, "max": 184.0247802734375, "pos_frac": 0.8125, "sample": [33.70911407470703, 55.4886474609375, 67.51826477050781, 184.0247802734375, 183.9303436279297, 89.94610595703125, 179.9046630859375, 12.080089569091797, 172.75704956054688, 37.060516357421875, 30.580345153808594, 169.8179168701172, -92.544189453125, 155.06849670410156, 42.87391662597656, 20.06439208984375, 172.16162109375, -14.301925659179688, 61.61737823486328, 18.4008731842041, 88.63932800292969, -14.57295036315918, -2.2306365966796875, 105.83647155761719, 63.378135681152344, 144.57861328125, 152.87661743164062, 81.79978942871094, 1.9488773345947266, 25.495697021484375, 0.40303993225097656, 75.77555084228516, 77.84169006347656, 19.892425537109375, 0.2008380889892578, 18.923847198486328, 3.575878143310547, -24.43048858642578, 21.958080291748047, 170.65370178222656, 125.61270141601562, 164.77342224121094, -48.805091857910156, 155.44815063476562, -2.769378662109375, 1.239776611328125, 179.38485717773438, 36.70452880859375, 165.52603149414062, -19.904205322265625, -68.29324340820312, -106.57707214355469, 129.8424072265625, 32.153350830078125, 147.87010192871094, 98.7369384765625, -2.760986328125, -20.843935012817383, 134.30142211914062, 15.503166198730469, 13.94317626953125, 174.07220458984375, 182.32228088378906, 49.130332946777344], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000440.npy"}
{"epoch": 0.6651549508692366, "step": 441, "batch_size": 64, "mean": 53.26115036010742, "std": 76.21244812011719, "min": -147.95639038085938, "p10": -24.998490142822263, "median": 43.232486724853516, "p90": 160.27684020996094, "max": 195.98715209960938, "pos_frac": 0.765625, "sample": [88.81800842285156, 123.98387145996094, 154.47300720214844, 122.81777954101562, 35.4508171081543, -15.174545288085938, 99.94142150878906, 10.384567260742188, 57.12193298339844, 2.9522323608398438, -27.355613708496094, 45.879150390625, 28.209705352783203, 160.27935791015625, 3.2374954223632812, 117.68905639648438, 132.19979858398438, 2.5509185791015625, -25.644630432128906, 160.27096557617188, 55.765716552734375, 50.88427734375, 195.98715209960938, 149.6136474609375, -2.6314048767089844, 179.8623046875, -0.5044708251953125, 151.2190704345703, 181.5753936767578, 13.93682861328125, -26.601234436035156, 6.829017639160156, -107.15868377685547, 168.409423828125, 102.31977081298828, -0.8257827758789062, -114.73524475097656, 169.98480224609375, 50.27584457397461, 22.05573272705078, -1.4654541015625, 137.29220581054688, 14.297317504882812, 83.16768646240234, 45.199859619140625, -25.835052490234375, 41.265113830566406, 0.9460487365722656, 54.6097412109375, -23.490829467773438, 163.89093017578125, -12.809371948242188, 6.541648864746094, 159.04054260253906, -147.95639038085938, 81.13729095458984, 30.305152893066406, 10.351791381835938, 33.936134338378906, 82.55592346191406, -7.22479248046875, 110.33736419677734, 46.841796875, 1.4315223693847656], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000441.npy"}
{"epoch": 0.6666666666666666, "step": 442, "batch_size": 64, "mean": 50.065460205078125, "std": 83.45018005371094, "min": -154.88133239746094, "p10": -45.724314498901364, "median": 37.09822463989258, "p90": 156.04876556396485, "max": 193.86192321777344, "pos_frac": 0.703125, "sample": [55.970420837402344, 11.087371826171875, -12.9383544921875, -76.58181762695312, 16.97095489501953, 156.5050506591797, 193.86192321777344, 174.8502197265625, 46.992862701416016, 145.04225158691406, 86.90618896484375, 125.4564208984375, 137.150634765625, -132.43557739257812, 115.00526428222656, 114.26687622070312, -2.1566619873046875, 24.82861328125, 7.4217529296875, -31.817604064941406, -46.776519775390625, 103.21672058105469, -65.67127227783203, -11.855056762695312, 133.9732666015625, 57.13666534423828, -37.14451599121094, 147.9690399169922, 143.90640258789062, 118.99348449707031, 97.25130462646484, 45.72003936767578, 154.98410034179688, 169.97767639160156, 162.56423950195312, -0.765045166015625, 7.325897216796875, -4.4278564453125, 169.676513671875, 44.30622100830078, 130.58209228515625, 28.30146026611328, -43.269168853759766, 7.640724182128906, 28.614013671875, -154.88133239746094, -14.675949096679688, 181.17762756347656, -76.59294891357422, 10.588668823242188, 148.54600524902344, 76.84139251708984, -105.480224609375, 102.85662841796875, 15.28594970703125, -27.366241455078125, 105.27477264404297, 29.890228271484375, -26.1812744140625, 69.58024597167969, 21.0283203125, 149.54551696777344, -6.410026550292969, 6.544776916503906], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000442.npy"}
{"epoch": 0.6681783824640968, "step": 443, "batch_size": 64, "mean": 60.69889831542969, "std": 89.70184326171875, "min": -149.8009490966797, "p10": -30.53099136352538, "median": 39.183921813964844, "p90": 178.3546310424805, "max": 244.88705444335938, "pos_frac": 0.765625, "sample": [2.272754669189453, 154.42013549804688, 55.562477111816406, 171.56056213378906, 20.117050170898438, 28.530853271484375, 183.57815551757812, -148.83621215820312, 48.601585388183594, 4.2701263427734375, 40.35459899902344, -74.7944107055664, 28.772747039794922, 140.14593505859375, -51.98313522338867, 170.991455078125, -4.641757965087891, 96.95620727539062, 129.52841186523438, 71.45762634277344, 244.88705444335938, -97.57857513427734, -17.100173950195312, 171.7258758544922, 2.7175941467285156, 188.94479370117188, 118.3302993774414, 13.468135833740234, 187.27734375, 171.87762451171875, 78.75789642333984, -20.497207641601562, 81.160888671875, 137.23660278320312, 9.168869018554688, 163.58030700683594, 63.43018341064453, 186.88551330566406, 176.02792358398438, 179.35179138183594, 45.86199188232422, 14.633769989013672, -14.309915542602539, 32.3657341003418, 174.819091796875, 159.4986114501953, -0.297821044921875, 38.01324462890625, 36.36054992675781, 19.947689056396484, 17.201513290405273, -4.224464416503906, 53.678070068359375, 3.8169479370117188, -5.883796691894531, 22.940689086914062, -149.8009490966797, -34.83118438720703, 189.78567504882812, 11.437105178833008, -79.0374755859375, -10.629447937011719, 86.118896484375, 170.74697875976562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000443.npy"}
{"epoch": 0.6696900982615268, "step": 444, "batch_size": 64, "mean": 21.587417602539062, "std": 79.17279815673828, "min": -169.8740234375, "p10": -58.70158195495605, "median": 7.146053314208984, "p90": 140.41323852539068, "max": 208.86932373046875, "pos_frac": 0.640625, "sample": [100.70498657226562, 24.246864318847656, -32.41572570800781, 2.546661376953125, 27.8319091796875, 3.9968605041503906, 110.15373992919922, 2.278045654296875, -41.380638122558594, 127.14761352539062, 15.226310729980469, 14.224838256835938, -77.56196594238281, -169.8740234375, 22.700897216796875, -0.770904541015625, 50.77631378173828, 7.329704284667969, 65.22537231445312, -11.195831298828125, 194.98062133789062, 115.83218383789062, -0.7006645202636719, -0.09740829467773438, -1.1342010498046875, 6.96240234375, -61.39017105102539, -6.23876953125, -1.4776687622070312, -11.927719116210938, 166.3954620361328, 0.87200927734375, 126.21399688720703, -32.58320617675781, 9.138534545898438, 44.252647399902344, 13.542953491210938, 172.77005004882812, -32.63860321044922, -168.24224853515625, 0.3996429443359375, 208.86932373046875, -149.5570526123047, -14.450767517089844, -105.65882110595703, 146.08607482910156, 4.361480712890625, 59.64556884765625, 127.17662048339844, 59.09974670410156, 10.202245712280273, 10.161521911621094, 10.058538436889648, 13.847343444824219, 157.24862670898438, -3.1587371826171875, -14.84659194946289, 23.126209259033203, 42.004905700683594, -52.42820739746094, -89.10706329345703, 153.510498046875, 3.9928722381591797, 5.289558410644531], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000444.npy"}
{"epoch": 0.671201814058957, "step": 445, "batch_size": 64, "mean": 69.86483764648438, "std": 86.01092529296875, "min": -112.41924285888672, "p10": -25.109025573730467, "median": 58.62308311462402, "p90": 180.21548004150392, "max": 190.17698669433594, "pos_frac": 0.765625, "sample": [109.46540832519531, 190.17698669433594, 5.72308349609375, -9.31884765625, -22.999038696289062, 177.56187438964844, 163.9589385986328, 157.2213592529297, 116.95851135253906, 78.04806518554688, -62.55775451660156, -59.9940071105957, 44.83643341064453, 0.23969268798828125, -21.070068359375, 157.31825256347656, -98.10691833496094, 1.7645339965820312, -26.28973388671875, 109.90206146240234, -0.40908050537109375, 182.13076782226562, -112.41924285888672, 27.240234375, -93.63871765136719, 178.58135986328125, 7.1456756591796875, 148.38577270507812, 36.76502990722656, -26.0133056640625, -8.6077880859375, -18.23120880126953, 61.56554412841797, 184.36068725585938, 169.72103881835938, 46.08296203613281, 188.70053100585938, 182.22938537597656, 55.68062210083008, 25.682235717773438, 153.24868774414062, 138.5349884033203, 63.051639556884766, 169.12954711914062, 134.77734375, 167.87977600097656, 180.9158172607422, 1.6159248352050781, 150.4144287109375, -7.677543640136719, 67.61811065673828, 3.3756561279296875, 3.7077484130859375, 46.20269775390625, 13.17844009399414, 144.57492065429688, 176.35604858398438, 108.2010269165039, 29.140098571777344, -1.1330718994140625, 153.0386962890625, 136.37725830078125, 182.9438018798828, 8.085891723632812], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000445.npy"}
{"epoch": 0.672713529856387, "step": 446, "batch_size": 64, "mean": 64.80374145507812, "std": 76.597900390625, "min": -111.48591613769531, "p10": -13.869842338562005, "median": 54.130210876464844, "p90": 175.0442306518555, "max": 192.3980712890625, "pos_frac": 0.828125, "sample": [4.091045379638672, -85.48606872558594, -40.404117584228516, 82.25576782226562, 22.03673553466797, 2.838897705078125, 75.03392791748047, 18.326004028320312, 184.74221801757812, 77.1452865600586, 110.89656066894531, 151.56716918945312, 79.87893676757812, 150.3810577392578, 107.61901092529297, 181.5703125, 3.8731231689453125, 56.737945556640625, 99.10479736328125, 28.823883056640625, -0.9788990020751953, 79.01475524902344, 3.451629638671875, 149.73663330078125, 166.1974639892578, 83.43561553955078, 51.52247619628906, 168.6331787109375, 189.17538452148438, -38.75068664550781, 126.71601867675781, 5.473728179931641, -37.718379974365234, 2.76513671875, -111.48591613769531, -8.034252166748047, 112.44901275634766, -6.028388977050781, 38.34809875488281, 21.43531036376953, 6.21746826171875, 184.77059936523438, 180.19338989257812, -17.326810836791992, 192.3980712890625, 2.1232948303222656, 172.56854248046875, 114.07450103759766, -16.37080955505371, 13.228445053100586, 59.92521667480469, 5.452339172363281, 22.357254028320312, 36.474082946777344, 161.63739013671875, 176.10523986816406, 7.838539123535156, 11.867134094238281, -0.5499839782714844, 160.37240600585938, 0.4828052520751953, 125.69957733154297, 86.39768981933594, 155.1125946044922], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000446.npy"}
{"epoch": 0.674225245653817, "step": 447, "batch_size": 64, "mean": 55.69664764404297, "std": 88.21941375732422, "min": -183.704345703125, "p10": -57.16727600097652, "median": 47.52202606201172, "p90": 171.08907623291014, "max": 199.14613342285156, "pos_frac": 0.78125, "sample": [189.0718231201172, 108.31672668457031, 109.86463928222656, 4.009098052978516, 52.46392059326172, -16.273223876953125, 57.65960693359375, 10.414299011230469, -100.35514831542969, 22.265823364257812, 32.80455017089844, 111.29844665527344, 3.8572311401367188, 30.2200927734375, 83.2269058227539, 170.56497192382812, -6.304872512817383, 161.5788116455078, 7.547229766845703, -0.5578575134277344, 31.958480834960938, -134.32583618164062, 79.43038940429688, -87.21321105957031, 5.647529602050781, 182.802734375, 168.27593994140625, 171.151611328125, 133.10055541992188, -74.69329833984375, 47.22419738769531, 123.11534118652344, 36.644927978515625, 147.756591796875, 31.54372787475586, 70.86686706542969, 176.72012329101562, 88.03692626953125, 175.4177703857422, -2.3949012756347656, 2.4556541442871094, 47.819854736328125, 150.18966674804688, -3.27459716796875, 2.2705421447753906, -117.00669860839844, 70.92886352539062, 170.9431610107422, 10.308921813964844, 159.11880493164062, -3.7097091674804688, 67.80255126953125, -95.87710571289062, 85.22576141357422, 169.0923614501953, 27.305984497070312, 140.64393615722656, -183.704345703125, 6.813648223876953, 199.14613342285156, 182.94517517089844, -11.411209106445312, 22.2637939453125, 61.55469512939453], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000447.npy"}
{"epoch": 0.6757369614512472, "step": 448, "batch_size": 64, "mean": 48.574462890625, "std": 82.53495025634766, "min": -169.05905151367188, "p10": -47.297231292724604, "median": 30.894834518432617, "p90": 160.71553955078124, "max": 203.7117919921875, "pos_frac": 0.734375, "sample": [162.89657592773438, -102.26704406738281, 160.90591430664062, 45.54411697387695, 11.921310424804688, 23.930191040039062, -0.14175796508789062, 131.29258728027344, 108.57192993164062, 6.00042724609375, 145.1802215576172, 74.36017608642578, 51.38441467285156, 36.52410125732422, 159.98757934570312, 147.71543884277344, -18.991880416870117, 177.58157348632812, 51.52131652832031, -50.58251953125, 50.90950012207031, 179.77328491210938, 1.4180831909179688, 10.968292236328125, 99.94129180908203, 0.07112884521484375, 147.7509307861328, 1.8608531951904297, 17.98926544189453, -44.889801025390625, -82.3140640258789, -16.547119140625, 38.3986701965332, 26.913589477539062, -3.0791549682617188, 50.828025817871094, -98.2188491821289, 34.87607955932617, 181.5753173828125, -4.833690643310547, 22.483142852783203, -169.05905151367188, 103.37409973144531, 10.729400634765625, 7.8800811767578125, 150.93588256835938, 122.9795913696289, 160.27133178710938, 55.83380889892578, -19.90923309326172, -54.70358657836914, -48.32898712158203, -24.69348907470703, 82.61089324951172, -4.143115997314453, 203.7117919921875, 5.527961730957031, -32.54669189453125, 134.22665405273438, 146.804931640625, 118.1885986328125, 25.332931518554688, 5.834770202636719, 188.69789123535156], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000448.npy"}
{"epoch": 0.6772486772486772, "step": 449, "batch_size": 64, "mean": 28.319255828857422, "std": 86.37510681152344, "min": -177.61215209960938, "p10": -83.79663619995117, "median": 18.34130859375, "p90": 156.45977478027345, "max": 190.27633666992188, "pos_frac": 0.65625, "sample": [0.18745803833007812, 18.603042602539062, -23.60955047607422, 61.399078369140625, 5.7031097412109375, -20.03162384033203, 7.367698669433594, 24.185165405273438, 30.720794677734375, 152.86402893066406, 146.12217712402344, 159.74237060546875, 2.525146484375, -124.65809631347656, -18.00664520263672, 19.124160766601562, 190.240966796875, 8.752706527709961, 4.9574737548828125, 84.26463317871094, 158.0008087158203, 46.0417366027832, 173.45596313476562, 178.94216918945312, -104.59683227539062, 33.898895263671875, 124.63121032714844, -53.39300537109375, 75.9218521118164, 117.69618225097656, 12.101165771484375, -44.204200744628906, 35.134979248046875, -177.61215209960938, 105.94467163085938, 58.2886962890625, 190.27633666992188, -37.142799377441406, -20.57790756225586, 11.51535415649414, -80.81840515136719, 86.76863098144531, -97.12548828125, 45.55206298828125, -11.413471221923828, 18.079574584960938, -28.376129150390625, 161.29364013671875, 1.5790729522705078, 104.4351577758789, 134.00152587890625, -22.194313049316406, 23.5972900390625, -1.002716064453125, -16.5821533203125, 48.598663330078125, 119.98871612548828, -118.03529357910156, 23.345623016357422, -85.0730209350586, 111.5365219116211, -173.28170776367188, -12.781112670898438, -34.437530517578125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000449.npy"}
{"epoch": 0.6787603930461074, "step": 450, "batch_size": 64, "mean": 47.054134368896484, "std": 86.29766082763672, "min": -167.95701599121094, "p10": -58.50404891967772, "median": 28.53791618347168, "p90": 161.3796127319336, "max": 200.80581665039062, "pos_frac": 0.671875, "sample": [-1.8328285217285156, -68.71688842773438, 53.252479553222656, -26.683643341064453, 131.97601318359375, 117.74628448486328, 103.9654312133789, 14.339767456054688, 120.83777618408203, 184.57723999023438, 155.3601531982422, 28.95050048828125, 28.12533187866211, 81.97811889648438, -5.438800811767578, 103.50403594970703, 5.9833984375, 3.2581863403320312, 101.4673080444336, 194.8621063232422, -12.033226013183594, 156.70208740234375, -0.4103240966796875, 59.19291687011719, -45.71485900878906, -5.188396453857422, 40.683746337890625, 159.1820831298828, 37.797218322753906, -63.985130310058594, 132.68487548828125, 179.85455322265625, -4.6226043701171875, 82.76231384277344, -167.95701599121094, 177.84239196777344, 159.59266662597656, -1.9891433715820312, 189.5067138671875, 78.48020935058594, 4.01202392578125, 44.79188537597656, 92.71623229980469, -86.71033477783203, 162.14544677734375, 10.262550354003906, 18.558513641357422, 1.94146728515625, 140.961181640625, -89.59977722167969, 21.47943878173828, -24.15001678466797, -12.933921813964844, 87.152587890625, 200.80581665039062, 17.042999267578125, -143.4419708251953, 122.59608459472656, -1.8335037231445312, 110.16292572021484, -90.78407287597656, 8.635211944580078, -18.83273696899414, -43.40646743774414], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000450.npy"}
{"epoch": 0.6802721088435374, "step": 451, "batch_size": 64, "mean": 40.36961364746094, "std": 80.75834655761719, "min": -143.10488891601562, "p10": -60.96357421874997, "median": 17.179691314697266, "p90": 168.2322998046875, "max": 195.28759765625, "pos_frac": 0.703125, "sample": [97.16427612304688, -30.5408935546875, -6.7861480712890625, 67.2276611328125, 181.33074951171875, 16.433380126953125, -108.71730041503906, 17.80999755859375, 29.567951202392578, 164.61618041992188, 39.037139892578125, 112.02125549316406, 177.61753845214844, 20.8646240234375, 178.25421142578125, 34.51277160644531, 5.867988586425781, 10.028915405273438, 3.710002899169922, 31.805593490600586, 96.93729400634766, -100.4313735961914, -2.454479217529297, 195.28759765625, 168.33804321289062, -16.20892333984375, 15.13555908203125, -143.10488891601562, -83.41195678710938, -3.883127212524414, 16.54938507080078, 133.4672393798828, -84.8056640625, 3.531219482421875, 131.18386840820312, 48.5755615234375, 167.98556518554688, 82.04331970214844, -10.523616790771484, 3.09088134765625, -31.092247009277344, 74.25218200683594, 2.7751693725585938, 71.62326049804688, 104.22312927246094, 16.5469970703125, 32.98543930053711, 166.47898864746094, 171.937744140625, 144.78053283691406, 91.63333129882812, 52.97869873046875, 9.824089050292969, -6.270599365234375, -20.304916381835938, -73.76557159423828, 6.7903594970703125, -4.2780609130859375, -4.4310302734375, 23.564109802246094, 183.37945556640625, 3.1639480590820312, -75.07151794433594, -17.195693969726562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000451.npy"}
{"epoch": 0.6817838246409675, "step": 452, "batch_size": 64, "mean": 55.43278121948242, "std": 88.30722045898438, "min": -165.94879150390625, "p10": -40.148468780517575, "median": 49.71333312988281, "p90": 168.90757751464844, "max": 193.71717834472656, "pos_frac": 0.6875, "sample": [50.355308532714844, -6.8984527587890625, 124.59590148925781, 106.57872772216797, -20.841800689697266, 139.54034423828125, 49.07135772705078, 188.7315673828125, 176.4354248046875, 159.18405151367188, 21.908790588378906, 125.88920593261719, -26.85443115234375, 160.8762664794922, 169.65689086914062, 41.311004638671875, 190.3997039794922, 91.36811828613281, 118.83735656738281, 34.4005126953125, 33.00603485107422, -42.900733947753906, 88.655517578125, 58.812591552734375, -1.4711742401123047, -7.060733795166016, 7.433677673339844, 5.9596099853515625, 64.1307144165039, 0.8321647644042969, -1.3910980224609375, 193.71717834472656, -108.235595703125, 167.1591796875, -46.548484802246094, 166.7820587158203, 132.27809143066406, 0.0027618408203125, -103.06103515625, 10.542404174804688, 134.79702758789062, -25.2523193359375, 142.62442016601562, 88.7253189086914, 151.5308837890625, -165.37258911132812, 183.1884765625, 83.98037719726562, 87.4944076538086, -7.014381408691406, 146.81785583496094, -4.443605422973633, 9.947517395019531, 117.89768981933594, -2.578685760498047, 26.87120819091797, -7.559349060058594, 126.04710388183594, 57.90193557739258, -64.88964080810547, -33.72651672363281, 173.51243591308594, -165.94879150390625, -20.041946411132812], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000452.npy"}
{"epoch": 0.6832955404383976, "step": 453, "batch_size": 64, "mean": 47.16241455078125, "std": 88.35204315185547, "min": -159.32772827148438, "p10": -36.19218063354492, "median": 24.311464309692383, "p90": 170.64566345214843, "max": 180.71563720703125, "pos_frac": 0.71875, "sample": [86.81568908691406, 61.16826248168945, -7.800930023193359, -35.43614196777344, 170.77487182617188, 18.16368865966797, -8.86709976196289, 170.34417724609375, 179.23040771484375, 31.318359375, -156.21755981445312, 3.8107986450195312, 109.11550903320312, -13.645355224609375, 173.5712890625, 170.09671020507812, 135.37396240234375, 60.027557373046875, 157.76223754882812, 43.94065856933594, 10.063156127929688, -81.31451416015625, 23.60394287109375, 21.48863983154297, 0.49413299560546875, -63.79008483886719, -91.61766052246094, 155.06002807617188, 4.861476898193359, 17.498306274414062, 158.10818481445312, 30.037349700927734, 48.907039642333984, -29.02727508544922, -11.212005615234375, 9.531425476074219, 146.28656005859375, 24.973522186279297, -159.32772827148438, 169.2266845703125, 23.64940643310547, 175.30079650878906, -14.475517272949219, -32.62179946899414, -0.5874252319335938, 46.34246826171875, 0.7911205291748047, 5.1314239501953125, 159.4774932861328, 17.10406494140625, 31.351043701171875, -36.516197204589844, 49.835662841796875, -0.6664333343505859, 47.10016632080078, -147.1744384765625, -1.2448806762695312, 156.3070526123047, 180.71563720703125, 14.682098388671875, 180.3228302001953, 151.24392700195312, 175.70440673828125, 103.22311401367188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000453.npy"}
{"epoch": 0.6848072562358276, "step": 454, "batch_size": 64, "mean": 64.81204986572266, "std": 81.46305847167969, "min": -190.52389526367188, "p10": -15.676013946533196, "median": 62.37913131713867, "p90": 168.46100463867188, "max": 183.9268341064453, "pos_frac": 0.796875, "sample": [174.71681213378906, -7.779998779296875, -27.845542907714844, 166.31170654296875, -190.52389526367188, -1.6267566680908203, 156.7074737548828, 10.391944885253906, 109.44781494140625, 3.3385486602783203, 155.86141967773438, 134.96559143066406, 37.638816833496094, 0.844512939453125, 182.3294677734375, 109.3865966796875, 169.10504150390625, 110.59210205078125, 30.5797119140625, 36.19915771484375, -7.174713134765625, -119.0649185180664, 3.1693878173828125, 145.31060791015625, 113.25108337402344, 176.31130981445312, -1.7283782958984375, 104.89326477050781, 5.2302093505859375, 6.726531982421875, 61.12220764160156, 124.0058364868164, 63.63605499267578, -94.44229125976562, 144.35316467285156, 112.5982666015625, 183.9268341064453, -31.218843460083008, 181.75485229492188, 104.0125732421875, 94.89712524414062, 34.87504577636719, 17.109825134277344, -5.06927490234375, -33.61381912231445, 58.68730926513672, 166.958251953125, 82.4635238647461, -19.060020446777344, 31.8662052154541, 127.15516662597656, -1.6927642822265625, 183.82980346679688, 153.88671875, 102.65216827392578, 97.73534393310547, 68.73133850097656, 140.855712890625, 6.05291748046875, 153.76141357421875, 39.89053726196289, 3.74395751953125, 4.390657424926758, 0.5503654479980469], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000454.npy"}
{"epoch": 0.6863189720332578, "step": 455, "batch_size": 64, "mean": 61.83723449707031, "std": 78.7041244506836, "min": -126.43310546875, "p10": -10.797367858886718, "median": 43.47693634033203, "p90": 175.4764617919922, "max": 208.84024047851562, "pos_frac": 0.796875, "sample": [46.429473876953125, 68.1259765625, 94.80459594726562, 191.21998596191406, 8.6932373046875, 2.3361854553222656, 135.16458129882812, 191.42312622070312, 125.44890594482422, 2.8057708740234375, 115.16558074951172, 2.153839111328125, 65.0578842163086, 167.37548828125, 120.81430053710938, 190.7671661376953, 22.79583740234375, -17.926010131835938, -9.77337646484375, -126.43310546875, 152.1208953857422, 0.2519989013671875, 161.47740173339844, 4.302906036376953, -0.08798027038574219, 177.08853149414062, 100.42552185058594, 101.90910339355469, 171.7149658203125, 11.930465698242188, 102.84757995605469, 14.16107177734375, 179.93316650390625, 9.81744384765625, 4.112640380859375, 40.73976135253906, -69.66543579101562, 117.34607696533203, 126.90401458740234, 133.3195343017578, -102.86453247070312, 3.0809478759765625, 81.49560546875, -28.28961181640625, 156.74313354492188, 100.61630249023438, 160.48178100585938, -9.346687316894531, 18.647247314453125, 10.875350952148438, 191.97097778320312, 18.031906127929688, 6.4161376953125, 96.59249114990234, 46.214111328125, -0.5715255737304688, 25.47482681274414, -7.346366882324219, -3.208099365234375, 5.218177795410156, -11.236221313476562, 64.68350982666016, -12.03585433959961, 208.84024047851562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000455.npy"}
{"epoch": 0.6878306878306878, "step": 456, "batch_size": 64, "mean": 52.61790084838867, "std": 78.68453979492188, "min": -125.450927734375, "p10": -20.49021072387695, "median": 30.55601692199707, "p90": 169.30971679687502, "max": 198.87399291992188, "pos_frac": 0.75, "sample": [106.98609924316406, 5.7040252685546875, 1.754037857055664, 93.96253967285156, 198.87399291992188, 2.865447998046875, -2.0043277740478516, 18.83814239501953, -5.187841415405273, 182.70468139648438, 48.71992492675781, 72.74663543701172, 152.940185546875, 120.98587799072266, 24.952392578125, 4.7390289306640625, 75.05148315429688, 186.58944702148438, -1.6340484619140625, 131.92239379882812, -22.67071533203125, 13.109786987304688, 45.00587463378906, -3.737884521484375, 106.65034484863281, 60.3104133605957, -125.450927734375, -42.23622131347656, 114.81787109375, -72.05880737304688, -2.4290847778320312, 80.07080078125, 25.347572326660156, 39.45940399169922, 3.92108154296875, 151.0720672607422, 110.0092544555664, -3.9582138061523438, 162.08612060546875, -1.2215251922607422, -112.58297729492188, 140.44923400878906, 38.2896842956543, 152.37989807128906, 31.364032745361328, -38.04132080078125, 103.28033447265625, 11.50555419921875, 5.1600494384765625, 1.6035919189453125, 171.34323120117188, 179.0010986328125, -15.402366638183594, 192.42306518554688, 29.748001098632812, 105.01998901367188, 176.74940490722656, 37.41435241699219, -5.736625671386719, 164.56484985351562, 7.097801208496094, -102.4053955078125, 5.010162353515625, 29.702880859375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000456.npy"}
{"epoch": 0.6893424036281179, "step": 457, "batch_size": 64, "mean": 44.53820037841797, "std": 73.54702758789062, "min": -141.48745727539062, "p10": -23.395715332031248, "median": 26.40020751953125, "p90": 152.17783966064454, "max": 189.52145385742188, "pos_frac": 0.75, "sample": [8.985603332519531, -3.66033935546875, -21.297725677490234, 1.7677001953125, 4.5155792236328125, -70.9025650024414, 28.90851402282715, 35.1138916015625, 41.289161682128906, -12.061897277832031, 4.700817108154297, 40.81549072265625, 92.9318618774414, -1.812652587890625, -7.599126815795898, -141.48745727539062, 169.05099487304688, 4.796180725097656, -27.597164154052734, 11.648155212402344, 186.02333068847656, 74.56759643554688, 189.52145385742188, 67.05557250976562, 5.198453903198242, -24.658344268798828, 2.682891845703125, 143.039794921875, 24.521377563476562, 27.412029266357422, -21.647506713867188, -132.80162048339844, 3.9425277709960938, 99.33772277832031, 98.08802795410156, -1.0275421142578125, -23.744659423828125, 15.580810546875, 149.50662231445312, 51.5511474609375, 130.9232177734375, -11.5020751953125, 177.61172485351562, 140.94967651367188, -22.581512451171875, 14.57232666015625, 153.32264709472656, 85.9736557006836, 25.388385772705078, 175.2488555908203, 59.13063430786133, 37.86726760864258, 120.78775024414062, 74.18777465820312, 102.29208374023438, 172.01406860351562, 114.05535888671875, 28.38779067993164, 9.41177749633789, 12.456764221191406, 42.39544677734375, 19.69036865234375, 147.05987548828125, -51.45372009277344], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000457.npy"}
{"epoch": 0.690854119425548, "step": 458, "batch_size": 64, "mean": 54.548736572265625, "std": 82.9202651977539, "min": -161.18116760253906, "p10": -44.45641021728515, "median": 36.383609771728516, "p90": 171.16892395019534, "max": 197.10101318359375, "pos_frac": 0.75, "sample": [-22.468292236328125, 150.95191955566406, 4.861545562744141, 24.112335205078125, 3.235902786254883, 34.272369384765625, 163.99722290039062, -161.18116760253906, -1.6229057312011719, 58.81378936767578, 38.50869369506836, -43.13783264160156, -4.906951904296875, 85.16961669921875, 147.40621948242188, 26.487987518310547, 56.52978515625, 38.494850158691406, 4.708683013916016, 22.120529174804688, 4.091072082519531, 9.976875305175781, 86.40550231933594, -53.20825958251953, 132.02896118164062, 129.9065704345703, 197.10101318359375, 106.5321044921875, 10.151229858398438, 139.31466674804688, 172.12362670898438, 115.55284881591797, 13.799995422363281, -45.021514892578125, 15.686248779296875, -0.9361763000488281, -4.088462829589844, 22.089927673339844, 39.198089599609375, 165.51422119140625, 78.7702865600586, 87.85926055908203, -123.12775421142578, -2.535825729370117, 168.9412841796875, -0.3462371826171875, 147.63284301757812, -4.633544921875, 181.09194946289062, 177.15350341796875, -47.059059143066406, 144.8197021484375, 86.759765625, 55.3099365234375, -113.3065414428711, 181.2724151611328, 91.21487426757812, 182.79605102539062, -49.1151123046875, 3.9654922485351562, 9.253335952758789, 25.641563415527344, 150.56781005859375, 175.6202392578125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000458.npy"}
{"epoch": 0.6923658352229781, "step": 459, "batch_size": 64, "mean": 62.62535095214844, "std": 79.98074340820312, "min": -131.44100952148438, "p10": -21.00505752563476, "median": 45.71685028076172, "p90": 174.44976043701175, "max": 216.5973663330078, "pos_frac": 0.75, "sample": [93.4144058227539, 95.47782897949219, -94.87804412841797, 180.0955047607422, 0.6111526489257812, -65.77377319335938, 50.99980163574219, 127.23818969726562, 149.224609375, 147.2840118408203, 59.469024658203125, -0.43640899658203125, 137.05831909179688, 5.156455993652344, 162.35504150390625, 203.46768188476562, 0.8251266479492188, 35.1978759765625, 41.794525146484375, 126.33651733398438, 6.312858581542969, 51.331634521484375, 15.407554626464844, -5.078330993652344, 20.57268524169922, 121.41793823242188, -1.834686279296875, 126.66338348388672, 161.10211181640625, 170.27049255371094, 216.5973663330078, 26.256683349609375, -17.37139892578125, 1.2686119079589844, -22.562339782714844, 118.51008605957031, 29.606613159179688, -131.44100952148438, 168.6177978515625, -2.2700042724609375, 2.8855857849121094, 110.11834716796875, 107.9949951171875, 1.9256610870361328, 133.16061401367188, -24.66144561767578, 73.64480590820312, 32.30341339111328, -44.86772155761719, 125.24384307861328, -1.7868194580078125, -8.421302795410156, -9.185745239257812, 86.30183410644531, 49.63917541503906, -22.923812866210938, 108.2978286743164, 34.81568908691406, 190.71604919433594, 176.24087524414062, 183.96963500976562, 14.761367797851562, 188.23809814453125, -8.684249877929688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000459.npy"}
{"epoch": 0.6938775510204082, "step": 460, "batch_size": 64, "mean": 57.389373779296875, "std": 79.22606658935547, "min": -81.02791595458984, "p10": -28.904821395874013, "median": 28.17438507080078, "p90": 172.24297637939452, "max": 212.4769744873047, "pos_frac": 0.78125, "sample": [35.71226501464844, 126.31472778320312, 212.4769744873047, -17.233657836914062, 12.328128814697266, 166.60447692871094, 38.09370422363281, 31.177459716796875, 167.52529907226562, 1.2978973388671875, -18.615039825439453, 7.622802734375, -33.314727783203125, 22.779579162597656, -7.7653656005859375, 36.083412170410156, 16.881160736083984, 138.0410919189453, -48.999298095703125, 118.13239288330078, 65.07604217529297, 191.898681640625, 1.1036834716796875, 106.17532348632812, 180.30389404296875, 76.33683776855469, 10.172203063964844, 25.171310424804688, 167.6488037109375, 5.3982391357421875, 13.0804443359375, 177.39566040039062, 71.95877075195312, 136.6671142578125, -37.4881477355957, 32.7454833984375, 3.9521846771240234, -66.10397338867188, 8.331771850585938, -3.2298316955566406, 171.8334503173828, 155.14736938476562, 9.642646789550781, -2.109233856201172, 98.80799865722656, 177.92657470703125, -43.77820587158203, -81.02791595458984, 183.66302490234375, -0.5965728759765625, 0.8526592254638672, 9.713348388671875, 146.25474548339844, 93.18579864501953, 52.746673583984375, 172.41848754882812, 0.7054367065429688, 160.67034912109375, -80.94832611083984, -0.2096405029296875, 7.328216552734375, 97.47926330566406, 2.92401123046875, 168.55191040039062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000460.npy"}
{"epoch": 0.6953892668178382, "step": 461, "batch_size": 64, "mean": 64.19013977050781, "std": 73.14051055908203, "min": -105.51813507080078, "p10": -23.122720718383786, "median": 48.47254180908203, "p90": 169.975114440918, "max": 197.34864807128906, "pos_frac": 0.765625, "sample": [142.09738159179688, 91.6171875, 42.45973205566406, 97.36138153076172, 172.32176208496094, 81.33987426757812, 74.35015869140625, 154.5904541015625, 76.87157440185547, -8.788036346435547, 29.777786254882812, 28.4073486328125, 136.76300048828125, -8.467575073242188, 12.495351791381836, -19.75731658935547, -7.82769775390625, 164.49960327148438, 4.894220352172852, 23.830772399902344, 33.97151184082031, 131.01962280273438, 48.49275207519531, -3.1081085205078125, 4.472663879394531, 115.73847961425781, -30.648712158203125, -27.735862731933594, -24.464065551757812, 19.941925048828125, 154.05889892578125, -2.0637893676757812, -24.71149444580078, 91.61903381347656, 117.58403015136719, 46.58258056640625, 120.78524780273438, 137.69796752929688, 59.08424377441406, 197.34864807128906, 155.9626007080078, 174.20010375976562, 156.79759216308594, 103.78871154785156, 4.921894073486328, 44.16517639160156, 27.32248306274414, -35.760677337646484, 153.58187866210938, 6.0562286376953125, 180.05828857421875, 10.861335754394531, 78.60149383544922, 64.53813171386719, 184.06094360351562, -105.51813507080078, 107.18653869628906, 2.1662826538085938, -0.8041286468505859, -41.195343017578125, 178.88925170898438, -19.992916107177734, 175.32608032226562, 48.45233154296875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000461.npy"}
{"epoch": 0.6969009826152683, "step": 462, "batch_size": 64, "mean": 67.24839782714844, "std": 90.12709045410156, "min": -179.9259033203125, "p10": -41.660612106323235, "median": 70.57659912109375, "p90": 177.68438262939452, "max": 202.4222869873047, "pos_frac": 0.734375, "sample": [177.4288787841797, 183.39251708984375, 170.95736694335938, 134.91217041015625, -1.4233856201171875, -2.724853515625, 177.9961700439453, -50.09086990356445, -0.9444122314453125, 3.51953125, 123.97119140625, 179.60809326171875, 177.79388427734375, 21.98406982421875, 163.81646728515625, 182.93515014648438, -36.295936584472656, 124.84611511230469, 102.07495880126953, 154.6112060546875, 161.4478759765625, 74.2025375366211, 154.73721313476562, 162.8125, 13.366104125976562, 0.22960662841796875, 158.83619689941406, 175.63136291503906, -0.8590202331542969, 134.30343627929688, -1.13128662109375, -70.28205871582031, 68.22398376464844, 156.87091064453125, 182.91651916503906, 4.795562744140625, 1.1742305755615234, 88.07012176513672, 59.92909240722656, 38.7491455078125, 37.2818603515625, 64.19926452636719, 114.41505432128906, 13.855499267578125, -2.48223876953125, 158.27105712890625, -43.32002258300781, 83.4784164428711, 114.18880462646484, -91.96102142333984, -3.8094406127929688, -179.9259033203125, 167.32669067382812, 59.82770538330078, 81.75140380859375, -74.70072937011719, 72.92921447753906, 202.4222869873047, 44.133018493652344, 116.00762176513672, -144.41754150390625, -37.78865432739258, -4.193168640136719, 4.015861511230469], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000462.npy"}
{"epoch": 0.6984126984126984, "step": 463, "batch_size": 64, "mean": 40.18242645263672, "std": 80.75404357910156, "min": -118.89228820800781, "p10": -43.94905166625976, "median": 12.596916198730469, "p90": 161.64551239013673, "max": 188.76156616210938, "pos_frac": 0.59375, "sample": [-118.89228820800781, 52.46648406982422, 146.16046142578125, -95.76055908203125, 137.68572998046875, 126.03443145751953, -13.93869400024414, 181.7217559814453, 181.2427978515625, 119.376708984375, 10.149986267089844, -8.908889770507812, 6.540557861328125, 7.557842254638672, 153.75328063964844, -1.812713623046875, 47.04267883300781, -21.1636962890625, -2.4106178283691406, -9.122200012207031, -99.72377014160156, 7.269828796386719, -32.76629638671875, 173.20286560058594, 86.7096176147461, -37.98284912109375, 49.32826232910156, 8.476272583007812, -4.991485595703125, -15.540641784667969, -0.02613067626953125, -5.1111602783203125, 133.8819122314453, 166.48878479003906, -42.951629638671875, 97.899169921875, 79.5136947631836, 114.81230926513672, 119.66441345214844, 188.76156616210938, 67.30123138427734, -64.31729125976562, 62.74072265625, -102.85945129394531, -24.807510375976562, -14.267303466796875, 184.6142578125, -69.0652084350586, 158.7895965576172, 73.34408569335938, 59.868247985839844, -37.91657257080078, 1.2443466186523438, -22.0726318359375, 22.385238647460938, 116.40687561035156, 91.81062316894531, -6.453399658203125, 16.95435333251953, 39.90052795410156, -44.37651824951172, 15.043846130371094, -0.10003280639648438, 162.86947631835938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000463.npy"}
{"epoch": 0.6999244142101285, "step": 464, "batch_size": 64, "mean": 56.42665481567383, "std": 96.4330825805664, "min": -194.61245727539062, "p10": -41.41757621765136, "median": 35.89184379577637, "p90": 182.35101470947265, "max": 193.37411499023438, "pos_frac": 0.734375, "sample": [-6.9848785400390625, 19.893714904785156, 134.5081024169922, -96.28135681152344, 179.88023376464844, -12.123725891113281, 1.3781089782714844, 171.54721069335938, 55.89613342285156, 28.589813232421875, 180.37286376953125, -14.386478424072266, 35.585609436035156, 73.3321533203125, 123.74173736572266, -21.297271728515625, 181.42930603027344, -31.47002410888672, 5.0662078857421875, 60.22373962402344, 14.571517944335938, 189.28634643554688, -23.8756103515625, 77.06880187988281, 115.47148132324219, 182.3770751953125, 2.8835105895996094, 192.65782165527344, 46.25437545776367, 124.21027374267578, 4.790130615234375, 4.907756805419922, 163.61703491210938, 191.497802734375, -194.61245727539062, 53.989654541015625, 160.57977294921875, 177.71388244628906, -138.7162322998047, -10.127037048339844, 175.16961669921875, 16.224517822265625, 99.70623779296875, 193.37411499023438, -125.81716918945312, -1.8005752563476562, -0.9911956787109375, -1.0255241394042969, -54.40391540527344, 168.5305938720703, 36.19807815551758, 182.2902069091797, -144.51226806640625, -45.68081283569336, 8.028762817382812, 1.741933822631836, 2.81158447265625, 184.5614013671875, 109.80401611328125, 184.0832061767578, 4.486980438232422, 137.2287139892578, 76.02909088134766, 1.8210601806640625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000464.npy"}
{"epoch": 0.7014361300075586, "step": 465, "batch_size": 64, "mean": 54.408992767333984, "std": 85.4114761352539, "min": -109.52354431152344, "p10": -41.86314544677734, "median": 33.38645553588867, "p90": 181.58215637207033, "max": 201.36239624023438, "pos_frac": 0.734375, "sample": [-45.73209762573242, 201.36239624023438, 144.21604919433594, 3.375762939453125, 169.11679077148438, 110.33721923828125, -16.005599975585938, 187.12860107421875, 17.141159057617188, 107.322509765625, 65.50678253173828, 182.28988647460938, 77.00102233886719, 189.1802215576172, -32.83559036254883, 110.54231262207031, 5.2594146728515625, 179.9307861328125, 66.83184814453125, 7.020233154296875, 23.723350524902344, -86.88321685791016, 135.3310546875, 189.45013427734375, 133.78016662597656, 188.55465698242188, -93.99252319335938, -96.14319610595703, -5.192253112792969, -5.854400634765625, -1.888519287109375, -3.4689483642578125, 19.7100830078125, 41.62787628173828, 89.19418334960938, 88.48654174804688, -109.52354431152344, -22.24502944946289, -1.8493309020996094, 153.203125, -11.706924438476562, 41.315574645996094, 101.05914306640625, 1.6266250610351562, 4.491447448730469, -108.41458129882812, -22.504173278808594, 12.857616424560547, 72.259033203125, 168.7401885986328, 25.45733642578125, 167.7645721435547, 5.197273254394531, 56.25369644165039, 5.9805145263671875, -61.728431701660156, 23.95648193359375, 1.6033096313476562, 16.381561279296875, 56.477569580078125, 190.90521240234375, 158.09890747070312, 160.34922790527344, 50.74443054199219], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000465.npy"}
{"epoch": 0.7029478458049887, "step": 466, "batch_size": 64, "mean": 36.63169479370117, "std": 78.48873138427734, "min": -165.91259765625, "p10": -25.506410217285154, "median": 9.628667831420898, "p90": 151.97116088867188, "max": 197.85818481445312, "pos_frac": 0.640625, "sample": [-13.511951446533203, 106.87510681152344, -24.197189331054688, -16.999984741210938, 97.86549377441406, -17.400524139404297, 7.235176086425781, -20.026947021484375, -123.27192687988281, -41.529541015625, 15.561088562011719, 113.43730926513672, 171.809326171875, -74.73442077636719, 87.14111328125, -4.505889892578125, 96.57028198242188, 179.29977416992188, 40.492919921875, 181.97799682617188, 151.12478637695312, 1.0578880310058594, 127.10846710205078, -2.2358760833740234, -5.7033843994140625, 2.6166305541992188, 40.41941452026367, -165.91259765625, 197.85818481445312, 11.345611572265625, 5.378253936767578, 18.859725952148438, -5.24822998046875, 75.22761535644531, 66.78526306152344, 5.2457122802734375, 15.356986999511719, 152.33389282226562, -18.118057250976562, 76.14573669433594, 0.26665496826171875, -26.0675048828125, -126.6962890625, -6.832237243652344, 128.59906005859375, -2.5078353881835938, 9.452484130859375, 143.29940795898438, 194.71856689453125, -11.890571594238281, -31.2236328125, 29.95672607421875, -22.448348999023438, 122.39287567138672, -2.231353759765625, 0.8473758697509766, 9.804851531982422, -11.34743881225586, 96.36416625976562, 95.43505859375, 174.55921936035156, 44.78826904296875, 14.79388427734375, 8.661766052246094], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000466.npy"}
{"epoch": 0.7044595616024187, "step": 467, "batch_size": 64, "mean": 55.23102569580078, "std": 77.14759826660156, "min": -117.17921447753906, "p10": -13.869064331054686, "median": 21.959816932678223, "p90": 170.3602783203125, "max": 224.60491943359375, "pos_frac": 0.796875, "sample": [10.777900695800781, 131.4083709716797, 90.82666778564453, 15.713516235351562, -3.296600341796875, 11.123287200927734, 3.856304168701172, -55.25264358520508, 47.68360900878906, 11.41143798828125, -0.09157562255859375, -9.78573226928711, 163.98880004882812, 21.49296760559082, 17.851089477539062, 81.2699203491211, 127.97940063476562, 18.309341430664062, -12.333328247070312, 9.020271301269531, 92.36285400390625, -117.17921447753906, -38.946048736572266, 193.41738891601562, 181.42941284179688, 78.7247314453125, 114.05287170410156, 113.48153686523438, -74.24313354492188, 182.39752197265625, -14.527236938476562, 188.65953063964844, 80.40618896484375, -2.6548004150390625, 138.1292266845703, 0.34808349609375, 104.8320541381836, 158.5829620361328, 51.7772216796875, 23.41510009765625, 3.4107532501220703, 0.6757869720458984, -26.14307403564453, 56.163509368896484, 10.645368576049805, -66.51759338378906, 22.426666259765625, 0.630340576171875, 165.81756591796875, 142.46078491210938, 224.60491943359375, 4.293373107910156, 13.89865493774414, -0.3011741638183594, 57.16608428955078, 20.073196411132812, 2.0495834350585938, 188.67291259765625, 77.08199310302734, 169.7351531982422, 27.187088012695312, 128.02552795410156, 170.62818908691406, 5.6807861328125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000467.npy"}
{"epoch": 0.7059712773998488, "step": 468, "batch_size": 64, "mean": 55.92658996582031, "std": 90.51982879638672, "min": -170.92617797851562, "p10": -39.814868164062474, "median": 30.521984100341797, "p90": 181.2524642944336, "max": 197.36090087890625, "pos_frac": 0.78125, "sample": [13.690345764160156, 5.284946441650391, 91.62734985351562, 183.4281768798828, 21.674922943115234, -52.383392333984375, -1.7776870727539062, 7.843029022216797, 52.43761444091797, 31.07421112060547, 172.8573455810547, 177.14471435546875, 18.97479248046875, -114.03919219970703, -0.2191314697265625, 95.05138397216797, -4.024131774902344, -10.488311767578125, -101.3428955078125, 26.715904235839844, 150.85289001464844, 7.9036865234375, 77.25053405761719, 186.58616638183594, 166.8203582763672, 146.77621459960938, 182.58541870117188, 46.98012924194336, -53.86284637451172, -170.92617797851562, 177.88421630859375, -70.96194458007812, 45.08427429199219, -8.180896759033203, 39.07655334472656, -148.29449462890625, 191.71128845214844, 25.189590454101562, -6.17364501953125, 178.60348510742188, 2.4971542358398438, 2.9346466064453125, 175.54946899414062, 139.61849975585938, 167.98452758789062, 1.6808357238769531, 7.849308013916016, 15.459403991699219, 53.587738037109375, 24.103118896484375, 182.3877410888672, 33.90638732910156, 41.04582977294922, -9.8251953125, 192.87974548339844, 29.969757080078125, 13.193374633789062, 197.36090087890625, 129.86834716796875, 170.43093872070312, 145.201171875, 74.47531127929688, 2.9960479736328125, 5.711479187011719], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000468.npy"}
{"epoch": 0.7074829931972789, "step": 469, "batch_size": 64, "mean": 51.060340881347656, "std": 74.31395721435547, "min": -102.92327880859375, "p10": -30.989360427856436, "median": 42.72084999084473, "p90": 173.32037200927735, "max": 186.6732940673828, "pos_frac": 0.703125, "sample": [-2.9209136962890625, 50.946449279785156, 76.90005493164062, 72.37763214111328, 6.9601287841796875, 42.73728561401367, -17.924163818359375, 163.99658203125, 6.6105194091796875, -83.72310638427734, 184.67620849609375, 57.498504638671875, 28.630523681640625, -1.9710159301757812, -50.632545471191406, 173.35569763183594, -2.346588134765625, 18.42462730407715, 95.86075592041016, 155.4244384765625, 177.76695251464844, 52.05390930175781, 107.6277847290039, -14.324745178222656, 101.68060302734375, -35.601287841796875, 134.47088623046875, 6.465118408203125, 74.33480834960938, 49.46299743652344, -22.377635955810547, 1.4416217803955078, 17.40081787109375, 185.56228637695312, 90.0600814819336, 37.28473663330078, -0.9833183288574219, -58.62114715576172, 175.89976501464844, 125.32466125488281, 163.84814453125, 6.383819580078125, 57.98127365112305, 26.029464721679688, -5.793792724609375, -20.753692626953125, -7.841064453125, -34.68009948730469, 186.6732940673828, 84.17822265625, 30.29460906982422, 42.70441436767578, 184.671630859375, 61.5640869140625, 173.23794555664062, -47.33417892456055, 44.532493591308594, 0.06739425659179688, 61.51277160644531, -102.92327880859375, 56.74309539794922, -8.268745422363281, -6.439056396484375, 141.66323852539062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000469.npy"}
{"epoch": 0.708994708994709, "step": 470, "batch_size": 64, "mean": 57.032936096191406, "std": 86.36649322509766, "min": -160.48721313476562, "p10": -25.807690429687497, "median": 44.34165000915527, "p90": 172.54454498291017, "max": 228.70941162109375, "pos_frac": 0.734375, "sample": [173.29092407226562, 99.06751251220703, 170.6264190673828, -160.48721313476562, 177.83438110351562, 54.71044921875, 46.445709228515625, 10.889930725097656, 149.85821533203125, 57.555145263671875, 65.75640869140625, -140.07400512695312, 180.77944946289062, 144.59298706054688, 180.0982666015625, -23.422119140625, 101.58150482177734, -8.453792572021484, -11.44158935546875, 125.45207977294922, 11.931468963623047, 70.52570343017578, 0.22723007202148438, -74.76983642578125, 22.086952209472656, 1.298095703125, 140.2996368408203, 55.412811279296875, 42.23759078979492, -2.7145843505859375, -10.231796264648438, -0.44770050048828125, 170.80299377441406, 10.051406860351562, -23.27764892578125, 10.20351791381836, 161.00794982910156, 4.85162353515625, 10.490249633789062, -11.546638488769531, 228.70941162109375, 126.37547302246094, 169.5565643310547, 23.074737548828125, -32.85474395751953, -10.992973327636719, 25.82948112487793, 152.70005798339844, 151.37533569335938, 179.37327575683594, -26.830078125, 170.00631713867188, 93.73127746582031, 66.2462158203125, -3.1126251220703125, 139.68185424804688, 10.012493133544922, 69.60147094726562, 2.8458633422851562, 102.40666198730469, 174.0032196044922, -33.96184158325195, -111.48339080810547, 0.7141494750976562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000470.npy"}
{"epoch": 0.7105064247921391, "step": 471, "batch_size": 64, "mean": 69.11367797851562, "std": 85.511474609375, "min": -200.4314727783203, "p10": -10.25310554504394, "median": 52.215538024902344, "p90": 184.19021301269532, "max": 198.52047729492188, "pos_frac": 0.8125, "sample": [-2.566946029663086, 188.6212921142578, 148.16546630859375, 24.714040756225586, 19.936378479003906, -6.3286285400390625, 120.65142822265625, 33.48424530029297, 177.68545532226562, 183.3059844970703, 12.656105041503906, 54.89306640625, 184.64520263671875, 13.615043640136719, -125.9442138671875, 1.3792448043823242, 27.61371612548828, 184.5691680908203, 102.74455261230469, -32.4344482421875, 14.471637725830078, 87.07795715332031, 106.73646545410156, 150.6439208984375, 172.52206420898438, -200.4314727783203, 49.53800964355469, 157.01025390625, 187.12600708007812, 2.9103851318359375, 123.67108154296875, 85.67971801757812, 3.96868896484375, -0.2775230407714844, 112.1912612915039, 104.99726867675781, 180.69271850585938, -60.22874450683594, 84.11680603027344, 163.38436889648438, 14.434646606445312, 30.529296875, 185.0514373779297, 178.76414489746094, 10.234071731567383, 14.879226684570312, 98.15827941894531, 45.00123977661133, -35.79692840576172, 142.27706909179688, -3.7455596923828125, 176.90501403808594, 21.57299041748047, 14.1292724609375, 137.35299682617188, -5.484073638916016, 55.3612060546875, 25.46029281616211, -16.676971435546875, 198.52047729492188, 9.263038635253906, 109.9966049194336, -11.93502426147461, 191.8155517578125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000471.npy"}
{"epoch": 0.7120181405895691, "step": 472, "batch_size": 64, "mean": 52.75737762451172, "std": 93.95744323730469, "min": -172.84445190429688, "p10": -35.09555740356445, "median": 34.823923110961914, "p90": 183.96452789306642, "max": 260.2996826171875, "pos_frac": 0.75, "sample": [0.8719635009765625, 3.86328125, 47.786781311035156, -14.670425415039062, 168.8144989013672, 145.85903930664062, -37.35003662109375, 156.02943420410156, 79.08966827392578, 30.0130615234375, 57.031288146972656, 206.67532348632812, 11.727123260498047, -16.85582733154297, 2.3420372009277344, 179.16432189941406, -13.938270568847656, 81.57649993896484, 188.15170288085938, 14.811569213867188, -19.56255340576172, 39.63478469848633, 11.429950714111328, 192.9169464111328, -7.87432861328125, 114.55364990234375, 7.165290832519531, 189.00308227539062, 186.02175903320312, -7.2844696044921875, 143.69032287597656, 126.04845428466797, 48.358055114746094, -113.28028869628906, -29.835105895996094, 49.85820007324219, 26.809768676757812, 260.2996826171875, -3.9199676513671875, -2.2957611083984375, 156.04150390625, 168.85427856445312, -172.84445190429688, 186.84054565429688, -80.42733764648438, -126.82461547851562, 173.432861328125, 14.52813720703125, -128.22549438476562, 58.69830322265625, 12.397340774536133, 1.9988765716552734, 2.5888137817382812, 14.512069702148438, 8.477569580078125, 22.88237762451172, 57.69850158691406, 55.847900390625, 157.61712646484375, 125.12165832519531, 105.80209350585938, 60.41130065917969, -137.49447631835938, 135.806640625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000472.npy"}
{"epoch": 0.7135298563869993, "step": 473, "batch_size": 64, "mean": 53.18375015258789, "std": 87.86287689208984, "min": -180.11708068847656, "p10": -38.83469390869139, "median": 39.68303108215332, "p90": 182.06712341308594, "max": 211.42169189453125, "pos_frac": 0.78125, "sample": [205.76231384277344, 99.2890625, 16.121910095214844, -139.18316650390625, -3.4163436889648438, 168.3291015625, 2.8590240478515625, 53.296485900878906, 31.459468841552734, 183.08258056640625, 154.51866149902344, 107.03746795654297, -1.9552078247070312, 98.57095336914062, 12.939308166503906, -22.807838439941406, 211.42169189453125, 188.02108764648438, 59.21165466308594, 151.80921936035156, -180.11708068847656, -45.703346252441406, 21.800750732421875, -11.866813659667969, -97.564453125, 160.60968017578125, 47.906593322753906, 54.310150146484375, -0.30683135986328125, 11.856636047363281, 18.778091430664062, -2.1954078674316406, -78.91751098632812, 80.53652954101562, -3.3431358337402344, 122.6086654663086, 188.798095703125, 179.69772338867188, 19.310131072998047, 63.93396759033203, 199.52423095703125, 170.47512817382812, 52.7379150390625, 78.68943786621094, 16.471389770507812, 97.04712677001953, 9.504938125610352, 1.4392166137695312, 77.63411712646484, 52.10480499267578, 8.075904846191406, 4.075462341308594, 108.68761444091797, 23.601913452148438, 1.0814151763916016, 23.017263412475586, 191.8253936767578, -119.86542510986328, 65.34564208984375, 7.682445526123047, 14.593513488769531, 164.59088134765625, 104.72557067871094, -75.80558776855469], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000473.npy"}
{"epoch": 0.7150415721844293, "step": 474, "batch_size": 64, "mean": 45.83870315551758, "std": 83.61859130859375, "min": -178.3271484375, "p10": -35.21809654235839, "median": 24.90569496154785, "p90": 179.30110015869144, "max": 209.20816040039062, "pos_frac": 0.75, "sample": [194.03402709960938, 89.37129974365234, 59.48527526855469, 10.12398910522461, -6.602729797363281, -14.423431396484375, 33.95555114746094, 190.68740844726562, 128.35906982421875, 35.07477569580078, 64.28034210205078, 0.6540603637695312, -178.3271484375, 51.23307800292969, 3.06524658203125, 185.95458984375, -0.7874603271484375, 8.824089050292969, 24.95376968383789, 19.6995849609375, -76.244140625, -85.55082702636719, 160.341796875, 1.6765193939208984, 169.55960083007812, 105.71968078613281, -1.4462871551513672, -4.44891357421875, 12.056648254394531, -10.426551818847656, 185.12786865234375, 4.147773742675781, -48.52269744873047, 47.493988037109375, -3.3029632568359375, 126.8722152709961, 12.367626190185547, 193.75636291503906, 119.72740173339844, 23.42974090576172, 1.4724178314208984, 11.8934326171875, 154.82810974121094, 40.46086883544922, -38.092899322509766, 90.17927551269531, 34.274017333984375, 24.857620239257812, 90.61087036132812, 1.4913215637207031, 183.4760284423828, 209.20816040039062, 15.714385986328125, 74.27645874023438, 31.52480697631836, -83.2288589477539, 158.68592834472656, -28.510223388671875, -11.886495590209961, 124.25762939453125, 129.558837890625, -144.1707763671875, 27.064804077148438, 3.78082275390625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000474.npy"}
{"epoch": 0.7165532879818595, "step": 475, "batch_size": 64, "mean": 73.72787475585938, "std": 77.5739974975586, "min": -148.23495483398438, "p10": -3.5362213134765623, "median": 69.46837615966797, "p90": 179.34135589599612, "max": 202.4755859375, "pos_frac": 0.828125, "sample": [-4.744045257568359, -2.7833404541015625, 59.265708923339844, 47.647518157958984, 170.8646697998047, 50.550498962402344, 180.44384765625, 188.24984741210938, 3.6970291137695312, -11.434268951416016, 126.60400390625, 1.7959747314453125, 140.1880645751953, 167.32611083984375, 10.026649475097656, 79.78970336914062, 137.58148193359375, 133.06373596191406, 2.324199676513672, 71.08060455322266, 68.08580780029297, -3.5748291015625, 1.3866653442382812, -3.027435302734375, 70.85094451904297, 7.007476806640625, 8.4078369140625, 109.7056884765625, 16.757522583007812, -3.446136474609375, 202.4755859375, 181.75921630859375, 105.98998260498047, 185.66580200195312, 64.85604858398438, -148.23495483398438, 118.61821746826172, 42.55973434448242, 108.86033630371094, 7.544788360595703, 194.3428955078125, 138.96002197265625, 103.75038146972656, 32.53523254394531, 125.66349029541016, -2.0554161071777344, 118.37208557128906, 41.36754608154297, 176.45587158203125, 1.658203125, 169.05404663085938, 190.50843811035156, 137.64732360839844, 117.84878540039062, 126.83949279785156, -27.2174072265625, -17.64360809326172, 158.14260864257812, 72.22645568847656, 176.7688751220703, 15.150928497314453, 1.0614204406738281, -88.07608032226562, 61.43584442138672], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000475.npy"}
{"epoch": 0.7180650037792895, "step": 476, "batch_size": 64, "mean": 50.27285385131836, "std": 82.12335205078125, "min": -161.95315551757812, "p10": -25.42253437042236, "median": 26.433183670043945, "p90": 168.48572692871093, "max": 199.75033569335938, "pos_frac": 0.671875, "sample": [138.41525268554688, 113.19046783447266, 28.463584899902344, 16.051471710205078, -7.398773193359375, -0.9701385498046875, -10.433059692382812, 169.17694091796875, 199.75033569335938, 36.12251281738281, 166.87289428710938, -1.7451648712158203, 181.2861328125, 117.6843032836914, 50.130775451660156, 7.9077911376953125, 28.081222534179688, 22.671844482421875, 120.12057495117188, -8.974609375, -41.057952880859375, 16.06275177001953, 186.067626953125, 144.1290283203125, 4.77789306640625, -161.95315551757812, -6.628929138183594, 8.391876220703125, -7.08331298828125, -6.926883697509766, 181.3204345703125, -0.7283706665039062, 14.590972900390625, 76.10969543457031, 24.785144805908203, 9.499427795410156, -148.66006469726562, -108.12857055664062, 74.43344116210938, -3.890430450439453, -9.092437744140625, 73.75494384765625, 41.422271728515625, 1.8491363525390625, 108.37129974365234, 69.55303955078125, -27.57535171508789, 102.32028198242188, -3.938385009765625, 166.29769897460938, 108.16569519042969, -32.266571044921875, 174.7252960205078, 18.54937744140625, 51.56834411621094, 190.81259155273438, -16.700180053710938, 80.98532104492188, 160.22047424316406, 91.80726623535156, -38.77492904663086, 144.89169311523438, 159.40003967285156, -20.399293899536133], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000476.npy"}
{"epoch": 0.7195767195767195, "step": 477, "batch_size": 64, "mean": 38.53519058227539, "std": 95.6515884399414, "min": -203.06536865234375, "p10": -53.78704757690429, "median": 13.332744598388672, "p90": 174.47607421875003, "max": 197.21820068359375, "pos_frac": 0.6875, "sample": [-20.596527099609375, -47.28717803955078, 176.2891845703125, 1.0159912109375, 2.6209564208984375, 11.816070556640625, -1.6151123046875, -40.70799255371094, 8.84014892578125, -5.463371276855469, 131.83474731445312, 148.103271484375, -34.67908477783203, 6.302219390869141, -25.74087142944336, 1.6597251892089844, -4.205047607421875, 163.87423706054688, 41.04486083984375, 3.4918079376220703, 128.08892822265625, 80.62776184082031, -3.6331138610839844, 20.768211364746094, 167.22265625, -51.964698791503906, -178.82351684570312, 5.611335754394531, 177.1032257080078, 14.531135559082031, -54.56805419921875, -14.221481323242188, 19.989871978759766, 3.2106781005859375, -177.70077514648438, -50.1612548828125, 4.8793182373046875, 170.2454833984375, 105.44817352294922, -9.092449188232422, 119.04933166503906, -86.50081634521484, 110.19915771484375, 192.08340454101562, 22.10564422607422, 180.44943237304688, -149.0360565185547, 29.686004638671875, 70.18199157714844, -60.425201416015625, 130.24591064453125, 76.23417663574219, 9.756423950195312, 182.53591918945312, 169.4521026611328, 64.85107421875, -203.06536865234375, 30.621337890625, 121.50682067871094, 197.21820068359375, 32.6885871887207, 180.86895751953125, 12.134353637695312, 159.2513427734375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000477.npy"}
{"epoch": 0.7210884353741497, "step": 478, "batch_size": 64, "mean": 41.69120788574219, "std": 95.63334655761719, "min": -181.1981201171875, "p10": -71.07748985290525, "median": 17.477063179016113, "p90": 178.46248474121097, "max": 197.06729125976562, "pos_frac": 0.65625, "sample": [-34.36296844482422, 187.4339599609375, 84.28216552734375, 172.3710479736328, 125.04431915283203, -17.00414276123047, 29.426849365234375, 40.77803039550781, -11.073684692382812, 172.5771484375, -2.3830642700195312, 2.031768798828125, 161.118408203125, 193.04165649414062, -84.90121459960938, 2.374919891357422, -170.07748413085938, 4.353607177734375, -2.0401268005371094, -0.7407455444335938, 2.1368484497070312, -127.46726989746094, 167.7489013671875, 137.83419799804688, 118.36088562011719, -95.29998779296875, 14.217201232910156, 99.33154296875, 34.93212127685547, -92.22026062011719, 75.6599349975586, 0.6350440979003906, -148.91030883789062, -33.865535736083984, 7.9463653564453125, 171.43817138671875, 158.3380889892578, -13.44732666015625, 25.642250061035156, 14.922760009765625, 187.08592224121094, -22.882755279541016, -181.1981201171875, -14.45339584350586, 189.40577697753906, -37.093875885009766, -0.23701858520507812, 197.06729125976562, 3.7004318237304688, 48.238861083984375, 129.5035858154297, 148.08370971679688, 0.6010208129882812, 72.12190246582031, 96.47245788574219, 190.93478393554688, 180.98477172851562, -18.283226013183594, 41.12135314941406, 67.25138092041016, -36.32654571533203, -38.8221321105957, 74.74565887451172, 20.0313663482666], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000478.npy"}
{"epoch": 0.7226001511715797, "step": 479, "batch_size": 64, "mean": 41.721351623535156, "std": 95.27742004394531, "min": -195.44732666015625, "p10": -101.80924072265621, "median": 25.116172790527344, "p90": 176.51039886474612, "max": 199.28900146484375, "pos_frac": 0.703125, "sample": [8.150737762451172, 26.826797485351562, 37.98567199707031, -128.77957153320312, -21.804637908935547, 21.274871826171875, 182.65142822265625, 129.19979858398438, 157.57373046875, 40.289955139160156, 45.96809768676758, -135.26370239257812, -1.2519607543945312, 2.180980682373047, 131.9068603515625, 94.66920471191406, 199.28900146484375, 71.37113952636719, -23.503154754638672, 6.678974151611328, -31.27922821044922, 84.48388671875, 149.92242431640625, 190.00550842285156, -9.903968811035156, -171.65325927734375, 9.396228790283203, 23.405548095703125, -64.70556640625, 6.793235778808594, 16.321990966796875, 142.10643005371094, -2.1031532287597656, 139.469482421875, 14.328933715820312, 19.805883407592773, -120.15436553955078, 125.28633880615234, 3.1126632690429688, 181.10955810546875, 74.76253509521484, -195.44732666015625, -7.14801025390625, 105.29499816894531, 119.50019836425781, 182.3991241455078, 48.69468307495117, -142.06607055664062, -12.273719787597656, 39.437156677246094, 10.477432250976562, 94.49617004394531, -117.7108154296875, 188.622802734375, -3.697122573852539, -0.3142051696777344, 172.7874298095703, 45.64060592651367, 74.96420288085938, 120.88716125488281, 0.3419628143310547, -30.04851531982422, 178.10595703125, 171.29714965820312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000479.npy"}
{"epoch": 0.7241118669690099, "step": 480, "batch_size": 64, "mean": 57.996116638183594, "std": 90.99842071533203, "min": -187.83297729492188, "p10": -40.31025733947754, "median": 47.80220031738281, "p90": 173.94383544921877, "max": 189.89706420898438, "pos_frac": 0.703125, "sample": [121.88954162597656, -0.683380126953125, 171.40676879882812, -41.217559814453125, 15.44059944152832, 1.7718658447265625, 172.38613891601562, -187.83297729492188, -2.7622299194335938, 148.943115234375, -38.19321823120117, -4.799751281738281, 136.5903778076172, -78.45172882080078, 176.47760009765625, -64.02777099609375, -12.040664672851562, 62.23753356933594, -6.177558898925781, 2.5295982360839844, -27.229923248291016, 158.65399169921875, 119.24325561523438, 4.304931640625, 152.086181640625, 127.56253051757812, 187.16783142089844, 140.0485382080078, 189.89706420898438, 26.80370330810547, 104.41598510742188, -0.4811573028564453, -0.2940521240234375, 69.95693969726562, 70.20237731933594, -17.93193817138672, 178.82769775390625, 174.61141967773438, -132.34124755859375, 171.47698974609375, 164.2885284423828, 113.23155212402344, 79.80802917480469, 110.98809814453125, -55.84954833984375, 169.5120391845703, 76.49848175048828, 45.23933410644531, 15.9046630859375, 50.36506652832031, 2.318187713623047, -31.937759399414062, 181.05343627929688, 70.13616943359375, 155.31178283691406, 18.676042556762695, 183.85548400878906, 153.91761779785156, 26.63424301147461, -135.8294219970703, 26.341617584228516, 42.49895477294922, 13.947486877441406, -35.62617492675781], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000480.npy"}
{"epoch": 0.7256235827664399, "step": 481, "batch_size": 64, "mean": 50.558876037597656, "std": 88.26538848876953, "min": -179.74658203125, "p10": -36.61605949401855, "median": 23.601917266845703, "p90": 175.83816528320312, "max": 236.2310791015625, "pos_frac": 0.71875, "sample": [169.33920288085938, -7.275840759277344, 186.34774780273438, -11.904495239257812, -41.202369689941406, 143.74533081054688, 125.62232208251953, -37.20468521118164, -8.056648254394531, -16.068614959716797, 176.93734741210938, 1.9495086669921875, 98.03684997558594, 164.87118530273438, -1.2122726440429688, 67.982177734375, 139.29537963867188, 25.17682647705078, 105.69389343261719, 74.13037109375, -179.74658203125, 10.243690490722656, 164.82859802246094, 72.69490814208984, 0.3658180236816406, -77.09356689453125, 47.42655944824219, 159.0443878173828, -0.06918716430664062, 91.45597076416016, 4.829174041748047, 34.876956939697266, 38.231712341308594, -35.24259948730469, 22.027008056640625, -1.276153564453125, 70.68165588378906, 18.44499969482422, 7.8445281982421875, 4.2428741455078125, 16.928466796875, 94.8321533203125, 173.27340698242188, 196.80035400390625, 10.060983657836914, 179.5900115966797, -31.488922119140625, -0.8407211303710938, 193.23451232910156, 108.50714874267578, 137.69346618652344, 3.4752044677734375, 19.409835815429688, -43.572845458984375, -124.34053802490234, 236.2310791015625, -146.6846160888672, 15.84185791015625, 45.01894760131836, 206.73602294921875, 96.17570495605469, 1.4024810791015625, -21.718170166015625, 59.18824768066406], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000481.npy"}
{"epoch": 0.72713529856387, "step": 482, "batch_size": 64, "mean": 50.410911560058594, "std": 100.51805877685547, "min": -189.80633544921875, "p10": -50.50357246398925, "median": 18.204498291015625, "p90": 182.8265396118164, "max": 251.399169921875, "pos_frac": 0.65625, "sample": [-25.15252685546875, -189.80633544921875, -15.163238525390625, 13.021614074707031, 12.6240234375, -7.804939270019531, 6.314697265625, -24.02007293701172, 6.625274658203125, 201.32818603515625, 78.37457275390625, 148.1552276611328, 152.20025634765625, 9.621871948242188, 137.8077392578125, 89.22479248046875, 159.41339111328125, 10.56820297241211, -179.09083557128906, 113.9893569946289, 187.55686950683594, -1.8423614501953125, 236.44119262695312, -4.18896484375, -1.5758514404296875, 97.70654296875, 179.11090087890625, 143.10934448242188, -7.126033782958984, 26.618877410888672, -53.8399658203125, 140.31455993652344, -42.71865463256836, 79.60157775878906, 183.20223999023438, 185.92315673828125, 147.901123046875, 13.625808715820312, -17.864425659179688, -23.39630126953125, 251.399169921875, 72.9256591796875, 9.2928466796875, 22.783187866210938, 69.60625457763672, 181.9499053955078, 114.25721740722656, 33.725215911865234, 197.12033081054688, -13.190170288085938, 119.57984924316406, 9.41943359375, 137.0523681640625, -1.206756591796875, 139.91543579101562, 1.0870361328125, -149.32749938964844, 181.19998168945312, -8.863494873046875, -134.0406494140625, 27.599937438964844, -18.113388061523438, -122.14987182617188, -62.514366149902344], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000482.npy"}
{"epoch": 0.7286470143613001, "step": 483, "batch_size": 64, "mean": 58.67215347290039, "std": 87.44818115234375, "min": -185.49769592285156, "p10": -14.711886596679687, "median": 33.66682243347168, "p90": 178.37866668701173, "max": 210.71104431152344, "pos_frac": 0.796875, "sample": [4.031333923339844, 127.07550048828125, -5.662361145019531, 145.28225708007812, 173.46646118164062, 178.76963806152344, 174.67617797851562, -15.075103759765625, -28.66729736328125, -1.3895721435546875, 6.651157379150391, 89.20249938964844, 56.56298828125, 197.21290588378906, 0.11371612548828125, 95.01272583007812, -59.07315444946289, -69.05464172363281, 3.9594287872314453, 182.9149169921875, 59.58343505859375, 182.9521484375, 132.9687042236328, -20.910537719726562, 177.46640014648438, 6.47332763671875, 182.0684814453125, 110.1409912109375, 0.356109619140625, 66.19450378417969, 174.1519012451172, -179.9893798828125, 175.3520965576172, 2.567689895629883, 16.384796142578125, 140.006591796875, 84.4024887084961, -185.49769592285156, 1.2129096984863281, -8.971824645996094, 5.9792327880859375, 41.93247604370117, 17.35230255126953, 14.976692199707031, -4.863040924072266, 73.7255859375, 210.71104431152344, 190.0642547607422, 143.7230224609375, 102.22776794433594, 94.94210052490234, 6.77642822265625, 17.430694580078125, 177.13363647460938, -5.453733444213867, -13.8643798828125, 20.20928955078125, 8.862934112548828, 135.10174560546875, 56.73442077636719, 1.1059341430664062, 25.401168823242188, 1.1699790954589844, 60.715362548828125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000483.npy"}
{"epoch": 0.7301587301587301, "step": 484, "batch_size": 64, "mean": 41.58033752441406, "std": 77.31977844238281, "min": -162.3724822998047, "p10": -19.708624267578124, "median": 13.114978790283203, "p90": 152.04169464111328, "max": 194.34335327148438, "pos_frac": 0.671875, "sample": [151.6515655517578, -19.68328857421875, -5.615562438964844, -1.9852180480957031, 183.56875610351562, 3.4332809448242188, 123.6303482055664, -162.3724822998047, 176.32713317871094, -12.874794006347656, 2.6391754150390625, -3.8461380004882812, -23.994403839111328, -5.391609191894531, 85.00775146484375, 14.811386108398438, -95.90292358398438, 16.870868682861328, -7.539886474609375, 176.64627075195312, -0.7170028686523438, 106.6922836303711, 145.3460693359375, 45.42334747314453, 16.984025955200195, 152.20889282226562, 91.65826416015625, 144.48023986816406, 107.35379028320312, 111.47430419921875, -4.353315353393555, 11.266460418701172, -2.1112213134765625, 35.392616271972656, 185.81997680664062, -2.3600997924804688, 0.05426025390625, 0.7784996032714844, 56.94481658935547, 7.558597564697266, 120.3950424194336, -2.8030776977539062, 189.09368896484375, -19.719482421875, 12.061893463134766, 12.814781188964844, 127.4216079711914, -38.382389068603516, 58.46418762207031, 77.044189453125, 194.34335327148438, 1.9087066650390625, 41.868309020996094, 13.415176391601562, -60.762664794921875, -137.7726593017578, 43.919532775878906, -2.9251251220703125, 0.4695758819580078, 39.41151428222656, 8.757293701171875, 78.4566879272461, 116.1129150390625, -17.726593017578125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000484.npy"}
{"epoch": 0.7316704459561603, "step": 485, "batch_size": 64, "mean": 46.33942413330078, "std": 79.59188842773438, "min": -179.71710205078125, "p10": -20.66305999755859, "median": 18.976619720458984, "p90": 168.24162902832032, "max": 204.72515869140625, "pos_frac": 0.734375, "sample": [60.053016662597656, -0.3837127685546875, 122.13678741455078, 127.07379913330078, 83.95719909667969, -2.4985122680664062, 33.450923919677734, -68.5853042602539, 19.681106567382812, -45.586090087890625, -179.71710205078125, 123.08533477783203, 187.42636108398438, 162.7449951171875, 12.145721435546875, 40.25775146484375, 51.85624313354492, 37.815093994140625, 174.56698608398438, 7.092052459716797, 51.78272247314453, 15.20025634765625, -52.001014709472656, 139.98052978515625, 25.683887481689453, 2.2057647705078125, -161.3001708984375, 98.84635162353516, 0.31982421875, 87.74701690673828, -23.361465454101562, 16.846603393554688, 77.187744140625, -1.9797286987304688, 88.92184448242188, -10.571624755859375, 204.72515869140625, 188.22824096679688, 11.240219116210938, 165.96929931640625, 18.272132873535156, -12.001922607421875, 3.933990478515625, -4.779472351074219, 6.837802886962891, -4.695316314697266, -0.9641189575195312, 169.21548461914062, 17.946056365966797, -19.577438354492188, 69.95573425292969, 10.982646942138672, 115.09075164794922, 185.2121124267578, -0.5881729125976562, 132.83621215820312, 142.4090118408203, 183.35104370117188, 13.513664245605469, -21.128326416015625, 49.625640869140625, 2.793304443359375, 2.9408111572265625, 32.297454833984375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000485.npy"}
{"epoch": 0.7331821617535903, "step": 486, "batch_size": 64, "mean": 48.51192092895508, "std": 93.84709930419922, "min": -188.5220947265625, "p10": -23.829610633850095, "median": 23.693538665771484, "p90": 188.02049865722657, "max": 228.49697875976562, "pos_frac": 0.71875, "sample": [73.60314178466797, 4.367240905761719, 204.62060546875, -2.1995201110839844, 1.9422988891601562, -2.391448974609375, 17.924835205078125, 173.74667358398438, 24.097244262695312, 8.550498962402344, -17.19961929321289, 123.80116271972656, 135.0522918701172, 56.69599914550781, 190.888916015625, 4.796775817871094, -4.710227966308594, 190.0027618408203, -1.7902545928955078, 169.95401000976562, 50.56047058105469, 44.79762268066406, -39.59745788574219, 1.0874404907226562, 190.85247802734375, -111.68899536132812, 32.383811950683594, -17.85879898071289, 48.55185317993164, -188.5220947265625, 193.56275939941406, 180.25778198242188, 14.58355712890625, -150.34136962890625, 2.440196990966797, 121.21465301513672, -12.816823959350586, 126.54844665527344, -75.05525970458984, -0.18628692626953125, 35.6218376159668, 166.7149200439453, 6.79193115234375, 186.96998596191406, 174.12265014648438, 9.311233520507812, 26.09613037109375, 129.81362915039062, 162.071533203125, 188.47071838378906, 49.5887451171875, 19.94457244873047, 11.344375610351562, 32.0933837890625, -21.945886611938477, -3.77545166015625, 1.6154212951660156, 44.83216857910156, 228.49697875976562, -24.636920928955078, 82.88209533691406, -181.910888671875, 23.289833068847656, -5.5674591064453125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000486.npy"}
{"epoch": 0.7346938775510204, "step": 487, "batch_size": 64, "mean": 52.0731201171875, "std": 83.12647247314453, "min": -154.57852172851562, "p10": -23.893676757812496, "median": 23.438507080078125, "p90": 184.34945220947267, "max": 221.35568237304688, "pos_frac": 0.765625, "sample": [-6.001861572265625, -74.76972961425781, 113.71527862548828, 181.87091064453125, 95.69957733154297, 4.764764785766602, -13.4820556640625, 54.76496505737305, 93.41224670410156, 135.67478942871094, 7.51953125, -25.46739959716797, 15.330093383789062, -75.43678283691406, -6.6496124267578125, 188.71231079101562, 24.536983489990234, 22.340030670166016, 212.284423828125, 70.5682373046875, -9.431930541992188, 36.718902587890625, 27.48833465576172, 182.50448608398438, 121.71467590332031, 0.96044921875, -10.31032943725586, -40.804107666015625, 92.81163787841797, 19.79302978515625, 163.89230346679688, 9.683303833007812, 58.56114959716797, -2.475139617919922, 189.70526123046875, 62.97618865966797, 144.25689697265625, -20.221656799316406, 149.1172637939453, 59.98373794555664, 21.022247314453125, 35.97540283203125, -122.05950927734375, 138.52464294433594, -8.6065673828125, 60.00090026855469, 21.10120391845703, 5.986900329589844, 1.4901580810546875, -154.57852172851562, 0.6683464050292969, 185.14015197753906, 9.267677307128906, -40.125877380371094, 18.470596313476562, 136.06707763671875, 1.9766502380371094, 43.198455810546875, 4.9941864013671875, 188.1715087890625, 189.05137634277344, 108.98345947265625, 221.35568237304688, 10.292449951171875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000487.npy"}
{"epoch": 0.7362055933484505, "step": 488, "batch_size": 64, "mean": 51.10912322998047, "std": 96.90269470214844, "min": -181.07583618164062, "p10": -58.616408538818355, "median": 25.44439125061035, "p90": 184.3613037109375, "max": 203.86875915527344, "pos_frac": 0.65625, "sample": [-42.2740478515625, -84.98678588867188, -32.93767166137695, -15.79629135131836, 61.92362976074219, -33.066680908203125, -181.07583618164062, 176.9732666015625, 184.75833129882812, 180.2156219482422, 1.0143280029296875, -33.903411865234375, 183.43490600585938, 13.85198974609375, 10.82916259765625, 195.82989501953125, -0.3474578857421875, 109.92987823486328, 20.18951416015625, -2.9986839294433594, 32.09870147705078, 34.29370880126953, 21.468353271484375, 2.43743896484375, -96.03079223632812, 12.232589721679688, -79.8046875, 180.0669708251953, -7.1305084228515625, 189.37428283691406, -8.231201171875, -92.22470092773438, 18.317638397216797, 181.77117919921875, 41.48625183105469, 24.85599136352539, 180.0633544921875, -8.88395881652832, 196.395751953125, 203.86875915527344, 185.15457153320312, -43.01531219482422, 173.93992614746094, -60.102508544921875, 173.97227478027344, 68.18035888671875, 1.8130111694335938, 49.579681396484375, -113.53146362304688, 162.7684326171875, -3.0238494873046875, 97.14373779296875, -13.988815307617188, 45.15516662597656, 30.612648010253906, 154.10145568847656, 96.49569702148438, 174.1001434326172, 170.63327026367188, -55.148841857910156, 26.032791137695312, 27.528488159179688, -5.533203125, 190.12759399414062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000488.npy"}
{"epoch": 0.7377173091458806, "step": 489, "batch_size": 64, "mean": 24.302478790283203, "std": 78.74761962890625, "min": -153.33868408203125, "p10": -74.2264633178711, "median": 8.350168228149414, "p90": 169.25046844482424, "max": 199.2371826171875, "pos_frac": 0.625, "sample": [-98.14115905761719, 20.392868041992188, 11.612281799316406, 166.23709106445312, 186.1873321533203, 176.6761474609375, -8.520637512207031, 101.20924377441406, 57.484046936035156, 199.2371826171875, 18.86334991455078, -72.82376098632812, 18.5703125, -153.33868408203125, -1.6224021911621094, -111.98609924316406, -8.098297119140625, -14.474418640136719, 15.59912109375, 7.499351501464844, 71.68838500976562, 20.960205078125, 103.02079010009766, 7.17156982421875, 86.6028823852539, 176.88116455078125, -2.1729164123535156, 13.907516479492188, -14.176239013671875, 47.45372772216797, -55.7430419921875, -94.82679748535156, 170.5419158935547, 15.076107025146484, 1.8427200317382812, -77.71871948242188, -12.385917663574219, 75.86294555664062, 46.15428924560547, 8.513416290283203, 16.31708526611328, 35.745574951171875, 36.69728088378906, 88.21078491210938, 56.810760498046875, -1.8311233520507812, 172.16030883789062, 6.22467041015625, 114.9373779296875, 7.882293701171875, -74.82762145996094, 3.4594650268554688, -3.1005096435546875, -0.7931690216064453, -0.9992637634277344, -141.2846221923828, 71.62680053710938, -6.1511688232421875, -27.751632690429688, 2.8519515991210938, -37.545108795166016, 173.18917846679688, -43.87446975708008, 8.186920166015625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000489.npy"}
{"epoch": 0.7392290249433107, "step": 490, "batch_size": 64, "mean": 54.10951232910156, "std": 93.9665756225586, "min": -175.83045959472656, "p10": -27.07459831237793, "median": 35.551002502441406, "p90": 183.27416381835937, "max": 198.13076782226562, "pos_frac": 0.6875, "sample": [-5.639396667480469, -27.243759155273438, -120.50625610351562, 14.198379516601562, -5.218486785888672, 65.7053451538086, 34.87733459472656, 86.00669860839844, 198.13076782226562, 191.00448608398438, -114.33841705322266, 189.35134887695312, -137.78436279296875, 1.6207504272460938, 123.5624771118164, 40.00726318359375, 158.94207763671875, 181.38169860839844, 167.64767456054688, -1.3001861572265625, 71.87911987304688, 145.2404327392578, 164.64508056640625, 164.12643432617188, -24.733291625976562, -6.466766357421875, 166.68087768554688, -1.9487543106079102, 63.22819519042969, 3.6889190673828125, -25.4744873046875, 7.7320098876953125, 9.061187744140625, 133.07699584960938, 3.4825477600097656, 28.944412231445312, 2.1089401245117188, 192.6861114501953, -140.07415771484375, 30.78277587890625, 172.98204040527344, 36.22467041015625, -65.6572265625, 184.08522033691406, 19.56900405883789, 36.725914001464844, 7.177955627441406, -2.0822792053222656, -13.517120361328125, -4.860466003417969, 196.33050537109375, 151.05239868164062, -1.3904037475585938, 117.47273254394531, 195.03860473632812, 115.04571533203125, 71.55580139160156, 115.8858413696289, -3.8271751403808594, 174.02169799804688, -175.83045959472656, -26.679889678955078, 70.84312438964844, 63.770835876464844], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000490.npy"}
{"epoch": 0.7407407407407407, "step": 491, "batch_size": 64, "mean": 53.826053619384766, "std": 79.62147521972656, "min": -128.8363037109375, "p10": -25.152639770507808, "median": 24.41572380065918, "p90": 176.34928131103516, "max": 213.94366455078125, "pos_frac": 0.734375, "sample": [190.79400634765625, -28.28522491455078, 39.556907653808594, 5.978599548339844, 85.98518371582031, 90.58882141113281, 25.70681381225586, 27.554306030273438, 16.331912994384766, 45.43730545043945, 9.964218139648438, -5.6269073486328125, 159.37001037597656, -0.7659988403320312, 170.12515258789062, -27.474288940429688, 20.178756713867188, 118.6895751953125, -1.7424392700195312, 124.16618347167969, 28.540817260742188, -8.4749755859375, 172.78407287597656, 0.587677001953125, -61.2030029296875, 84.14689636230469, 3.1176605224609375, 13.835678100585938, -19.735458374023438, 6.181602478027344, 187.78123474121094, 101.35763549804688, -16.842086791992188, 55.088287353515625, 173.0555877685547, 137.23233032226562, 7.8965911865234375, 213.94366455078125, 23.1246337890625, -73.84585571289062, 185.9661865234375, -0.28882598876953125, 1.8002243041992188, 12.300493240356445, 151.9034881591797, 82.92036437988281, -5.667205810546875, -3.747528076171875, 138.272216796875, 170.88375854492188, -28.2252197265625, 28.25696563720703, 177.7608642578125, 183.09561157226562, 112.7055435180664, 18.314849853515625, 15.089656829833984, 67.07246398925781, 51.5361213684082, -128.8363037109375, -61.08128356933594, 0.7392997741699219, -9.663185119628906, 188.65289306640625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000491.npy"}
{"epoch": 0.7422524565381708, "step": 492, "batch_size": 64, "mean": 36.33604431152344, "std": 86.40914916992188, "min": -155.98519897460938, "p10": -50.515120315551755, "median": 15.636575698852539, "p90": 173.41317443847657, "max": 206.3406982421875, "pos_frac": 0.625, "sample": [206.3406982421875, 189.2835693359375, -0.6332130432128906, 1.646627426147461, 127.74828338623047, -10.783355712890625, 172.61610412597656, -50.666080474853516, 20.360733032226562, -23.863880157470703, 41.11390686035156, 58.01667785644531, 162.7309112548828, -10.806266784667969, 20.853797912597656, 28.6312255859375, -50.162879943847656, 71.04167175292969, 0.9669399261474609, 24.37940216064453, 40.74586868286133, -76.68869018554688, -32.486572265625, -45.58177947998047, 30.528106689453125, 110.9608154296875, 83.37934875488281, 193.564208984375, -40.16508483886719, -114.17987823486328, 185.34457397460938, 7.878364562988281, 146.1673583984375, -1.1927719116210938, -108.78228759765625, 26.113624572753906, -155.98519897460938, 186.75045776367188, -29.926605224609375, -17.961181640625, 118.62461853027344, 121.46774291992188, 7.957489013671875, 3.4882431030273438, -4.703826904296875, 86.31442260742188, 101.92393493652344, 6.422698974609375, 3.7288970947265625, 74.12542724609375, -33.00891876220703, -124.43488311767578, -23.232017517089844, 88.66381072998047, 173.75477600097656, 10.912418365478516, 169.1251983642578, -71.15751647949219, 69.15505981445312, -16.553009033203125, 34.100494384765625, -7.444820404052734, -14.440940856933594, 183.42007446289062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000492.npy"}
{"epoch": 0.7437641723356009, "step": 493, "batch_size": 64, "mean": 48.00779724121094, "std": 95.20880889892578, "min": -177.85198974609375, "p10": -77.39762268066406, "median": 41.58210754394531, "p90": 183.45252380371096, "max": 195.028076171875, "pos_frac": 0.765625, "sample": [38.729427337646484, 7.279178619384766, -78.98408508300781, 137.0159149169922, 9.480567932128906, 22.097213745117188, -171.95071411132812, 0.19727325439453125, 47.27922439575195, -54.3314323425293, 34.639984130859375, 187.28512573242188, 68.48635864257812, 186.1639404296875, 44.43478775024414, -90.61874389648438, 139.9935302734375, 0.23547935485839844, 134.25233459472656, 137.20318603515625, 1.8818645477294922, 9.143775939941406, -40.879180908203125, -2.60595703125, -5.95648193359375, -73.69587707519531, 188.3984375, 5.297977447509766, 192.55441284179688, 67.40241241455078, -69.92977905273438, 77.6560287475586, -92.81338500976562, 63.679664611816406, -122.06246948242188, 119.23158264160156, 168.83639526367188, -28.073890686035156, 142.0711669921875, 0.23909378051757812, 6.1532440185546875, 174.05978393554688, 87.2750015258789, 73.74102783203125, 12.009239196777344, 191.87403869628906, -109.72198486328125, 162.57073974609375, 177.12588500976562, -60.71781921386719, 3.3857650756835938, 170.14471435546875, 27.6605224609375, 52.09571838378906, 192.2340087890625, 107.9471435546875, 107.55227661132812, 5.87457275390625, 65.86813354492188, 195.028076171875, 37.39567565917969, 73.653564453125, 97.87749481201172, -177.85198974609375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000493.npy"}
{"epoch": 0.745275888133031, "step": 494, "batch_size": 64, "mean": 43.766326904296875, "std": 93.06819915771484, "min": -187.8507080078125, "p10": -30.792031097412107, "median": 18.810569763183594, "p90": 179.27427825927734, "max": 212.29498291015625, "pos_frac": 0.671875, "sample": [145.4261474609375, 3.9376907348632812, -1.972747802734375, -12.298141479492188, 158.26910400390625, -0.7720489501953125, 57.82177734375, 156.30712890625, 40.722530364990234, -1.3847198486328125, -4.206779479980469, 0.7691612243652344, 0.2669525146484375, 8.205673217773438, 9.870861053466797, -1.2509841918945312, 179.53851318359375, -92.58415985107422, 54.228240966796875, -22.26434326171875, 6.5911102294921875, 212.29498291015625, 83.73101806640625, 41.6693115234375, 21.09747314453125, 178.65773010253906, 40.61791229248047, -26.428585052490234, 161.23239135742188, 173.23497009277344, 109.95195770263672, -177.72543334960938, -150.77035522460938, 111.33370971679688, -0.27999114990234375, -97.82972717285156, -29.157691955566406, 150.48455810546875, 16.003395080566406, 24.072315216064453, 28.681869506835938, -10.339836120605469, -187.8507080078125, -8.833641052246094, 22.885181427001953, 81.62934875488281, 190.90896606445312, 4.551429748535156, -6.304389953613281, 18.00372314453125, 172.22830200195312, 19.617416381835938, 97.49664306640625, 200.0290069580078, 9.390853881835938, 160.26834106445312, -31.492462158203125, -23.355350494384766, 194.9365234375, 28.1461181640625, 182.58779907226562, -67.12092590332031, 9.110969543457031, 188.45887756347656], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000494.npy"}
{"epoch": 0.7467876039304611, "step": 495, "batch_size": 64, "mean": 44.22233581542969, "std": 93.25865173339844, "min": -181.74020385742188, "p10": -59.97677307128905, "median": 22.759769439697266, "p90": 180.77564544677736, "max": 222.8406982421875, "pos_frac": 0.671875, "sample": [-181.74020385742188, -45.681602478027344, -88.91120910644531, -177.28860473632812, 184.60006713867188, -9.084419250488281, -114.87794494628906, 19.1358642578125, 59.94677734375, 110.78498840332031, -64.9566650390625, -48.357025146484375, -2.0418853759765625, 81.30599975585938, 124.39173889160156, 172.61874389648438, -102.48246002197266, 115.24903869628906, 184.29689025878906, -0.0140228271484375, 19.209125518798828, 36.769439697265625, 8.129196166992188, 116.0113754272461, 28.13065528869629, 178.923095703125, -4.281658172607422, -2.448089599609375, 21.238021850585938, 121.73297119140625, 74.03985595703125, 19.559711456298828, 160.95086669921875, -36.636512756347656, 24.281517028808594, -45.46002197265625, 78.05242156982422, 71.34623718261719, 220.0635986328125, 171.58876037597656, -110.77923583984375, 15.473930358886719, 102.4362564086914, 48.4759521484375, 181.56959533691406, 107.8881607055664, 27.730270385742188, -0.06463623046875, -24.513626098632812, 192.8035125732422, -5.3347015380859375, 150.37269592285156, 1.6946868896484375, 48.82415008544922, 1.04486083984375, 222.8406982421875, 65.82936096191406, -16.981651306152344, -3.9682846069335938, 18.229843139648438, 115.29318237304688, 6.049541473388672, 194.20860290527344, 13.011512756347656], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000495.npy"}
{"epoch": 0.7482993197278912, "step": 496, "batch_size": 64, "mean": 50.637855529785156, "std": 90.99951171875, "min": -176.47894287109375, "p10": -51.55950813293457, "median": 35.52454376220703, "p90": 182.97374420166017, "max": 239.9857177734375, "pos_frac": 0.734375, "sample": [53.49983215332031, 85.76303100585938, 197.91346740722656, 174.91226196289062, -47.63768005371094, 27.87358856201172, -53.240291595458984, 239.9857177734375, 28.350784301757812, 83.95706176757812, -32.07465744018555, 18.81188201904297, 75.42726135253906, 97.22677612304688, 6.212215423583984, 32.4378662109375, 6.512889862060547, 112.45401000976562, 180.39756774902344, 57.75054931640625, -32.66053009033203, 63.712249755859375, 0.0057315826416015625, 171.6024932861328, 43.218101501464844, -54.33294677734375, 214.85604858398438, 65.93644714355469, -0.7236709594726562, 88.22795104980469, -153.90406799316406, 4.63824462890625, 187.66917419433594, 193.58865356445312, 0.9824371337890625, 7.479274749755859, -135.3942413330078, 10.627079010009766, 13.2496337890625, -29.151260375976562, 8.179290771484375, 156.8450927734375, 184.07781982421875, 99.66781616210938, 26.2496337890625, 54.553794860839844, 76.79104614257812, 85.66344451904297, -39.82493209838867, 38.61122131347656, 193.771484375, -176.47894287109375, -17.283981323242188, 4.933172225952148, 160.25802612304688, 139.74862670898438, 103.78174591064453, 163.1315155029297, -55.53839874267578, -1.1872787475585938, -54.890533447265625, 125.30689239501953, -31.648765563964844, -10.055740356445312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000496.npy"}
{"epoch": 0.7498110355253212, "step": 497, "batch_size": 64, "mean": 58.869712829589844, "std": 80.69425964355469, "min": -163.42428588867188, "p10": -12.649478149414062, "median": 34.29178810119629, "p90": 171.37969665527345, "max": 203.1080780029297, "pos_frac": 0.78125, "sample": [191.4551544189453, 5.823444366455078, 203.1080780029297, 166.30526733398438, 158.72015380859375, 5.385414123535156, 103.38339233398438, 36.23490524291992, 1.7584972381591797, 7.4672698974609375, 29.421762466430664, 140.44882202148438, 155.32936096191406, -12.274009704589844, 3.667144775390625, 7.196220397949219, 132.92083740234375, -163.42428588867188, 107.30729675292969, 32.348670959472656, -4.518959045410156, 170.08595275878906, 72.22935485839844, 2.1990509033203125, -3.5380859375, -12.428581237792969, -33.924652099609375, 119.90974426269531, 39.15752410888672, 106.79132080078125, -2.4489498138427734, 54.59185791015625, 31.536239624023438, 139.57199096679688, -12.744148254394531, 169.5162811279297, 191.91680908203125, 132.32862854003906, 16.85778045654297, -137.0807647705078, 171.9341583251953, 15.973464965820312, 45.4227180480957, 39.988525390625, 14.599639892578125, -5.579914093017578, 194.05499267578125, 192.72994995117188, -26.24645233154297, 50.46174621582031, 24.045799255371094, 38.295440673828125, 31.396873474121094, 92.88710021972656, 94.52548217773438, 187.90731811523438, -40.37628173828125, 18.163570404052734, -1.8381881713867188, 122.44223022460938, 19.950149536132812, 139.57423400878906, -21.260665893554688, 16.017974853515625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000497.npy"}
{"epoch": 0.7513227513227513, "step": 498, "batch_size": 64, "mean": 53.72976303100586, "std": 85.15476989746094, "min": -206.02175903320312, "p10": -16.193793773651123, "median": 22.94260025024414, "p90": 186.18101959228517, "max": 207.4112548828125, "pos_frac": 0.75, "sample": [-94.088134765625, 56.716915130615234, 159.9085235595703, 44.10688018798828, 40.994171142578125, -19.495162963867188, 69.12871551513672, 106.3162841796875, 29.717731475830078, -15.461273193359375, 177.43572998046875, 13.508018493652344, 2.685577392578125, 1.8222732543945312, 180.5222625732422, 50.29651641845703, 1.7209320068359375, 13.790931701660156, 0.7555046081542969, 57.26163864135742, 207.4112548828125, 2.045654296875, 177.8155517578125, -13.039722442626953, 18.796592712402344, 43.0224609375, -5.6778564453125, 95.2845687866211, 57.37504577636719, -5.590446472167969, 148.63461303710938, 176.45654296875, 2.7230052947998047, 4.1104583740234375, -1.90899658203125, -41.532073974609375, -4.663154602050781, 158.09890747070312, 0.29093170166015625, 88.95272827148438, 188.606201171875, 162.51182556152344, -206.02175903320312, 157.8602294921875, 114.59342956542969, 1.6916618347167969, 67.56096649169922, 67.90568542480469, 8.815330505371094, 2.1367130279541016, 190.6217041015625, 192.69566345214844, 27.088607788085938, -42.92286682128906, 191.4780731201172, -15.800423622131348, 189.60252380371094, 4.9484100341796875, -6.008541107177734, -22.470245361328125, -16.362380981445312, -11.89935302734375, 4.009479522705078, 201.81387329101562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000498.npy"}
{"epoch": 0.7528344671201814, "step": 499, "batch_size": 64, "mean": 43.43333053588867, "std": 101.6302719116211, "min": -175.74774169921875, "p10": -88.61650314331052, "median": 30.443466186523438, "p90": 184.8170135498047, "max": 203.62901306152344, "pos_frac": 0.671875, "sample": [174.89385986328125, 152.65496826171875, -35.86888885498047, 73.97398376464844, 80.76583099365234, 91.94338989257812, -57.966209411621094, -175.7152099609375, -29.624393463134766, 127.86236572265625, 30.593311309814453, 182.0103302001953, 38.159942626953125, -6.265096664428711, 30.445449829101562, 8.103851318359375, 147.16845703125, -107.79369354248047, 187.23626708984375, 176.271240234375, 64.84408569335938, -15.351638793945312, -128.18487548828125, -133.85885620117188, -95.83898162841797, 49.513153076171875, 14.069534301757812, 150.09707641601562, -71.76405334472656, -44.275360107421875, 3.083820343017578, -7.178764343261719, 161.99331665039062, 25.0421142578125, 32.56498336791992, -1.6528396606445312, 80.65269470214844, 25.00617218017578, 2.6742477416992188, 30.441482543945312, 196.1130828857422, -25.111766815185547, 191.2371826171875, 50.30503845214844, -50.22560119628906, -17.414474487304688, 22.07733917236328, 124.62710571289062, 200.91778564453125, 203.62901306152344, 159.41717529296875, 109.21792602539062, 21.903308868408203, -63.708099365234375, -175.74774169921875, 22.57733154296875, 184.00979614257812, 140.93751525878906, -144.2090606689453, 192.38705444335938, -45.547828674316406, 185.1629638671875, 41.96189880371094, 24.488922119140625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000499.npy"}
{"epoch": 0.7543461829176115, "step": 500, "batch_size": 64, "mean": 49.34884262084961, "std": 82.37059020996094, "min": -162.46975708007812, "p10": -38.605413055419916, "median": 25.727523803710938, "p90": 181.44808349609377, "max": 191.8201904296875, "pos_frac": 0.734375, "sample": [0.2612762451171875, 3.9697113037109375, 19.863574981689453, -1.9733123779296875, 24.61376953125, 24.664535522460938, 64.72584533691406, 24.31427001953125, 7.1141510009765625, 186.41632080078125, 52.49095153808594, 184.6188201904297, 177.01248168945312, -7.016807556152344, 182.70834350585938, 44.42985534667969, 191.8201904296875, -44.445213317871094, 33.709625244140625, 5.3663177490234375, 178.50747680664062, -82.88677215576172, 3.6017227172851562, 23.447845458984375, -14.095081329345703, 150.78817749023438, 172.04283142089844, 8.36187744140625, 24.38042449951172, -9.139442443847656, -84.59358215332031, 34.88617706298828, 3.1455764770507812, 34.658966064453125, -162.46975708007812, 27.620220184326172, 26.790512084960938, 118.4276123046875, 142.35858154296875, -2.754974365234375, 72.0357894897461, -32.91358184814453, 103.71946716308594, 42.48027420043945, -1.223419189453125, 95.14208221435547, 30.940277099609375, -79.993408203125, 186.796142578125, -53.180809020996094, 188.5278778076172, -41.044769287109375, 0.6258049011230469, 34.73368835449219, 94.3267822265625, -1.3471527099609375, 0.2484607696533203, 171.6296844482422, 158.96417236328125, -2.211700439453125, -13.5948486328125, 129.44569396972656, 185.59649658203125, 120.8797607421875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000500.npy"}
{"epoch": 0.7558578987150416, "step": 501, "batch_size": 64, "mean": 41.89936065673828, "std": 67.49636840820312, "min": -66.16767120361328, "p10": -13.44840269088745, "median": 13.786312103271484, "p90": 156.8212417602539, "max": 187.47386169433594, "pos_frac": 0.671875, "sample": [-6.671962738037109, 172.33334350585938, -11.420623779296875, 30.527130126953125, 178.32791137695312, 51.19373321533203, -4.779632568359375, 13.501350402832031, 17.310470581054688, 185.53628540039062, 130.03936767578125, -66.16767120361328, 96.64645385742188, 75.19025421142578, -45.2647819519043, -0.5865745544433594, 14.071273803710938, 156.894287109375, -25.69451904296875, 1.727071762084961, 9.645217895507812, 1.3096351623535156, -1.6503667831420898, 69.7309799194336, 162.3872833251953, 156.6508026123047, 93.69447326660156, 187.47386169433594, -5.670753479003906, 2.3913650512695312, -63.70848846435547, -3.025115966796875, -1.7138175964355469, 5.060028076171875, 122.86495971679688, 1.4887847900390625, 21.18132781982422, 10.61518669128418, 1.9145278930664062, -11.679450988769531, 67.58758544921875, 178.55548095703125, -0.03781890869140625, 86.39586639404297, -10.344345092773438, -48.505775451660156, 19.280654907226562, 14.61505126953125, -21.888961791992188, 134.63526916503906, 37.56056213378906, 79.28447723388672, -0.7224788665771484, 14.8516845703125, 8.434616088867188, -14.206524848937988, 1.837432861328125, 18.75454330444336, 153.472900390625, 82.21653747558594, 112.07101440429688, -0.25753021240234375, 56.128265380859375, -9.8330078125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000501.npy"}
{"epoch": 0.7573696145124716, "step": 502, "batch_size": 64, "mean": 38.604698181152344, "std": 70.01765441894531, "min": -127.00946044921875, "p10": -26.21578903198242, "median": 26.059114456176758, "p90": 149.90409240722659, "max": 199.05621337890625, "pos_frac": 0.734375, "sample": [-84.3150634765625, 7.709123611450195, 4.214336395263672, -10.822975158691406, 96.01885986328125, 150.82601928710938, 39.91188049316406, 16.641632080078125, -3.92279052734375, 120.54693603515625, 2.3122406005859375, 157.57257080078125, 1.9738845825195312, -127.00946044921875, 48.579437255859375, 199.05621337890625, 11.321533203125, -73.07728576660156, -1.3886947631835938, -67.97439575195312, 91.72049713134766, -73.20893859863281, 12.312923431396484, 69.0577621459961, 120.05877685546875, -0.06464385986328125, 152.36720275878906, -9.23196792602539, 67.76631164550781, 19.075302124023438, 5.960548400878906, 1.8290748596191406, 61.2239990234375, -1.7121658325195312, 33.04292678833008, 50.01264953613281, 187.30259704589844, 196.34739685058594, -23.878341674804688, -10.027435302734375, 6.3093414306640625, -27.217552185058594, 40.43998336791992, 6.5218658447265625, 61.097564697265625, 4.50067138671875, -0.40918731689453125, 90.03185272216797, 164.3713836669922, -96.55181884765625, 4.848377227783203, 40.222900390625, 147.7529296875, 75.94078063964844, 87.30035400390625, 62.106048583984375, 48.12040710449219, 102.52299499511719, -10.599929809570312, 72.00188446044922, 33.677024841308594, 9.471611022949219, 71.92166137695312, 38.190948486328125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000502.npy"}
{"epoch": 0.7588813303099018, "step": 503, "batch_size": 64, "mean": 60.310672760009766, "std": 83.18994903564453, "min": -158.6104736328125, "p10": -14.179367065429686, "median": 50.23724555969238, "p90": 178.99725952148438, "max": 187.6707763671875, "pos_frac": 0.78125, "sample": [42.67572021484375, -2.4197044372558594, 88.87799072265625, 119.87435913085938, 54.536033630371094, 21.8631591796875, 181.16136169433594, -150.6747589111328, -8.78668212890625, 163.34275817871094, 86.84410095214844, -12.072189331054688, 119.63557434082031, 184.20022583007812, -28.60445213317871, 60.011566162109375, 1.1734161376953125, 18.713714599609375, 173.83499145507812, -51.919921875, 2.7740631103515625, 82.07652282714844, 182.91937255859375, -36.910301208496094, 51.28530502319336, 74.59786987304688, 49.189186096191406, 187.6707763671875, 106.70577239990234, 170.46807861328125, -12.841659545898438, 15.899024963378906, 172.39002990722656, 181.16493225097656, 180.87451171875, -1.9801292419433594, 24.798168182373047, 1.9329376220703125, -14.752670288085938, 14.529464721679688, -10.212738037109375, 9.908470153808594, 8.580419540405273, 80.41383361816406, 45.84291076660156, 4.398811340332031, 67.69676208496094, 180.06060791015625, 114.2298583984375, 4.622400283813477, 176.30186462402344, 147.60739135742188, -109.03569793701172, 5.943098068237305, 166.67782592773438, 102.6245346069336, 124.11431121826172, -2.2660045623779297, 39.47572326660156, 176.51611328125, 48.96400833129883, 74.45245361328125, 66.51824188232422, -158.6104736328125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000503.npy"}
{"epoch": 0.7603930461073318, "step": 504, "batch_size": 64, "mean": 65.74114227294922, "std": 82.44983673095703, "min": -125.499755859375, "p10": -11.032443237304685, "median": 42.917484283447266, "p90": 179.6244918823242, "max": 200.4154815673828, "pos_frac": 0.859375, "sample": [-125.499755859375, 170.8358612060547, 0.869354248046875, 49.066871643066406, 9.869117736816406, 0.04462432861328125, 156.87928771972656, 28.535572052001953, 190.09422302246094, 4.874717712402344, 2.935455322265625, 2.9210968017578125, 3.583873748779297, 200.4154815673828, 10.180931091308594, 140.482177734375, 161.06851196289062, -18.262359619140625, 190.3814697265625, 172.5692901611328, 32.073978424072266, 51.9093017578125, 2.463848114013672, 188.00904846191406, 179.3582000732422, -40.668853759765625, 2.104419708251953, 145.49224853515625, 162.46014404296875, 190.01724243164062, 179.73861694335938, 60.51091003417969, 124.41497039794922, 26.379985809326172, 5.117485046386719, 18.16313934326172, -1.3120155334472656, -12.344764709472656, 150.6463623046875, -7.970359802246094, 157.58131408691406, 25.894973754882812, 2.3367271423339844, -63.38311767578125, 135.72714233398438, 91.45318603515625, 4.455738067626953, 25.123323440551758, 151.7335205078125, 64.71886444091797, 3.3785171508789062, -47.68752670288086, -105.99099731445312, 64.114501953125, 60.13485336303711, 11.786857604980469, 3.1718530654907227, 36.768096923828125, 147.34219360351562, 132.73358154296875, 102.24533081054688, 199.12527465820312, 134.96353149414062, 61.29578399658203], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000504.npy"}
{"epoch": 0.7619047619047619, "step": 505, "batch_size": 64, "mean": 34.941402435302734, "std": 99.43045043945312, "min": -208.34011840820312, "p10": -100.4461585998535, "median": 22.089561462402344, "p90": 177.96698303222658, "max": 204.271484375, "pos_frac": 0.671875, "sample": [120.06117248535156, 17.97515869140625, 26.203964233398438, -19.36266326904297, 0.8393707275390625, 167.07476806640625, -139.21604919433594, 195.70449829101562, 60.28739547729492, 74.60840606689453, 178.69207763671875, 30.391529083251953, 72.89889526367188, -22.32208251953125, 150.52197265625, 8.750335693359375, -13.110015869140625, -4.219505310058594, 45.01618576049805, 8.17547607421875, -34.02107238769531, 167.13247680664062, 88.58253479003906, 91.35250091552734, 95.65011596679688, 204.271484375, 71.90962219238281, -18.797332763671875, 2.7168922424316406, -0.6756610870361328, -16.31353759765625, -72.20809173583984, 178.36618041992188, -156.517333984375, 183.0596466064453, 123.38002014160156, 96.6317138671875, 154.25164794921875, -26.737472534179688, -87.8291015625, 17.934799194335938, -133.59677124023438, 82.3897476196289, -135.74989318847656, 27.98388671875, 12.328601837158203, 59.62342071533203, -1.613616943359375, 187.76295471191406, -164.4881591796875, -49.08002471923828, 128.1683349609375, 186.08311462402344, 33.618080139160156, 147.9439697265625, -105.81132507324219, 49.39912414550781, 3.485504150390625, 3.0188255310058594, 177.0355224609375, 2.273956298828125, 0.6311435699462891, -87.92743682861328, -208.34011840820312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000505.npy"}
{"epoch": 0.763416477702192, "step": 506, "batch_size": 64, "mean": 43.43827819824219, "std": 104.36103820800781, "min": -199.85052490234375, "p10": -105.67426223754883, "median": 26.113271713256836, "p90": 177.82196502685548, "max": 211.14846801757812, "pos_frac": 0.734375, "sample": [66.90818786621094, 184.04818725585938, -2.212360382080078, 0.3420867919921875, 104.76593017578125, 79.28236389160156, 57.06439208984375, -136.85906982421875, 164.96804809570312, -10.690704345703125, 3.113117218017578, 65.89884948730469, 69.24008178710938, -15.257331848144531, 116.97994232177734, 26.57049560546875, -189.27993774414062, 211.14846801757812, 209.245849609375, 4.642204284667969, 111.12013244628906, -153.91012573242188, -19.91175079345703, -199.85052490234375, -158.91346740722656, -162.28256225585938, 4.40521240234375, 173.94886779785156, 194.3863525390625, 101.58815002441406, 175.22015380859375, 2.614013671875, 192.89012145996094, 4.4291839599609375, 182.64356994628906, 4.670745849609375, 169.85626220703125, 23.39600372314453, 108.70805358886719, -14.827850341796875, -60.249847412109375, 0.31441497802734375, 11.193550109863281, 15.212120056152344, 131.77081298828125, 25.656047821044922, -71.07542419433594, 8.777679443359375, 44.1805419921875, 14.927528381347656, 164.58982849121094, 29.472320556640625, -3.17181396484375, 95.60971069335938, 150.79942321777344, 47.37016296386719, -6.987998962402344, 178.93702697753906, -106.97675323486328, -102.63511657714844, 5.317802429199219, 166.2291259765625, 130.82411193847656, 159.86532592773438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000506.npy"}
{"epoch": 0.764928193499622, "step": 507, "batch_size": 64, "mean": 47.5564079284668, "std": 89.0269546508789, "min": -159.1668243408203, "p10": -67.8832046508789, "median": 26.957420349121094, "p90": 181.85188446044924, "max": 227.52557373046875, "pos_frac": 0.75, "sample": [161.70516967773438, 66.79914855957031, 8.12103271484375, -16.538909912109375, 211.59999084472656, 21.306835174560547, 33.39958190917969, -94.57272338867188, 0.7474136352539062, 91.47119140625, -71.8272476196289, -1.8337936401367188, -59.37351989746094, 50.00556182861328, -38.19151306152344, 26.612091064453125, 10.792938232421875, -5.61749267578125, 173.46090698242188, 6.434967041015625, 73.25291442871094, 12.55499267578125, -159.1668243408203, -50.803165435791016, 192.09249877929688, 0.3235931396484375, 51.66558837890625, 97.7493896484375, 45.81562805175781, 9.607765197753906, 15.210197448730469, 91.26144409179688, 215.32815551757812, -71.53021240234375, 165.29067993164062, 120.17486572265625, 29.69538116455078, 36.737571716308594, -107.9977798461914, 5.904336929321289, 81.6159896850586, 15.270431518554688, 49.326454162597656, 175.77064514160156, 141.8661651611328, -2.287872314453125, 27.302749633789062, 4.6455078125, 227.52557373046875, 198.40191650390625, -71.71031188964844, 177.4680938720703, 81.76409912109375, -1.2567596435546875, 22.028701782226562, 11.897911071777344, 11.17789077758789, -47.030738830566406, 89.94146728515625, -103.31446838378906, 115.98400115966797, 184.4480743408203, 121.37527465820312, 183.73065185546875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000507.npy"}
{"epoch": 0.7664399092970522, "step": 508, "batch_size": 64, "mean": 49.9659538269043, "std": 93.2735366821289, "min": -139.55691528320312, "p10": -44.98701438903808, "median": 31.598224639892578, "p90": 191.36370391845705, "max": 216.63424682617188, "pos_frac": 0.703125, "sample": [121.76730346679688, -20.21826934814453, 216.63424682617188, 121.75535583496094, 11.052154541015625, -119.03777313232422, 206.93267822265625, 0.5389022827148438, 99.40797424316406, 187.03048706054688, 213.2183380126953, 175.37493896484375, -9.711402893066406, 42.650325775146484, -132.87139892578125, 32.85192108154297, -139.55691528320312, 80.08946228027344, 111.42935180664062, -37.47412872314453, 30.344528198242188, -39.537445068359375, 91.9474105834961, -10.828239440917969, -5.937049865722656, -107.30647277832031, 18.377351760864258, -5.121879577636719, 206.13729858398438, 99.11500549316406, 10.160903930664062, 20.045806884765625, 74.9266357421875, 85.24641418457031, 170.46200561523438, 20.759857177734375, -47.32254409790039, 39.67524719238281, 189.40455627441406, 7.5009002685546875, -11.582611083984375, -2.2503585815429688, 192.20333862304688, 175.23739624023438, 10.475383758544922, -112.53250122070312, 64.02542114257812, 33.767333984375, 200.52549743652344, 5.3644256591796875, 153.3765106201172, 0.6784400939941406, 34.76002502441406, -1.2782173156738281, -4.852134704589844, -113.24385833740234, -5.35931396484375, 176.0821533203125, 76.913818359375, 50.0242919921875, 14.965404510498047, 21.900238037109375, 33.11711883544922, 195.58935546875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000508.npy"}
{"epoch": 0.7679516250944822, "step": 509, "batch_size": 64, "mean": 58.69524383544922, "std": 83.02764129638672, "min": -170.68435668945312, "p10": -28.761257171630856, "median": 55.93313789367676, "p90": 159.2442657470703, "max": 197.11117553710938, "pos_frac": 0.75, "sample": [54.79634094238281, -6.982551574707031, -0.7279167175292969, 164.60946655273438, 113.0843505859375, 0.3098182678222656, 98.79410552978516, -30.71894073486328, -70.05303955078125, 76.96463012695312, 57.0699348449707, 74.58708190917969, -4.660400390625, 61.83873748779297, 170.10498046875, -170.68435668945312, 71.19796752929688, 152.81141662597656, 140.63523864746094, -12.003189086914062, 138.18771362304688, -139.27659606933594, 45.05963134765625, 0.2508049011230469, 155.3031463623047, 3.757720947265625, 18.0987548828125, -43.92588806152344, 147.39254760742188, 2.9362640380859375, 197.11117553710938, 133.0127410888672, 21.662216186523438, 57.33160400390625, 63.12867736816406, -0.44632720947265625, 146.60232543945312, 83.39846801757812, 175.64918518066406, 142.65328979492188, 25.12000274658203, -8.378887176513672, 174.20457458496094, -123.1570053100586, 146.09616088867188, 127.84656524658203, 9.078262329101562, 46.524261474609375, 192.89669799804688, 159.43032836914062, 134.41603088378906, 106.18727111816406, -1.6614456176757812, -24.193328857421875, 27.658981323242188, 149.73204040527344, 7.742042541503906, -37.0971565246582, -2.7984161376953125, 29.22296905517578, 3.965333938598633, 48.7969970703125, 158.81011962890625, 117.19198608398438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000509.npy"}
{"epoch": 0.7694633408919124, "step": 510, "batch_size": 64, "mean": 62.599971771240234, "std": 99.58258819580078, "min": -186.55007934570312, "p10": -44.26875228881835, "median": 47.34967803955078, "p90": 185.8598587036133, "max": 205.35888671875, "pos_frac": 0.734375, "sample": [6.7222442626953125, 140.82647705078125, 169.73558044433594, 91.04656219482422, 2.8570098876953125, 48.234710693359375, 46.46464538574219, 190.4112091064453, 11.647132873535156, 5.332183837890625, -4.070167541503906, -1.581817626953125, 17.717662811279297, 23.489757537841797, -81.27278900146484, 8.87746810913086, 1.4541091918945312, 154.51724243164062, -65.14642333984375, 94.8959732055664, 110.22412109375, 171.7457275390625, 183.5249481201172, -0.5388107299804688, 6.20391845703125, 157.3485107421875, 20.81456756591797, -2.1861114501953125, 7.350883483886719, 74.567138671875, 132.7762451171875, 151.40884399414062, 200.9825439453125, -49.77880096435547, -0.8115234375, 128.7834930419922, 110.03154754638672, 201.02438354492188, 1.5237655639648438, -31.411972045898438, 10.615772247314453, -171.26792907714844, 163.3717498779297, 161.80075073242188, 205.35888671875, 97.7430648803711, 126.56407165527344, 196.7322540283203, 162.69668579101562, -9.910202026367188, 33.71605682373047, 169.7421112060547, 183.51754760742188, 138.6248779296875, -186.55007934570312, -0.23853302001953125, 55.863739013671875, 186.86053466796875, -25.749122619628906, 174.41207885742188, 197.87286376953125, -12.687921524047852, -153.61888122558594, -134.81455993652344], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000510.npy"}
{"epoch": 0.7709750566893424, "step": 511, "batch_size": 64, "mean": 66.58783721923828, "std": 97.78633117675781, "min": -152.13516235351562, "p10": -49.97957763671874, "median": 41.7305965423584, "p90": 192.58462219238282, "max": 216.55743408203125, "pos_frac": 0.75, "sample": [216.55743408203125, 171.04202270507812, 44.72441864013672, 183.83493041992188, 15.521717071533203, -4.1314544677734375, 36.93751525878906, -36.67510223388672, 98.59689331054688, 74.83557891845703, 85.35396575927734, -31.380508422851562, 11.094314575195312, 4.553535461425781, 177.15023803710938, 197.15814208984375, 162.17596435546875, 9.94495964050293, 89.26372528076172, 163.2782745361328, 138.84091186523438, -152.13516235351562, 32.60942077636719, -102.80816650390625, -61.37065124511719, -2.231597900390625, 192.5836181640625, 152.55422973632812, 105.04109191894531, -23.593894958496094, 35.85734558105469, 90.37010192871094, -127.21173095703125, 23.452285766601562, 173.49571228027344, 61.76188659667969, 175.98719787597656, -16.68468475341797, -0.06766510009765625, 11.759666442871094, 15.425018310546875, 38.73677444458008, 192.58505249023438, 193.88748168945312, 203.0266876220703, 172.9980010986328, -55.681495666503906, 9.778945922851562, 24.606674194335938, 186.57772827148438, 9.841995239257812, -1.07318115234375, 105.8547592163086, 126.96904754638672, 1.0567140579223633, 167.26829528808594, -82.32339477539062, 196.34042358398438, 143.88909912109375, 188.69100952148438, 195.5144500732422, 3.7134971618652344, -151.94618225097656, -2.1620101928710938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000511.npy"}
{"epoch": 0.7724867724867724, "step": 512, "batch_size": 64, "mean": 40.587364196777344, "std": 88.07730865478516, "min": -179.84014892578125, "p10": -61.10141716003417, "median": 14.466949462890625, "p90": 183.90200500488282, "max": 208.63821411132812, "pos_frac": 0.703125, "sample": [25.52344512939453, 16.18658447265625, 9.82730484008789, -129.3345947265625, -82.2847900390625, 188.61831665039062, -48.83810043334961, 7.239902496337891, 19.388519287109375, 32.14588165283203, 95.80360412597656, 134.2362060546875, -179.84014892578125, -66.35712432861328, 5.2284088134765625, 90.57579040527344, -3.6701889038085938, 4.652549743652344, 128.36984252929688, 2.7562789916992188, 183.77175903320312, 135.76791381835938, 6.7402801513671875, 154.4336700439453, 12.747314453125, 181.82177734375, -109.1351547241211, 152.50473022460938, -6.7357177734375, 192.4167938232422, -2.334766387939453, 36.31267547607422, 202.99530029296875, 0.04327583312988281, 93.64654541015625, 40.15013122558594, 2.883586883544922, 106.176513671875, 22.216541290283203, 183.95782470703125, -7.5873260498046875, -66.79349517822266, -1.3732147216796875, 5.2046051025390625, 56.277381896972656, 208.63821411132812, -15.814231872558594, 4.659812927246094, 3.7998046875, -72.2281723022461, 97.5387954711914, -1.5093574523925781, 22.9019775390625, -0.7686767578125, 21.352035522460938, 151.84747314453125, 4.907615661621094, 33.35768127441406, 193.87432861328125, 185.9627227783203, -36.06451416015625, -42.107574462890625, 40.508888244628906, -29.602386474609375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000512.npy"}
{"epoch": 0.7739984882842026, "step": 513, "batch_size": 64, "mean": 52.076133728027344, "std": 79.35090637207031, "min": -129.79168701171875, "p10": -13.644640922546387, "median": 21.41493034362793, "p90": 185.66649017333984, "max": 207.43450927734375, "pos_frac": 0.765625, "sample": [203.8406524658203, 177.46597290039062, 189.54898071289062, 3.5138626098632812, 94.81355285644531, 43.8280143737793, 107.48062133789062, -0.21537017822265625, 0.6966476440429688, 30.74591827392578, 12.309951782226562, 1.5436668395996094, 174.42587280273438, 20.300373077392578, 11.280479431152344, 164.9832763671875, 40.2529182434082, -12.17803955078125, 194.4271240234375, 20.4874267578125, 9.994632720947266, 185.98268127441406, 21.001792907714844, 76.40184020996094, 96.87832641601562, -70.1290283203125, -8.315673828125, 92.78154754638672, -21.024158477783203, 13.054275512695312, 24.975584030151367, 94.86913299560547, 65.4830322265625, -13.749271392822266, 55.5302734375, 6.001075744628906, 21.828067779541016, 79.906982421875, 207.43450927734375, 30.413246154785156, -13.400503158569336, -12.414093017578125, 189.12960815429688, -24.123268127441406, 178.1407470703125, 184.9287109375, 191.15528869628906, -129.79168701171875, 98.208740234375, 63.01854705810547, 89.10918426513672, -41.083465576171875, 12.899765014648438, 129.50064086914062, -2.480316162109375, -118.9410400390625, 10.360572814941406, 10.86993408203125, 8.524078369140625, 48.43254089355469, 8.507492065429688, -5.329490661621094, -2.729644775390625, 11.50921630859375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000513.npy"}
{"epoch": 0.7755102040816326, "step": 514, "batch_size": 64, "mean": 64.87118530273438, "std": 87.221923828125, "min": -178.94412231445312, "p10": -12.87427997589111, "median": 67.32944869995117, "p90": 173.55655517578126, "max": 215.14471435546875, "pos_frac": 0.765625, "sample": [13.176422119140625, 91.14083862304688, 52.13901901245117, 144.74691772460938, 36.2428092956543, -30.12248992919922, 130.94296264648438, -4.143318176269531, 19.994903564453125, -4.281078338623047, 85.84222412109375, 112.9817886352539, 136.26702880859375, 82.30319213867188, 1.8086471557617188, -6.773918151855469, -9.1788330078125, -9.350454330444336, 4.078861236572266, 6.086265563964844, -178.94412231445312, -8.243423461914062, 149.04803466796875, 115.95561218261719, 198.42210388183594, 2.4610748291015625, 15.093292236328125, 159.71141052246094, 122.81336975097656, 168.30189514160156, 44.342567443847656, -93.39167785644531, 172.84580993652344, 148.06015014648438, 58.83992004394531, 94.59110260009766, 28.87030029296875, 127.3677978515625, 103.471923828125, 6.3278656005859375, 53.870567321777344, -125.20429992675781, -55.452396392822266, 99.33668518066406, 85.39804077148438, 151.23428344726562, 75.81897735595703, -1.110137939453125, 173.8611602783203, 123.74430847167969, 8.132781982421875, 90.70950317382812, 26.565235137939453, 180.99478149414062, -139.29087829589844, 179.13546752929688, 183.53884887695312, 32.461517333984375, 170.31710815429688, -7.784740447998047, 215.14471435546875, 172.40618896484375, -14.384490966796875, 182.466064453125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000514.npy"}
{"epoch": 0.7770219198790628, "step": 515, "batch_size": 64, "mean": 51.578346252441406, "std": 90.62377166748047, "min": -127.67256927490234, "p10": -69.15394210815428, "median": 27.435338973999023, "p90": 177.38635711669923, "max": 204.2125701904297, "pos_frac": 0.71875, "sample": [16.947784423828125, 13.290462493896484, 2.0895614624023438, 93.73223876953125, 165.9445037841797, 155.8496551513672, -1.1285247802734375, 119.16415405273438, 90.54061889648438, 172.66232299804688, 20.471817016601562, 179.09788513183594, 67.72916412353516, -29.90575408935547, 88.33990478515625, -109.83865356445312, -101.37467193603516, 144.17984008789062, -57.174041748046875, -89.62236022949219, 4.6279296875, 2.5312957763671875, -32.60502624511719, 104.62281799316406, -34.588294982910156, 14.603689193725586, 187.7935333251953, 73.82007598876953, 66.78230285644531, -1.9329910278320312, -17.15680694580078, 86.74729919433594, 33.98219680786133, 21.884090423583984, 94.66098022460938, 184.2252960205078, -85.8981704711914, 163.87722778320312, -2.7502822875976562, 13.967071533203125, -127.67256927490234, 65.4275894165039, 186.39682006835938, -74.2881851196289, 165.48464965820312, 204.2125701904297, 0.6584053039550781, 101.21052551269531, 15.505401611328125, 173.39279174804688, 8.256805419921875, 145.77720642089844, 32.98658752441406, 193.25042724609375, 199.88922119140625, -55.3374137878418, 148.10870361328125, 10.406539916992188, 166.37640380859375, -34.991050720214844, 52.56612777709961, 10.218719482421875, -7.910810470581055, -99.1015625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000515.npy"}
{"epoch": 0.7785336356764928, "step": 516, "batch_size": 64, "mean": 46.28776550292969, "std": 84.87876892089844, "min": -182.4265899658203, "p10": -44.646211242675776, "median": 45.77364730834961, "p90": 165.02914733886723, "max": 195.90789794921875, "pos_frac": 0.796875, "sample": [173.443359375, 191.29669189453125, 98.01907348632812, -48.28398132324219, -100.11849975585938, 38.65037536621094, -10.963348388671875, 24.74573516845703, 140.63111877441406, 82.28997039794922, 22.273300170898438, 62.87202835083008, 34.786285400390625, 195.90789794921875, 1.3718223571777344, 4.14361572265625, 57.66355895996094, 7.491886138916016, 85.10862731933594, 54.03106689453125, 24.48583221435547, -182.4265899658203, 115.39949798583984, 46.94477081298828, 179.4237518310547, 8.973026275634766, 57.40500259399414, 5.077014923095703, 155.89968872070312, 30.867759704589844, -175.242431640625, -70.91921997070312, 0.6853561401367188, -127.64041900634766, -1.3844528198242188, 109.49522399902344, 173.09188842773438, 57.17839813232422, 6.022796630859375, 105.04049682617188, 149.99884033203125, -160.83103942871094, 24.872283935546875, 168.9417724609375, 144.8563995361328, -36.1580810546875, 72.35726165771484, 70.00822448730469, 58.5380859375, 38.418243408203125, 51.583229064941406, 145.984619140625, 33.177406311035156, 44.60252380371094, 3.8577880859375, 111.41651916503906, 62.58470916748047, 103.32713317871094, -1.2199020385742188, -1.4733848571777344, 78.75970458984375, 0.5040321350097656, -6.943077087402344, 171.5155029296875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000516.npy"}
{"epoch": 0.780045351473923, "step": 517, "batch_size": 64, "mean": 31.928131103515625, "std": 91.82672119140625, "min": -180.13795471191406, "p10": -57.904582214355464, "median": 6.924577713012695, "p90": 181.4659622192383, "max": 211.17588806152344, "pos_frac": 0.640625, "sample": [139.94888305664062, -115.48810577392578, 17.86834716796875, 27.377418518066406, -152.31040954589844, 28.924896240234375, 68.0959243774414, -25.496639251708984, 164.35638427734375, 60.14104461669922, 66.61578369140625, -1.4975624084472656, -2.4821701049804688, 1.007781982421875, -14.975296020507812, -7.892250061035156, 4.629631042480469, -1.3328475952148438, 14.369709014892578, 0.21895980834960938, 170.17074584960938, -50.171630859375, 28.49420166015625, 195.60162353515625, -47.825103759765625, 15.089225769042969, 6.055629730224609, 116.80247497558594, -9.925994873046875, 131.9627685546875, 33.40601348876953, 211.17588806152344, 19.53882598876953, 6.054729461669922, -65.83562469482422, -180.13795471191406, 5.679298400878906, 195.1230926513672, -55.77288055419922, -58.81816864013672, 12.871681213378906, 156.93666076660156, 196.15615844726562, 17.030223846435547, 5.224784851074219, 183.6213836669922, 7.793525695800781, 3.2717132568359375, -31.11821174621582, -0.28912353515625, -11.496309280395508, 1.9830455780029297, -0.9299468994140625, 189.96998596191406, 197.73004150390625, 176.4366455078125, 15.642616271972656, -0.5921669006347656, 23.066120147705078, -0.15000152587890625, 170.88710021972656, -156.53961181640625, 55.63713836669922, -108.48965454101562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000517.npy"}
{"epoch": 0.781557067271353, "step": 518, "batch_size": 64, "mean": 52.375755310058594, "std": 91.17594146728516, "min": -178.08099365234375, "p10": -59.22515373229978, "median": 43.19001388549805, "p90": 182.27405090332036, "max": 216.2803955078125, "pos_frac": 0.734375, "sample": [83.89505004882812, -14.721481323242188, -5.155723571777344, 112.41063690185547, 41.05982208251953, 1.7329483032226562, 28.6131591796875, 118.23111724853516, -2.8351593017578125, 141.10940551757812, 75.38493347167969, -119.7231674194336, 29.30884552001953, 77.00540161132812, 172.67974853515625, -13.687149047851562, 117.11260223388672, 76.82057189941406, 1.038106918334961, 61.503787994384766, 216.2803955078125, -178.08099365234375, -0.7462997436523438, 152.58966064453125, 189.37899780273438, 99.82268524169922, -106.11310577392578, 55.24324035644531, -171.12289428710938, 71.12342834472656, 6.735034942626953, -33.869991302490234, -0.7899112701416016, 86.57975006103516, 133.40199279785156, 1.557830810546875, 2.2236175537109375, 11.195770263671875, 156.54766845703125, 198.9188232421875, -0.3236083984375, 45.32020568847656, 0.8195075988769531, 196.61773681640625, 113.60443115234375, 207.35238647460938, -70.0916519165039, 172.03631591796875, 52.19854736328125, -74.59429931640625, -14.894363403320312, 50.11345291137695, 168.09744262695312, 0.4910469055175781, -0.4337959289550781, 186.38589477539062, 157.38905334472656, 37.83742904663086, 90.85702514648438, -74.2452621459961, 1.9633293151855469, 0.24843597412109375, 193.57882690429688, 39.06127166748047], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000518.npy"}
{"epoch": 0.783068783068783, "step": 519, "batch_size": 64, "mean": 31.934837341308594, "std": 109.01180267333984, "min": -203.7298583984375, "p10": -124.27926177978516, "median": 14.79214859008789, "p90": 176.7946472167969, "max": 192.5522003173828, "pos_frac": 0.640625, "sample": [-147.41732788085938, 70.47254943847656, -10.4500732421875, -195.45448303222656, -1.8879318237304688, 146.63002014160156, 15.57107925415039, 190.88656616210938, -98.79718017578125, 94.29421997070312, 1.992462158203125, 73.14906311035156, 142.87583923339844, 5.6501007080078125, 15.124801635742188, 48.43521499633789, -4.659626007080078, 14.459495544433594, -124.54296112060547, -14.562759399414062, 166.55288696289062, 120.43938446044922, -137.2213134765625, 163.03233337402344, 180.216796875, 186.89361572265625, -10.260250091552734, 74.47792053222656, 159.60093688964844, 102.80987548828125, -80.29554748535156, 157.79562377929688, 6.081382751464844, 97.06251525878906, 10.257076263427734, -19.349609375, 1.4968948364257812, 4.089469909667969, 168.80963134765625, -162.4114532470703, -41.98340606689453, 2.853832244873047, 163.66131591796875, 42.818416595458984, 113.87226104736328, -203.7298583984375, 75.06583404541016, -108.43582153320312, 86.15082550048828, -185.22349548339844, 184.70806884765625, -1.8457756042480469, -108.78450012207031, 192.04837036132812, 26.372024536132812, 5.736869812011719, 162.53762817382812, 192.5522003173828, -17.416244506835938, -123.6639633178711, 54.96441650390625, -53.907020568847656, 184.45401000976562, -10.82358169555664], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000519.npy"}
{"epoch": 0.7845804988662132, "step": 520, "batch_size": 64, "mean": 47.34003829956055, "std": 74.83524322509766, "min": -155.78546142578125, "p10": -23.380928802490228, "median": 28.715938568115234, "p90": 162.00275573730474, "max": 195.3688507080078, "pos_frac": 0.796875, "sample": [150.8857421875, -11.878232955932617, -0.9701404571533203, 23.18115234375, 91.08476257324219, 82.30203247070312, 127.91060638427734, 3.2255706787109375, 8.632225036621094, 51.91740417480469, 147.28939819335938, 23.307601928710938, 12.103195190429688, 55.807655334472656, 24.885940551757812, 26.178482055664062, 7.0812530517578125, -89.76325225830078, 3.4670677185058594, 166.7398223876953, 95.92242431640625, 150.94960021972656, 46.75468444824219, 177.15768432617188, 85.32296752929688, 26.29674530029297, 168.3860321044922, 41.07179260253906, -1.8243408203125, 86.33222961425781, 43.678070068359375, 5.149478912353516, 34.366905212402344, -16.926559448242188, 12.999595642089844, 0.15764808654785156, 184.3666534423828, 131.39166259765625, 70.58317565917969, -77.24964904785156, -102.49089813232422, -35.90425109863281, 48.523681640625, -155.78546142578125, 20.67552947998047, 18.129459381103516, 86.89239501953125, 7.162849426269531, 130.48861694335938, 31.1351318359375, 37.379364013671875, 186.44615173339844, 98.46774291992188, 12.958932876586914, 195.3688507080078, 46.412841796875, 11.25384521484375, 1.0021438598632812, -26.14708709716797, -1.2261505126953125, -4.505519866943359, 137.37754821777344, 174.89833068847656, -57.02661895751953], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000520.npy"}
{"epoch": 0.7860922146636432, "step": 521, "batch_size": 64, "mean": 53.414710998535156, "std": 86.0402603149414, "min": -185.24301147460938, "p10": -27.2446907043457, "median": 29.864362716674805, "p90": 171.34705505371093, "max": 221.290283203125, "pos_frac": 0.71875, "sample": [99.05517578125, 65.70306396484375, 171.37158203125, 9.812164306640625, 151.78073120117188, 1.9127197265625, 120.68583679199219, -3.07269287109375, -82.66021728515625, 157.220947265625, 45.805870056152344, -42.14067840576172, 75.977294921875, 64.20108032226562, 140.0276336669922, 53.68975067138672, 27.489166259765625, 6.441825866699219, -13.789337158203125, 185.41354370117188, -1.07354736328125, 161.10891723632812, -39.533935546875, 5.479644775390625, 116.7109146118164, 171.28982543945312, -13.112625122070312, 174.07168579101562, 40.21238708496094, 26.69330596923828, 204.29595947265625, 221.290283203125, -175.9828338623047, 32.239559173583984, 17.666580200195312, -28.915489196777344, 9.507614135742188, -7.772552490234375, 3.2186622619628906, 196.75770568847656, 156.980712890625, 6.437889099121094, -23.346160888671875, 17.625167846679688, -21.1667537689209, -0.3549785614013672, 131.140869140625, 152.9511260986328, -2.0016098022460938, 165.4368896484375, -38.59794998168945, 53.62091827392578, -185.24301147460938, 185.18934631347656, 10.064178466796875, 78.26663208007812, 68.39292907714844, -18.46209716796875, 87.33132934570312, -1.8265914916992188, 99.92440795898438, 93.37106323242188, 26.714492797851562, 27.015167236328125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000521.npy"}
{"epoch": 0.7876039304610734, "step": 522, "batch_size": 64, "mean": 37.793296813964844, "std": 92.25397491455078, "min": -190.698974609375, "p10": -39.79955673217773, "median": 12.086687088012695, "p90": 183.81448974609376, "max": 197.5213165283203, "pos_frac": 0.671875, "sample": [12.145233154296875, 46.474056243896484, 1.1595191955566406, 149.6936492919922, 38.310569763183594, 183.39031982421875, 88.04933166503906, 180.64224243164062, 184.02476501464844, 59.47746276855469, -148.61349487304688, -41.99976348876953, 47.627113342285156, 191.65606689453125, 2.8331451416015625, 40.35851287841797, -34.47455596923828, 6.529485702514648, 71.1773681640625, -32.93285369873047, 49.55604934692383, -0.01084136962890625, -34.665740966796875, -190.698974609375, 1.179595947265625, 1.967193603515625, -61.716556549072266, 23.948165893554688, -129.92538452148438, 2.3025436401367188, -0.40833282470703125, 171.2386474609375, -30.517913818359375, 183.99627685546875, 118.27902221679688, -118.12001037597656, -7.494091033935547, 17.156517028808594, -18.82917022705078, 19.10521697998047, 13.681327819824219, -6.0461883544921875, 3.1663665771484375, -7.947120666503906, 74.89107513427734, 0.2420063018798828, 162.38821411132812, -115.25222778320312, 184.882080078125, 12.746212005615234, 51.85302734375, -27.551467895507812, 144.19483947753906, 197.5213165283203, -1.755584716796875, 3.43072509765625, -32.714012145996094, 187.33193969726562, 168.3090057373047, 185.34912109375, -10.608482360839844, 12.028141021728516, 3.7742652893066406, 172.98587036132812], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000522.npy"}
{"epoch": 0.7891156462585034, "step": 523, "batch_size": 64, "mean": 43.630802154541016, "std": 84.65016174316406, "min": -182.69400024414062, "p10": -56.792382049560544, "median": 37.963035583496094, "p90": 166.60168762207033, "max": 190.54751586914062, "pos_frac": 0.734375, "sample": [-53.58070373535156, 185.46121215820312, 37.979026794433594, -23.155227661132812, 104.16975402832031, 127.10247802734375, 10.046653747558594, 27.524932861328125, 101.9557113647461, 84.26497650146484, 0.00423431396484375, 13.874130249023438, 190.54751586914062, 17.215316772460938, 64.9489974975586, 86.40357971191406, 47.9544677734375, -0.5808143615722656, 89.26233673095703, -106.69146728515625, 23.03003692626953, 154.617919921875, -61.448692321777344, 6.953437805175781, -114.16973876953125, 105.56474304199219, 8.654472351074219, -50.59303283691406, 1.5755691528320312, 5.142856597900391, 56.93193817138672, 163.24913024902344, 78.0316162109375, 77.84443664550781, 168.0384979248047, 188.70681762695312, 177.59432983398438, -7.838623046875, 42.541717529296875, -114.598388671875, 120.55496215820312, -112.96965026855469, 182.23074340820312, 35.6748046875, 75.5887451171875, -8.979240417480469, -39.052818298339844, 30.983016967773438, 97.97605895996094, -58.16881561279297, 159.9914093017578, 11.29684066772461, 143.10861206054688, -17.78205108642578, 81.13597869873047, 177.05364990234375, 37.947044372558594, 46.63482666015625, 4.618686676025391, -182.69400024414062, -45.77362060546875, 65.54191589355469, -11.54595947265625, 84.46392822265625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000523.npy"}
{"epoch": 0.7906273620559335, "step": 524, "batch_size": 64, "mean": 47.05378723144531, "std": 88.02437591552734, "min": -182.35946655273438, "p10": -52.4455108642578, "median": 35.05872917175293, "p90": 164.08522338867192, "max": 187.37759399414062, "pos_frac": 0.765625, "sample": [126.24168395996094, 14.216789245605469, 138.10736083984375, 28.43697738647461, -7.6183319091796875, 0.47365570068359375, -32.06043243408203, 72.73857116699219, 168.15440368652344, 35.39912796020508, 120.14327239990234, 0.18975448608398438, -99.9519271850586, 33.512474060058594, -40.20426940917969, 183.14584350585938, -170.43104553222656, -57.69175720214844, 38.589141845703125, 104.24600982666016, 145.9957275390625, -182.35946655273438, 25.788307189941406, 2.0256195068359375, 1.3017196655273438, 4.828310012817383, 176.96011352539062, 115.47858428955078, 125.68336486816406, -2.0168304443359375, -7.490268707275391, 89.51668548583984, -84.11752319335938, 18.311241149902344, 35.469390869140625, 48.643890380859375, 152.6812744140625, 34.71833038330078, 17.119110107421875, 0.014801025390625, 20.944244384765625, -4.3134613037109375, -1.0022201538085938, 176.00746154785156, 72.32467651367188, 187.37759399414062, 154.59046936035156, -2.762481689453125, 43.444664001464844, 141.61819458007812, 99.60399627685547, 7.904541015625, 136.5395050048828, 129.03224182128906, 50.21617126464844, 84.36315155029297, 182.93417358398438, 116.52719116210938, -163.70660400390625, 1.1287460327148438, 171.71827697753906, -96.33719635009766, 124.16514587402344, 4.934169769287109], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000524.npy"}
{"epoch": 0.7921390778533636, "step": 525, "batch_size": 64, "mean": 30.131698608398438, "std": 84.59285736083984, "min": -160.97589111328125, "p10": -72.90135040283201, "median": 11.728128433227539, "p90": 164.48472290039066, "max": 208.059814453125, "pos_frac": 0.671875, "sample": [35.723785400390625, -39.03901672363281, -2.7713470458984375, -80.09715270996094, -105.47145080566406, 150.2737579345703, 50.073585510253906, 71.45526123046875, 196.17857360839844, -6.87591552734375, 119.20925903320312, -5.5565948486328125, 35.49195098876953, 7.897174835205078, 44.253028869628906, 80.03985595703125, 167.7729034423828, -21.247695922851562, 208.059814453125, -12.009208679199219, 43.677337646484375, -6.2560577392578125, 32.458770751953125, 172.18753051757812, -101.93537139892578, 15.571813583374023, 46.83854675292969, -36.004764556884766, 2.131134033203125, 76.05888366699219, -126.39933776855469, 44.52894592285156, 6.555751800537109, 138.94146728515625, 87.08848571777344, 1.499725341796875, 11.22054672241211, 3.5826187133789062, 202.44415283203125, -144.6941680908203, -56.11114501953125, 12.235710144042969, -0.2240009307861328, -160.97589111328125, -10.806999206542969, 69.89463806152344, 2.7046470642089844, 1.1498947143554688, 45.657676696777344, 186.6573944091797, 14.0408935546875, 173.78964233398438, 74.33426666259766, -1.4470176696777344, 48.20957946777344, 110.15243530273438, -16.57916259765625, 51.70855712890625, 4.669563293457031, 7.5922088623046875, -7.922977447509766, 156.8123016357422, 9.153827667236328, -149.12387084960938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000525.npy"}
{"epoch": 0.7936507936507936, "step": 526, "batch_size": 64, "mean": 39.6178092956543, "std": 83.4588394165039, "min": -171.1062469482422, "p10": -44.56566772460937, "median": 30.26951313018799, "p90": 154.6100311279297, "max": 201.2532501220703, "pos_frac": 0.6875, "sample": [201.2532501220703, 22.781295776367188, 1.9754714965820312, 1.786529541015625, -10.23154067993164, 35.62423324584961, 31.4337100982666, 165.40350341796875, -134.1378936767578, -40.89393615722656, 144.2645263671875, 149.08938598632812, -23.022735595703125, 11.210433959960938, 155.74221801757812, -30.21009063720703, 30.531373977661133, 186.7510223388672, 34.0257568359375, 144.65403747558594, 40.38505554199219, 188.97866821289062, 43.94847869873047, 96.37989807128906, 102.74253845214844, 20.2381591796875, 40.31605529785156, 28.951927185058594, -83.77925109863281, 33.335205078125, -4.284313201904297, 0.59466552734375, -0.7070541381835938, -5.862361907958984, 85.67172241210938, -21.181198120117188, 22.780956268310547, -0.33099365234375, 81.73753356933594, -122.2227783203125, 89.31733703613281, 30.007652282714844, 41.07997131347656, -22.705915451049805, 106.58504486083984, -171.1062469482422, -7.052337646484375, -46.13926696777344, 151.96826171875, -11.817474365234375, 102.47428894042969, -0.473236083984375, 198.9493865966797, 11.374946594238281, 188.58106994628906, 76.6861572265625, -46.31635284423828, 70.32380676269531, 85.35273742675781, 112.9793701171875, 22.91179656982422, 3.9980545043945312, 82.90719604492188, -160.07003784179688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000526.npy"}
{"epoch": 0.7951625094482238, "step": 527, "batch_size": 64, "mean": 33.732025146484375, "std": 77.55470275878906, "min": -167.71524047851562, "p10": -23.583172607421872, "median": 15.322269439697266, "p90": 155.8092956542969, "max": 202.19937133789062, "pos_frac": 0.6875, "sample": [-1.8803558349609375, 3.302928924560547, 30.614028930664062, 115.00194549560547, 11.691856384277344, -2.5346145629882812, 173.4287872314453, 80.28627014160156, 158.989990234375, 202.19937133789062, 74.1304931640625, 14.765869140625, 31.557701110839844, -30.35845947265625, 64.25874328613281, 5.772869110107422, 117.34746551513672, 2.1543426513671875, 67.4285888671875, -18.83559799194336, 22.81804656982422, -130.5108642578125, -95.24976348876953, -10.805099487304688, 11.76043701171875, 50.309879302978516, 189.4468994140625, 1.1220703125, 22.576675415039062, 16.54969024658203, 24.766708374023438, 15.97979736328125, -78.90573120117188, 9.888839721679688, 104.27946472167969, 175.0224609375, 148.53414916992188, -10.9364013671875, -1.2445106506347656, 130.16700744628906, 140.34735107421875, 4.320034027099609, 131.05072021484375, -24.150177001953125, -22.260162353515625, -12.269752502441406, -17.010393142700195, 183.5297393798828, -5.939788818359375, 55.17424011230469, 19.126708984375, 1.7085762023925781, -16.31121826171875, -4.948337554931641, 6.429901123046875, 0.052074432373046875, 23.041534423828125, 25.32624053955078, 15.878669738769531, 100.47319030761719, -167.71524047851562, -124.74105834960938, 158.92721557617188, -6.082515716552734], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000527.npy"}
{"epoch": 0.7966742252456538, "step": 528, "batch_size": 64, "mean": 39.19126892089844, "std": 90.2035903930664, "min": -179.86404418945312, "p10": -35.12727165222168, "median": 22.24875831604004, "p90": 169.42491912841797, "max": 207.86328125, "pos_frac": 0.6875, "sample": [101.24154663085938, -158.382080078125, 126.971923828125, 19.494922637939453, 94.822998046875, 83.19457244873047, 19.069080352783203, -1.8560867309570312, 25.967700958251953, 11.07436752319336, -4.9323883056640625, -16.623947143554688, 176.24310302734375, -178.7474365234375, 5.215553283691406, -1.2388763427734375, 135.64791870117188, 36.45841598510742, 6.143962860107422, -0.35907745361328125, -14.630767822265625, 198.06492614746094, 1.3484745025634766, -26.066741943359375, 0.9395332336425781, 0.19240570068359375, -3.8012542724609375, -78.9773178100586, -34.47370529174805, 183.95721435546875, 55.95210266113281, 50.373985290527344, 25.080795288085938, 207.86328125, 44.511505126953125, -9.481330871582031, 195.75753784179688, 11.951065063476562, 129.52195739746094, 25.722614288330078, 168.8601531982422, 157.77862548828125, 67.58661651611328, 175.2616424560547, -147.73947143554688, 88.74925231933594, 0.8926353454589844, 7.75396728515625, 94.41542053222656, -4.578472137451172, 75.79528045654297, -179.86404418945312, -110.93758392333984, 138.321044921875, 25.002593994140625, 45.3301887512207, -35.407371520996094, -3.3631744384765625, 154.20274353027344, -13.64638900756836, 25.242416381835938, 169.66696166992188, 154.10011291503906, 11.605743408203125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000528.npy"}
{"epoch": 0.7981859410430839, "step": 529, "batch_size": 64, "mean": 78.05900573730469, "std": 91.92823791503906, "min": -156.1966094970703, "p10": -11.131934356689452, "median": 74.24507904052734, "p90": 190.07024841308595, "max": 207.57931518554688, "pos_frac": 0.78125, "sample": [180.24130249023438, -80.7116470336914, 79.51831817626953, -1.2570877075195312, 136.67831420898438, -12.772300720214844, 207.57931518554688, 179.92227172851562, 72.4487533569336, 11.441200256347656, -6.425422668457031, -3.0452537536621094, -2.405029296875, 115.93814086914062, 124.72666931152344, 5.963962554931641, -151.13653564453125, 191.7731475830078, -0.5887641906738281, 181.0919647216797, 191.360107421875, 4.906768798828125, 28.7916259765625, -3.1714019775390625, 193.84457397460938, 182.15707397460938, 17.479812622070312, 76.0414047241211, -11.721755981445312, -36.473045349121094, 174.75282287597656, 150.60562133789062, 19.69831085205078, 187.97360229492188, 157.99853515625, 166.3839111328125, 166.50704956054688, 160.12713623046875, 76.90701293945312, 136.95437622070312, 48.17124938964844, 170.600830078125, 195.70669555664062, 4.527732849121094, 129.856201171875, 163.27389526367188, 8.455841064453125, 5.210941314697266, 14.702400207519531, 142.53756713867188, 45.311676025390625, -22.89788818359375, 38.05580520629883, 180.8319549560547, 9.699817657470703, 190.96881103515625, 164.91702270507812, 171.41580200195312, 2.4953384399414062, 8.929916381835938, -156.1966094970703, 18.36376953125, -9.755683898925781, 200.488037109375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000529.npy"}
{"epoch": 0.799697656840514, "step": 530, "batch_size": 64, "mean": 57.46356964111328, "std": 81.23524475097656, "min": -108.89424896240234, "p10": -19.64773406982421, "median": 23.98493766784668, "p90": 189.04689025878906, "max": 229.27130126953125, "pos_frac": 0.6875, "sample": [229.27130126953125, 135.01766967773438, 30.27679443359375, -11.910125732421875, 162.3437042236328, 52.827857971191406, 198.07797241210938, 1.3888702392578125, 194.0347900390625, 112.213134765625, -1.0618515014648438, 193.17086791992188, -91.20777130126953, -1.353546142578125, 102.1610107421875, 27.34380340576172, 143.8431854248047, 2.5181350708007812, 189.31417846679688, 9.333389282226562, -8.569023132324219, -5.760581970214844, -41.0916748046875, 101.54637908935547, 88.80960845947266, 2.0533828735351562, 63.08030700683594, 132.17840576171875, 16.010963439941406, 191.44772338867188, 15.662063598632812, -7.7296295166015625, 160.04531860351562, 200.01824951171875, 11.317184448242188, 12.058887481689453, 86.27750396728516, 183.0362548828125, -108.89424896240234, 44.94606018066406, -7.072624206542969, 16.65869140625, -4.9680328369140625, 98.0494613647461, -34.754493713378906, -31.573989868164062, 154.02734375, -5.280067443847656, -5.1894683837890625, -30.6556396484375, -4.747943878173828, -22.963851928710938, 148.80947875976562, 22.311275482177734, 188.4232177734375, 59.386932373046875, 128.87411499023438, -5.5210418701171875, -6.450508117675781, 7.613090515136719, 86.73770141601562, 79.33123016357422, 6.918205261230469, 25.658599853515625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000530.npy"}
{"epoch": 0.8012093726379441, "step": 531, "batch_size": 64, "mean": 41.554649353027344, "std": 75.44834899902344, "min": -166.68698120117188, "p10": -29.66850395202636, "median": 15.514896392822266, "p90": 160.11764373779297, "max": 192.18304443359375, "pos_frac": 0.703125, "sample": [3.585712432861328, 13.30263900756836, 104.51917266845703, -9.141059875488281, 59.04803466796875, 48.882545471191406, 2.083953857421875, 89.80047607421875, 62.218353271484375, -4.096611022949219, 91.06373596191406, 3.377042770385742, 163.11813354492188, -5.6809539794921875, 192.18304443359375, 130.56802368164062, -98.31262969970703, -51.388999938964844, -6.636810302734375, 97.12935638427734, 33.90025329589844, 157.61557006835938, -10.5069580078125, 11.515762329101562, -13.891181945800781, 46.32109069824219, 131.59677124023438, -32.31941223144531, 1.0469818115234375, -51.703033447265625, 16.162033081054688, -0.5057773590087891, -23.483051300048828, 1.5110855102539062, -166.68698120117188, -4.073947906494141, 128.91531372070312, 16.380523681640625, 94.18756103515625, -86.05125427246094, 20.705291748046875, 26.264633178710938, 164.15260314941406, 160.2023162841797, -4.25347900390625, -41.140167236328125, 176.05462646484375, 165.5201416015625, 37.74310302734375, 14.867759704589844, 4.783649444580078, 159.92007446289062, 4.720039367675781, 126.49417114257812, 95.79647827148438, -6.689460754394531, 30.927871704101562, 47.182525634765625, 12.882057189941406, 133.19598388671875, 4.5626373291015625, -2.5524864196777344, 190.4180145263672, 2.184539794921875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000531.npy"}
{"epoch": 0.8027210884353742, "step": 532, "batch_size": 64, "mean": 49.61630630493164, "std": 87.505126953125, "min": -161.22994995117188, "p10": -60.70648956298827, "median": 26.76003646850586, "p90": 172.46218719482422, "max": 193.20167541503906, "pos_frac": 0.734375, "sample": [148.65916442871094, 120.35801696777344, 170.82823181152344, 156.7669677734375, 30.134567260742188, -12.923511505126953, 0.03857421875, 189.96060180664062, 8.529533386230469, 34.153778076171875, -20.15667724609375, 110.76156616210938, 172.72003173828125, 51.55810546875, 26.14287567138672, 16.672225952148438, 38.968040466308594, 165.44097900390625, 18.154098510742188, -115.83808135986328, -66.3760986328125, -1.9435691833496094, 145.83177185058594, 83.24577331542969, 1.119842529296875, 9.867828369140625, 22.63286590576172, 171.8605499267578, 153.34555053710938, 164.55715942382812, 38.47857666015625, -7.132469177246094, 24.462188720703125, 191.99732971191406, 11.007011413574219, 27.377197265625, 183.8905029296875, -18.440452575683594, -3.6908111572265625, 59.98220443725586, -75.05064392089844, -47.47740173339844, -80.65052032470703, -6.823146820068359, 187.79525756835938, 10.05596923828125, 188.6674041748047, 72.91492462158203, 19.8939208984375, 129.88723754882812, 47.69426727294922, 109.36234283447266, 1.2178573608398438, 193.20167541503906, -161.22994995117188, 35.4626579284668, -105.45880126953125, 8.573333740234375, 169.47677612304688, -85.71017456054688, -2.326099395751953, -21.60184097290039, 19.7578125, 64.80879211425781], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000532.npy"}
{"epoch": 0.8042328042328042, "step": 533, "batch_size": 64, "mean": 48.72636413574219, "std": 88.21988677978516, "min": -170.35516357421875, "p10": -55.09531478881834, "median": 36.13223648071289, "p90": 179.20318603515625, "max": 202.6968994140625, "pos_frac": 0.703125, "sample": [23.18347930908203, 175.4124298095703, -3.8540496826171875, 42.25434112548828, 50.943603515625, 72.31794738769531, -19.054412841796875, -68.13563537597656, 80.08492279052734, 120.52337646484375, -11.824432373046875, -72.78128051757812, 20.04494857788086, -10.08155632019043, 159.8719940185547, 74.96792602539062, 11.72265625, 63.74128723144531, -80.59304809570312, -1.0868911743164062, 6.773128509521484, 2.828266143798828, 111.44224548339844, 13.251296997070312, -4.4421234130859375, -39.74704360961914, 101.7784652709961, -6.32635498046875, 37.46411895751953, -170.35516357421875, 4.6733551025390625, 68.97770690917969, 186.65496826171875, 156.89291381835938, -4.663810729980469, 154.07174682617188, -1.0205841064453125, 199.25978088378906, 6.527536392211914, -61.67314529418945, 42.133140563964844, 154.23590087890625, 48.03318786621094, 176.46090698242188, 202.6968994140625, 188.65036010742188, 34.80035400390625, -95.82714080810547, 64.13552856445312, -33.3734130859375, 1.6909523010253906, 176.23159790039062, 95.54248046875, 5.389671325683594, 46.200592041015625, -32.269500732421875, 19.47290802001953, 149.5062255859375, 5.966461181640625, 188.2001953125, 180.37844848632812, 81.24432373046875, -153.28993225097656, 182.25201416015625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000533.npy"}
{"epoch": 0.8057445200302343, "step": 534, "batch_size": 64, "mean": 55.12841033935547, "std": 83.3495864868164, "min": -136.02291870117188, "p10": -39.96037292480469, "median": 39.2275333404541, "p90": 171.8565200805664, "max": 207.92019653320312, "pos_frac": 0.703125, "sample": [75.01033020019531, -4.880603790283203, 130.38589477539062, 108.25733947753906, 9.194217681884766, 100.8731689453125, 23.331527709960938, -71.22149658203125, 168.12115478515625, -1.6681594848632812, -4.061422348022461, -16.33277130126953, 130.97393798828125, 172.36563110351562, -65.14335632324219, 196.47476196289062, 83.25224304199219, 93.29884338378906, 106.60541534423828, -12.795478820800781, 178.28799438476562, 183.20994567871094, 49.20200729370117, 122.29103088378906, -5.4057769775390625, -37.19342041015625, -136.02291870117188, 17.759475708007812, 155.75527954101562, 26.781517028808594, 12.661613464355469, -105.11894226074219, 177.41685485839844, 191.16616821289062, 143.65609741210938, 26.115219116210938, -131.18435668945312, 36.668792724609375, 207.92019653320312, 170.66859436035156, 139.75674438476562, 133.7513427734375, -41.146209716796875, -0.7751617431640625, 35.371131896972656, 108.11289978027344, -3.0358314514160156, 56.674076080322266, 40.04730224609375, 13.814071655273438, 67.64889526367188, 82.61433410644531, 4.501106262207031, -11.014106750488281, 3.894500732421875, 71.76376342773438, 138.28201293945312, -56.55964660644531, 38.40776443481445, -12.012443542480469, 85.04069519042969, 32.537261962890625, 129.5075225830078, -35.64056396484375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000534.npy"}
{"epoch": 0.8072562358276644, "step": 535, "batch_size": 64, "mean": 42.33102035522461, "std": 89.57905578613281, "min": -185.6868896484375, "p10": -49.15388183593748, "median": 23.664752960205078, "p90": 187.3534362792969, "max": 204.0084228515625, "pos_frac": 0.734375, "sample": [4.1491241455078125, -56.35076904296875, 204.0084228515625, 39.40532684326172, -5.102508544921875, 39.78355407714844, -136.29708862304688, 24.3240966796875, -68.33187866210938, 188.81710815429688, 20.308563232421875, 19.73822784423828, 41.50257873535156, 133.13510131835938, -6.738834381103516, -6.1466217041015625, -17.259963989257812, 66.5563735961914, 1.4962615966796875, -32.36114501953125, 26.623451232910156, 101.1274642944336, 70.0309829711914, -15.502487182617188, 145.22311401367188, 4.266654968261719, 156.97598266601562, -172.215576171875, -13.3297119140625, 25.939245223999023, 190.6175537109375, 191.39645385742188, 85.342529296875, 67.24115753173828, 9.258472442626953, 72.98298645019531, 15.604873657226562, 192.30152893066406, -107.49525451660156, 134.9601593017578, 45.81884765625, -1.9441280364990234, 202.74774169921875, 21.677749633789062, 52.94151306152344, -185.6868896484375, 46.923492431640625, 164.2547607421875, 1.4112701416015625, 141.96099853515625, 1.060211181640625, -2.5904083251953125, 4.979114532470703, 192.13037109375, 19.92021942138672, 23.005409240722656, 24.44072723388672, 48.37348937988281, 175.611328125, -11.437332153320312, -94.42938232421875, 0.0171051025390625, 183.93820190429688, 18.0751953125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000535.npy"}
{"epoch": 0.8087679516250945, "step": 536, "batch_size": 64, "mean": 61.04181671142578, "std": 73.70304107666016, "min": -74.32759857177734, "p10": -3.9705570220947264, "median": 32.37965393066406, "p90": 180.855810546875, "max": 200.98184204101562, "pos_frac": 0.84375, "sample": [23.68106460571289, 2.8637847900390625, -74.32759857177734, 180.09817504882812, 39.409912109375, -57.99212646484375, 37.327308654785156, -3.7125701904296875, 78.7637710571289, 2.467517852783203, -5.411109924316406, 79.25786590576172, 185.88790893554688, 119.05439758300781, 130.5590362548828, 115.09355926513672, 22.04571533203125, 187.19764709472656, 159.85621643066406, 15.300657272338867, 5.3436279296875, 14.32400131225586, 173.54022216796875, 105.2843017578125, 2.6573028564453125, 196.12255859375, 149.46044921875, 7.209449768066406, 6.763130187988281, 153.25436401367188, 183.17941284179688, 1.6418037414550781, 2.0944442749023438, 22.618824005126953, 62.337486267089844, 43.787353515625, 168.5428466796875, -13.83163833618164, 99.3345947265625, 0.8603591918945312, 109.3283462524414, 44.722938537597656, -0.06440544128417969, -4.542104721069336, 2.7606239318847656, 12.024246215820312, 42.44777297973633, 50.50935363769531, 27.43199920654297, 200.98184204101562, -4.0723419189453125, 53.58302307128906, 10.4234619140625, 181.18051147460938, 131.91513061523438, 8.44204330444336, -3.7330589294433594, 192.99734497070312, 0.8311653137207031, 0.6670684814453125, 59.72040557861328, 174.3448028564453, -9.633628845214844, 2.4640274047851562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000536.npy"}
{"epoch": 0.8102796674225246, "step": 537, "batch_size": 64, "mean": 48.52098846435547, "std": 85.2042007446289, "min": -151.1260223388672, "p10": -40.493727111816405, "median": 14.829193115234375, "p90": 175.85051574707032, "max": 204.88580322265625, "pos_frac": 0.671875, "sample": [96.21290588378906, 9.802375793457031, -17.744735717773438, 47.3734130859375, 204.88580322265625, 158.3429412841797, 185.9864044189453, 1.2284622192382812, -1.582550048828125, 68.08320617675781, 148.09564208984375, 4.208580017089844, 184.27285766601562, 122.46656799316406, 134.48312377929688, 97.5780029296875, 190.01519775390625, 14.360454559326172, 6.268421173095703, -0.5655593872070312, -10.089719772338867, 14.71966552734375, -3.0394439697265625, -60.9172248840332, -40.8560791015625, 171.4283905029297, 12.87225341796875, 116.05384826660156, -34.29877471923828, -5.450340270996094, 97.14099884033203, 43.51995849609375, 98.42942810058594, -9.627494812011719, -48.821685791015625, 5.0062255859375, 201.43338012695312, 53.21894073486328, 1.291748046875, -6.657402038574219, 59.510772705078125, 186.5619354248047, -95.76141357421875, -30.53942108154297, 11.107887268066406, -10.195114135742188, 22.859634399414062, -38.815269470214844, 54.22900390625, 35.348976135253906, 8.164093017578125, 158.9006805419922, -39.64823913574219, 146.07171630859375, 163.5052490234375, 177.74571228027344, -151.1260223388672, 38.954864501953125, -69.26573944091797, 14.938720703125, 137.84817504882812, 170.72650146484375, -77.89067077636719, -17.01683807373047], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000537.npy"}
{"epoch": 0.8117913832199547, "step": 538, "batch_size": 64, "mean": 55.653079986572266, "std": 86.61102294921875, "min": -127.32811737060547, "p10": -38.840789794921875, "median": 23.636302947998047, "p90": 180.32415008544922, "max": 221.0658416748047, "pos_frac": 0.75, "sample": [-90.93179321289062, 100.8113784790039, -20.04001235961914, 15.813247680664062, 153.8939971923828, -127.32811737060547, 131.14471435546875, -9.928596496582031, -6.0295257568359375, 150.14010620117188, 137.35067749023438, 184.71409606933594, -49.20574951171875, 169.50564575195312, 4.433742523193359, 14.913551330566406, 166.5795440673828, 170.3031005859375, 11.165748596191406, 7.128358840942383, 64.98043823242188, -3.4465465545654297, 144.04217529296875, 59.91192626953125, 27.585433959960938, -35.360870361328125, 181.83181762695312, 14.2862548828125, 185.656982421875, 10.112503051757812, 60.9334716796875, 117.61654663085938, 5.59661865234375, 8.69940185546875, 2.966583251953125, 1.4836196899414062, -70.38169860839844, 24.88367462158203, -4.065605163574219, 118.8193588256836, 176.9442596435547, 31.004425048828125, 184.09359741210938, 36.08595275878906, -60.32766342163086, 157.13880920410156, -7.058380126953125, 181.77267456054688, 139.6851806640625, 221.0658416748047, 17.612842559814453, 67.87739562988281, -119.069091796875, -40.332183837890625, 91.74502563476562, -3.568511962890625, 208.9170379638672, 0.11029815673828125, 60.731689453125, -19.69970703125, 22.388931274414062, 165.6034698486328, 1.3556900024414062, 17.133407592773438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000538.npy"}
{"epoch": 0.8133030990173847, "step": 539, "batch_size": 64, "mean": 44.81317138671875, "std": 77.67262268066406, "min": -132.49684143066406, "p10": -28.969389724731435, "median": 23.080495834350586, "p90": 169.67456512451173, "max": 198.88259887695312, "pos_frac": 0.71875, "sample": [69.65125274658203, 170.68182373046875, -10.475410461425781, 191.46876525878906, 23.79851531982422, 31.360504150390625, 162.1407470703125, 1.851898193359375, 161.79095458984375, -4.933704376220703, 112.32440185546875, 15.8548583984375, 2.2591094970703125, 2.0286712646484375, -33.46220397949219, 12.852611541748047, -85.866943359375, 5.7964019775390625, 188.18447875976562, 42.883338928222656, 65.47262573242188, -3.4540882110595703, 35.32684326171875, -132.49684143066406, -2.3131866455078125, -64.22676086425781, -18.486156463623047, 84.57121276855469, 190.3731231689453, 68.74114227294922, 2.05633544921875, -2.2576904296875, -36.84147644042969, -3.878082275390625, 190.68423461914062, 77.41030883789062, 18.94402313232422, 7.445068359375, 184.6639404296875, 36.00025939941406, 3.1223373413085938, -15.083404541015625, 135.45120239257812, -16.871780395507812, 160.22100830078125, 55.785072326660156, -33.99127960205078, 198.88259887695312, 1.0908966064453125, -8.118125915527344, 149.8189239501953, 36.27984619140625, 22.362476348876953, -9.288360595703125, 115.18753051757812, 28.260955810546875, 15.25860595703125, 10.0701904296875, -96.21978759765625, 88.7406997680664, 36.80416488647461, 167.3242950439453, 30.916763305664062, 34.11304473876953], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000539.npy"}
{"epoch": 0.8148148148148148, "step": 540, "batch_size": 64, "mean": 56.455596923828125, "std": 87.75103759765625, "min": -164.16766357421875, "p10": -46.316155242919905, "median": 40.596370697021484, "p90": 185.376969909668, "max": 204.47360229492188, "pos_frac": 0.78125, "sample": [-29.840782165527344, 63.479217529296875, 68.34817504882812, 186.4866485595703, 3.932781219482422, -146.73260498046875, 25.090057373046875, 73.56546020507812, 32.671356201171875, 15.833675384521484, 63.46260452270508, 29.5413818359375, 156.73715209960938, 110.43441009521484, -74.3800277709961, 39.744102478027344, 160.00900268554688, 73.25408172607422, -1.4846611022949219, 114.53434753417969, 179.543212890625, 1.4103546142578125, 99.29727172851562, 173.3662567138672, 196.98361206054688, 81.57259368896484, -63.43899154663086, 96.59111785888672, -53.37702941894531, 0.260284423828125, -10.54534912109375, 182.7877197265625, 175.4569549560547, 41.448638916015625, 204.47360229492188, 2.199237823486328, -27.757965087890625, 175.22781372070312, -72.09303283691406, 1.427499771118164, -1.3113021850585938, 169.29556274414062, 60.144813537597656, 190.70034790039062, 194.70640563964844, -2.1019058227539062, 57.72882080078125, 39.44849395751953, -18.543350219726562, -164.16766357421875, 13.721214294433594, 60.032989501953125, 1.21160888671875, 24.1790771484375, 12.772945404052734, 27.00177001953125, -72.41486358642578, 194.97296142578125, 1.439361572265625, 67.97162628173828, 64.69145202636719, 10.497591018676758, 141.94247436523438, 189.7174072265625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000540.npy"}
{"epoch": 0.8163265306122449, "step": 541, "batch_size": 64, "mean": 64.70645141601562, "std": 79.88262176513672, "min": -70.96732330322266, "p10": -36.66834602355956, "median": 49.04121208190918, "p90": 174.36521606445314, "max": 186.90084838867188, "pos_frac": 0.765625, "sample": [165.54434204101562, 166.210693359375, 44.7757568359375, 185.51670837402344, -6.673225402832031, 173.35398864746094, -70.96732330322266, 174.79859924316406, 151.17584228515625, -11.252326965332031, 22.787918090820312, -59.503753662109375, 6.437248229980469, 64.18305969238281, 4.826271057128906, -68.89451599121094, 130.457763671875, 152.00778198242188, 32.19537353515625, 184.30242919921875, -13.209754943847656, -10.038167953491211, 183.3794403076172, 122.53678894042969, 144.72531127929688, 50.77140808105469, 145.04322814941406, -57.992652893066406, -41.427310943603516, 1.1494693756103516, 153.15948486328125, 33.65422058105469, 176.85928344726562, 70.59967041015625, -20.445709228515625, 121.08297729492188, 3.057271957397461, -56.927093505859375, 48.4222412109375, 60.040382385253906, 49.66018295288086, -10.590827941894531, 74.68550109863281, 146.5267333984375, 17.633682250976562, 21.047264099121094, 142.01248168945312, 138.81912231445312, -25.56409454345703, 108.44935607910156, 1.1937437057495117, 120.12287902832031, 24.781452178955078, 12.099220275878906, -16.131690979003906, 39.4498291015625, 169.5133056640625, 120.07794189453125, 3.477224349975586, 103.99964141845703, 182.80780029296875, -48.31647491455078, 186.90084838867188, 22.834548950195312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000541.npy"}
{"epoch": 0.817838246409675, "step": 542, "batch_size": 64, "mean": 69.86752319335938, "std": 89.23687744140625, "min": -178.46231079101562, "p10": -29.007142639160154, "median": 50.84021186828613, "p90": 188.06854095458985, "max": 201.85299682617188, "pos_frac": 0.765625, "sample": [178.57366943359375, 167.67758178710938, 9.731536865234375, 188.220458984375, -2.148517608642578, 10.225425720214844, 38.399261474609375, 130.59146118164062, 64.38023376464844, 65.95549774169922, 80.4105224609375, -7.215644836425781, -39.04408264160156, 176.63662719726562, 16.09441375732422, 105.97969055175781, 30.512657165527344, 118.14344787597656, -17.205810546875, 185.13673400878906, 188.31564331054688, 194.49252319335938, 3.524801254272461, 1.2593612670898438, -76.78707885742188, -178.46231079101562, 1.3147773742675781, 7.1661376953125, 142.34100341796875, -28.118759155273438, 184.58164978027344, 189.41404724121094, 106.42878723144531, 136.61111450195312, -73.60045623779297, -6.381965637207031, -0.788360595703125, 49.18208312988281, -29.38787841796875, 164.0792236328125, 19.66461944580078, 22.922077178955078, 179.03048706054688, -51.243186950683594, 121.06358337402344, -0.6631011962890625, 187.7140655517578, 52.49834060668945, 56.60216522216797, 201.85299682617188, 171.18185424804688, 189.92822265625, -71.75381469726562, -1.1433258056640625, 18.84386444091797, 10.641220092773438, 30.054988861083984, 190.8678436279297, 125.59121704101562, 45.24358367919922, 48.712318420410156, 175.951904296875, 97.98564147949219, 173.73452758789062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000542.npy"}
{"epoch": 0.8193499622071051, "step": 543, "batch_size": 64, "mean": 56.92228698730469, "std": 91.6534423828125, "min": -195.70631408691406, "p10": -20.48514823913574, "median": 25.176929473876953, "p90": 186.45392150878908, "max": 210.70730590820312, "pos_frac": 0.75, "sample": [126.910400390625, 24.694828033447266, -17.705429077148438, 0.9063587188720703, 18.24481201171875, -17.39950942993164, 131.4998779296875, 194.1056365966797, -139.99673461914062, 111.06243896484375, 26.685287475585938, 29.65392303466797, 162.3190460205078, 24.742942810058594, 181.50558471679688, 21.074417114257812, 184.11996459960938, -164.70126342773438, 7.886760711669922, 179.35694885253906, -1.7548961639404297, -0.23204421997070312, 18.202850341796875, 87.75658416748047, 15.327045440673828, 184.94381713867188, -195.70631408691406, 1.9147224426269531, -33.323368072509766, -65.16348266601562, 14.2796630859375, 161.05056762695312, 191.44097900390625, 69.93206787109375, 158.59066772460938, -0.8713226318359375, 187.1855926513672, -2.316192626953125, -21.676456451416016, 0.7901363372802734, 25.610916137695312, 4.263919830322266, 210.70730590820312, 161.02188110351562, 100.81539916992188, 65.89894104003906, 185.40155029296875, 206.31167602539062, 17.59728240966797, 95.65670776367188, -7.148231506347656, 59.54615783691406, -1.4593353271484375, 7.0059356689453125, 100.49197387695312, 9.72406005859375, 76.24235534667969, 52.93072509765625, 40.24561309814453, -3.4121360778808594, 186.90493774414062, 196.23007202148438, -27.00926971435547, 24.110881805419922], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000543.npy"}
{"epoch": 0.8208616780045351, "step": 544, "batch_size": 64, "mean": 59.45600891113281, "std": 89.49799346923828, "min": -132.64801025390625, "p10": -52.60212173461913, "median": 44.48595428466797, "p90": 188.90399475097658, "max": 205.60305786132812, "pos_frac": 0.6875, "sample": [44.974822998046875, -96.73503112792969, 17.613601684570312, -0.6777439117431641, 123.95045471191406, 37.635765075683594, -2.9443283081054688, 4.441963195800781, 18.555023193359375, 30.77269744873047, -16.978073120117188, 58.74542999267578, 35.614471435546875, 145.99700927734375, 49.204193115234375, 0.39762115478515625, -72.39251708984375, 88.07351684570312, 24.394546508789062, -2.8482208251953125, 84.92737579345703, 54.97581481933594, 33.791656494140625, 190.0560302734375, 190.31942749023438, -19.128843307495117, 62.20281982421875, -74.21888732910156, -5.011962890625, -37.371063232421875, 205.60305786132812, -132.64801025390625, 157.6724853515625, -1.6331825256347656, 184.23556518554688, 102.20370483398438, 156.6478271484375, -60.1534423828125, 129.1690673828125, 178.740966796875, 92.64779663085938, 118.14369201660156, 194.31436157226562, 12.801254272460938, -57.31269073486328, 127.90056610107422, 78.20236206054688, 80.28472900390625, 190.31509399414062, -98.53868865966797, -13.760454177856445, 194.31179809570312, 202.6618194580078, 168.0505828857422, -34.053192138671875, -20.339920043945312, 21.657333374023438, -41.61079406738281, 167.943359375, 182.08108520507812, -13.64847183227539, 43.99708557128906, 134.74447631835938, 186.21591186523438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000544.npy"}
{"epoch": 0.8223733938019653, "step": 545, "batch_size": 64, "mean": 42.74488067626953, "std": 87.78929138183594, "min": -150.3452911376953, "p10": -60.72207832336422, "median": 14.931385040283203, "p90": 171.30422973632812, "max": 219.74777221679688, "pos_frac": 0.671875, "sample": [132.41738891601562, 45.655418395996094, -20.835708618164062, 178.44598388671875, 13.949718475341797, 58.412559509277344, 7.6562652587890625, -74.65565490722656, -20.366580963134766, -0.9566192626953125, 119.27587890625, -8.858551025390625, 137.8099365234375, -2.00323486328125, 5.912040710449219, 38.26624298095703, 39.627471923828125, 160.32948303222656, -5.524833679199219, 149.23997497558594, 3.6542587280273438, 219.74777221679688, -11.011955261230469, 172.89247131347656, 42.733245849609375, 51.67692565917969, 104.32316589355469, 75.3935546875, 5.319831848144531, 19.887496948242188, -28.210399627685547, -150.3452911376953, -5.7576446533203125, 88.97965240478516, -101.58790588378906, -6.073101043701172, 141.03518676757812, 194.06857299804688, 36.581756591796875, -11.690704345703125, -2.8691253662109375, 178.41952514648438, 120.6827163696289, 14.824165344238281, 7.420604705810547, 12.251808166503906, -93.56624603271484, 129.34246826171875, 59.257415771484375, -120.21577453613281, 4.4851837158203125, 124.26583862304688, 2.9467315673828125, 202.98321533203125, 111.3947982788086, -140.63011169433594, -2.5461349487304688, 102.68067169189453, -125.47494506835938, -1.9736061096191406, 15.038604736328125, 172.73117065429688, 167.97470092773438, 0.8346786499023438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000545.npy"}
{"epoch": 0.8238851095993953, "step": 546, "batch_size": 64, "mean": 46.011558532714844, "std": 79.89759063720703, "min": -123.45346069335938, "p10": -33.030906677246094, "median": 21.23223114013672, "p90": 170.7607849121094, "max": 201.45008850097656, "pos_frac": 0.734375, "sample": [170.83477783203125, -23.45098876953125, 181.8831024169922, 12.977706909179688, 28.28866195678711, 67.18470001220703, 17.39270782470703, -10.752029418945312, 6.455169677734375, -18.130191802978516, -38.681884765625, 158.02899169921875, 54.626522064208984, 27.086933135986328, -123.45346069335938, 142.33929443359375, -1.1112442016601562, 0.7881889343261719, -29.254119873046875, 30.10446548461914, -15.381721496582031, 14.121660232543945, 6.772449493408203, 31.91717529296875, 25.359447479248047, 6.93951416015625, 170.588134765625, -88.518310546875, 139.46237182617188, 201.45008850097656, 54.561553955078125, 57.312530517578125, 156.75015258789062, 11.176040649414062, 3.9394454956054688, -45.23207092285156, 20.708953857421875, 184.89065551757812, 17.511760711669922, -31.097244262695312, 2.494647979736328, 21.755508422851562, -74.92835998535156, 181.6107940673828, 2.9581871032714844, 49.090476989746094, -19.186901092529297, 6.494991302490234, 92.60076904296875, 23.210891723632812, 108.61027526855469, 64.05329132080078, 198.246337890625, -33.859619140625, -1.8211593627929688, -3.6843223571777344, 139.6920166015625, 56.04743194580078, 193.0518035888672, 126.7331771850586, 163.53346252441406, -84.4830322265625, 148.7837371826172, 7.345468521118164], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000546.npy"}
{"epoch": 0.8253968253968254, "step": 547, "batch_size": 64, "mean": 65.12631225585938, "std": 87.38771057128906, "min": -126.86322021484375, "p10": -51.06466445922849, "median": 65.67594146728516, "p90": 178.07215423583986, "max": 206.2588348388672, "pos_frac": 0.828125, "sample": [7.932636260986328, 2.3074493408203125, -0.01265716552734375, 6.4263763427734375, 163.65463256835938, 92.13516235351562, -126.86322021484375, 206.2588348388672, 147.14889526367188, 167.8837127685547, 67.70423889160156, 9.828216552734375, -62.62171173095703, 74.14879608154297, 93.63363647460938, -8.406177520751953, 174.6393280029297, 183.128662109375, -108.31266021728516, 2.2527313232421875, 189.0166473388672, 149.244384765625, 191.02133178710938, 15.610595703125, 148.48487854003906, 80.6847915649414, 63.64764404296875, 133.299560546875, 28.742767333984375, 36.449180603027344, 13.151206970214844, 0.53851318359375, -99.4882583618164, 130.53509521484375, 27.43254852294922, 45.66069030761719, 179.54336547851562, 9.888784408569336, 127.86692810058594, 86.84710693359375, 149.61508178710938, 160.08807373046875, 72.23522186279297, 155.55238342285156, -84.78213500976562, 43.125091552734375, 46.021244049072266, -73.3817138671875, -24.098220825195312, 146.12216186523438, 181.52345275878906, 126.78885650634766, 164.17156982421875, 2.6811294555664062, 75.56938171386719, -118.963134765625, 15.073596954345703, 122.38330841064453, 29.969764709472656, 1.9300155639648438, 16.469417572021484, 150.66500854492188, 181.183837890625, -22.904144287109375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000547.npy"}
{"epoch": 0.8269085411942555, "step": 548, "batch_size": 64, "mean": 43.605491638183594, "std": 83.42790985107422, "min": -161.27896118164062, "p10": -51.40347671508788, "median": 25.280799865722656, "p90": 171.2315689086914, "max": 202.35023498535156, "pos_frac": 0.71875, "sample": [47.42625427246094, 99.24330139160156, -7.227447509765625, -3.2291717529296875, -34.97551727294922, 143.45803833007812, 0.6110305786132812, 168.88223266601562, 126.10576629638672, 175.74114990234375, -5.126556396484375, 150.04837036132812, -1.6939420700073242, 19.762557983398438, 0.2380523681640625, 14.0302734375, -129.6793212890625, 12.730777740478516, 9.984512329101562, 36.06938171386719, 56.96376037597656, 119.03868103027344, -74.40695190429688, 121.63417053222656, 41.378082275390625, 72.39837646484375, 194.10304260253906, 171.62928771972656, 104.59115600585938, 104.17206573486328, 17.050148010253906, 41.23971176147461, 15.111595153808594, 93.2365951538086, 7.4935455322265625, 22.98858642578125, 179.51400756835938, 17.06588363647461, -94.51212310791016, 18.566566467285156, 89.80410766601562, 2.2298507690429688, 73.209716796875, -22.502655029296875, 202.35023498535156, 39.42087936401367, -4.9609832763671875, -41.74298095703125, -78.93992614746094, 61.0341796875, 35.0301513671875, -77.88029479980469, 0.2093372344970703, 127.6150894165039, 27.573013305664062, 76.87527465820312, 170.30355834960938, -18.016502380371094, -33.910552978515625, -53.78504943847656, 197.33975219726562, 174.96463012695312, -161.27896118164062, -45.846473693847656], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000548.npy"}
{"epoch": 0.8284202569916855, "step": 549, "batch_size": 64, "mean": 42.149147033691406, "std": 92.90361022949219, "min": -184.97158813476562, "p10": -59.60014572143554, "median": 25.356876373291016, "p90": 179.48594970703127, "max": 202.7900390625, "pos_frac": 0.703125, "sample": [7.0048828125, 18.637054443359375, -44.875518798828125, 28.772109985351562, -43.75592803955078, -48.583343505859375, 124.77491760253906, 4.706184387207031, 26.71429443359375, 186.19515991210938, 202.7900390625, 0.92919921875, 23.21304702758789, 179.78915405273438, -40.427581787109375, -10.565216064453125, -157.31600952148438, 21.72339630126953, 158.22509765625, 80.6387939453125, 134.44882202148438, 87.37956237792969, 49.4554443359375, -126.2385482788086, 15.376052856445312, 16.106712341308594, 38.890037536621094, 189.39047241210938, 139.22821044921875, -64.3216323852539, 144.0286102294922, 198.19386291503906, 48.6343994140625, -33.11254119873047, 115.36204528808594, 178.77847290039062, -0.6060295104980469, -30.30987548828125, 42.070098876953125, 102.62578582763672, -90.539794921875, 3.0767822265625, -1.328104019165039, 69.84286499023438, 25.47466278076172, 115.08622741699219, -11.862800598144531, 93.18281555175781, 189.08151245117188, 25.239089965820312, -5.683692932128906, 160.58291625976562, 54.093658447265625, 9.421401977539062, 94.56587219238281, 19.976051330566406, 4.189292907714844, -0.2628936767578125, -68.73492431640625, 172.30990600585938, 184.267822265625, -184.97158813476562, -178.01626586914062, 54.58482360839844], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000549.npy"}
{"epoch": 0.8299319727891157, "step": 550, "batch_size": 64, "mean": 40.88159942626953, "std": 91.70925903320312, "min": -168.38162231445312, "p10": -66.09005508422851, "median": 14.18609619140625, "p90": 167.35958251953127, "max": 193.88165283203125, "pos_frac": 0.671875, "sample": [131.58038330078125, 179.99932861328125, 193.40536499023438, 80.45527648925781, -148.21136474609375, -19.86553192138672, -0.3374519348144531, -68.34806823730469, -6.861106872558594, 4.176177978515625, 169.72268676757812, 34.67809295654297, 61.339881896972656, 4.143129348754883, -3.27496337890625, 173.0177459716797, 193.88165283203125, 120.85986328125, 11.154754638671875, 13.9981689453125, 152.771240234375, -8.950332641601562, 96.14183044433594, 161.84567260742188, 139.76866149902344, -60.82135772705078, 127.64739990234375, -30.819202423095703, 50.01834487915039, -46.026641845703125, -148.11883544921875, 11.089973449707031, 5.238006591796875, 67.09978485107422, 9.102283477783203, 123.44535064697266, 0.849639892578125, 13.845680236816406, -14.699737548828125, 6.359352111816406, 128.26268005371094, 159.00765991210938, 14.3740234375, 111.73126220703125, -3.9422149658203125, 53.77083206176758, 28.405670166015625, 71.02623748779297, -41.71916961669922, 188.71157836914062, -85.48567962646484, 146.93270874023438, 29.48822021484375, 114.64694213867188, 136.81881713867188, 6.011035919189453, -138.7728271484375, -0.9502677917480469, 173.50982666015625, -168.38162231445312, 70.07400512695312, -4.167869567871094, -128.31370544433594, -25.91712188720703], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000550.npy"}
{"epoch": 0.8314436885865457, "step": 551, "batch_size": 64, "mean": 43.25688934326172, "std": 74.91301727294922, "min": -187.32012939453125, "p10": -16.05178241729736, "median": 26.844501495361328, "p90": 156.43879089355468, "max": 264.5069580078125, "pos_frac": 0.765625, "sample": [8.081153869628906, 106.51902770996094, 90.15306091308594, -8.233352661132812, 138.68951416015625, 83.71426391601562, 264.5069580078125, 5.207622528076172, 27.61309051513672, 86.0196304321289, -11.638740539550781, 6.674430847167969, 171.1838836669922, 20.256526947021484, 169.2554931640625, 56.234413146972656, 27.133453369140625, 155.2969970703125, -187.32012939453125, -0.4829826354980469, 7.806663513183594, -7.450355529785156, 156.92813110351562, 17.798553466796875, 52.72357940673828, 200.34939575195312, -45.024620056152344, 69.9764404296875, 26.55554962158203, 49.27766799926758, 34.553306579589844, 68.21195983886719, 0.2741661071777344, 41.59498596191406, 17.870346069335938, 120.0323486328125, -24.87865447998047, 72.4019775390625, 1.2876205444335938, 4.08404541015625, 7.23944091796875, 67.18342590332031, 17.20233917236328, -102.07681274414062, 172.34365844726562, 108.54088592529297, 61.46998596191406, 96.56509399414062, 2.441608428955078, -4.232643127441406, 10.821807861328125, 40.458221435546875, 24.025482177734375, -49.12810516357422, -0.798980712890625, -13.983261108398438, -11.977645874023438, 168.72488403320312, 48.31593322753906, 6.224937438964844, -85.76383972167969, -16.938291549682617, 110.02880859375, 38.516845703125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000551.npy"}
{"epoch": 0.8329554043839759, "step": 552, "batch_size": 64, "mean": 43.128273010253906, "std": 86.80941009521484, "min": -187.80831909179688, "p10": -60.28287582397459, "median": 35.76553535461426, "p90": 144.28108673095704, "max": 197.37855529785156, "pos_frac": 0.734375, "sample": [-88.62307739257812, -5.807769775390625, -42.83466339111328, -3.53369140625, 131.3121337890625, 2.3093204498291016, -2.646472930908203, 25.052581787109375, 158.19931030273438, 116.12544250488281, 109.1680908203125, 197.37855529785156, 84.8301773071289, 158.3365478515625, -3.376018524169922, -13.936277389526367, 2.8040428161621094, 7.251457214355469, 144.3730010986328, 27.895980834960938, 84.93455505371094, 144.06661987304688, 31.111251831054688, 112.61593627929688, 180.8018035888672, 6.479972839355469, 3.1068038940429688, 28.018321990966797, 129.031005859375, -134.9757843017578, 12.320344924926758, 0.2820892333984375, 111.20279693603516, -173.8789825439453, 137.14511108398438, -70.95042419433594, 40.03260803222656, 119.7110595703125, -15.788890838623047, -0.8344497680664062, 160.8980255126953, 7.0779266357421875, -67.76068115234375, 60.3292236328125, 43.701812744140625, 18.89666748046875, 48.66266632080078, 120.53911590576172, -187.53225708007812, 85.63258361816406, 184.10968017578125, 61.9658203125, 32.5953483581543, -187.80831909179688, 133.10443115234375, 114.37503051757812, 82.60411071777344, 38.93572235107422, 135.8530731201172, 49.18213653564453, 23.34935760498047, -1.13397216796875, 79.25267791748047, -25.331005096435547], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000552.npy"}
{"epoch": 0.8344671201814059, "step": 553, "batch_size": 64, "mean": 51.320491790771484, "std": 99.87693786621094, "min": -182.69229125976562, "p10": -70.5093879699707, "median": 42.35885429382324, "p90": 179.9053955078125, "max": 202.16676330566406, "pos_frac": 0.671875, "sample": [202.16676330566406, -3.6079177856445312, -72.3613052368164, 171.4139404296875, -24.496170043945312, -17.111934661865234, 135.10877990722656, 178.58050537109375, -0.43025970458984375, -5.370370864868164, 39.86906051635742, 57.02154541015625, 195.43606567382812, 148.2538604736328, 12.421112060546875, 44.84864807128906, -8.52176284790039, 155.18673706054688, -16.018550872802734, -182.69229125976562, -6.037517547607422, 142.0694580078125, 55.29518127441406, 58.42313003540039, 130.7188720703125, 28.609939575195312, 78.39649963378906, 15.817947387695312, -0.9932098388671875, -62.141845703125, 197.86795043945312, 178.44261169433594, 95.35429382324219, 199.09088134765625, 120.03993225097656, 109.23883056640625, 170.04348754882812, 20.724578857421875, 6.3666839599609375, 180.47320556640625, 87.64783477783203, 3.7455062866210938, 17.459142684936523, 44.91291809082031, -137.99502563476562, -66.18824768066406, 0.6129150390625, 196.51834106445312, -164.79647827148438, -136.97962951660156, 160.9277801513672, 131.66302490234375, 167.16403198242188, -1.4294586181640625, -163.75558471679688, 110.22545623779297, 96.6303939819336, -26.97515869140625, 97.72796630859375, 35.29595947265625, -17.560762405395508, -81.67430114746094, 190.88067626953125, 12.956649780273438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000553.npy"}
{"epoch": 0.8359788359788359, "step": 554, "batch_size": 64, "mean": 56.50690841674805, "std": 88.63822937011719, "min": -172.86068725585938, "p10": -49.55473899841308, "median": 39.57394027709961, "p90": 180.0571060180664, "max": 195.31402587890625, "pos_frac": 0.71875, "sample": [70.74906921386719, 176.27493286132812, 193.04632568359375, 148.48397827148438, 4.408821105957031, 68.98673248291016, 7.773937225341797, 135.94012451171875, -142.14402770996094, 164.30471801757812, -3.142932891845703, -31.5843505859375, -39.761497497558594, 127.35151672363281, -172.86068725585938, 6.789684295654297, 2.034242630004883, -6.758125305175781, 82.55914306640625, 23.753732681274414, 46.805824279785156, 133.74322509765625, 83.21089172363281, -19.39349365234375, 29.805099487304688, 54.834381103515625, 188.65032958984375, 157.37071228027344, -17.909887313842773, 11.695686340332031, 195.31402587890625, -7.1152801513671875, 103.35589599609375, -73.6099853515625, 2.449604034423828, 6.580928802490234, 133.7913818359375, 32.34205627441406, 121.85326385498047, -11.88116455078125, -74.80422973632812, 118.06676483154297, 94.4522933959961, 5.661773681640625, 26.041175842285156, -38.454166412353516, 10.192472457885742, -53.80840301513672, 29.69247055053711, -4.13848876953125, 182.5311279296875, -23.7833251953125, -53.7518424987793, 180.91090393066406, 182.017578125, 107.11649322509766, 113.49871826171875, 123.00334167480469, 73.85987091064453, 164.35215759277344, 193.16607666015625, -57.96385192871094, 152.41937255859375, 178.06491088867188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000554.npy"}
{"epoch": 0.8374905517762661, "step": 555, "batch_size": 64, "mean": 52.293487548828125, "std": 97.28158569335938, "min": -191.5152587890625, "p10": -62.43030776977538, "median": 44.46229934692383, "p90": 174.93335571289063, "max": 194.2918701171875, "pos_frac": 0.796875, "sample": [1.9091720581054688, -114.6353530883789, 48.015846252441406, 125.01276397705078, 175.42074584960938, 22.040084838867188, -65.55378723144531, -44.982452392578125, -92.84737396240234, 28.166839599609375, 173.79611206054688, -40.692230224609375, 64.230712890625, 18.53079605102539, 185.55474853515625, 164.15402221679688, 5.449195861816406, 91.82040405273438, 16.592117309570312, 140.18624877929688, 40.90875244140625, 5.967979431152344, 187.55447387695312, 49.414398193359375, 108.32542419433594, 55.19677734375, 151.31680297851562, 140.97862243652344, 19.2301025390625, 192.7008056640625, 144.79388427734375, 59.877655029296875, 157.89901733398438, 184.61959838867188, -164.3386993408203, 161.89222717285156, -191.5152587890625, 132.73643493652344, -4.4845733642578125, 4.171531677246094, 166.27565002441406, 7.400444030761719, 5.1614227294921875, 19.08563995361328, -55.142189025878906, 9.690776824951172, 49.035682678222656, 19.684371948242188, 182.60777282714844, -38.63194274902344, 95.69610595703125, -47.102752685546875, -189.4938507080078, 3.4866905212402344, 159.96063232421875, 32.28692626953125, 94.64869689941406, 12.054168701171875, 160.1954345703125, -108.66348266601562, 194.2918701171875, 82.08271789550781, 128.8723602294922, 23.88552474975586], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000555.npy"}
{"epoch": 0.8390022675736961, "step": 556, "batch_size": 64, "mean": 48.65314483642578, "std": 85.75850677490234, "min": -130.6085205078125, "p10": -53.39694061279297, "median": 26.130724906921387, "p90": 184.35338439941407, "max": 209.94036865234375, "pos_frac": 0.671875, "sample": [116.7390365600586, -1.58343505859375, 16.518905639648438, 1.2258567810058594, -29.471572875976562, 178.58441162109375, 193.01771545410156, 38.245967864990234, -5.180837631225586, -46.11250305175781, -4.517185211181641, -5.322765350341797, 58.08531188964844, 102.0633316040039, 64.57568359375, 31.59186553955078, 30.40484619140625, 4.173973083496094, 131.39340209960938, 152.59329223632812, -2.6298675537109375, 81.93905639648438, 193.09178161621094, -50.16266632080078, -65.47828674316406, 1.5683746337890625, -58.16626739501953, 33.60847091674805, 101.8448257446289, -19.768508911132812, 2.030193328857422, -85.45237731933594, 47.149497985839844, 11.085731506347656, 3.562255859375, 19.87401008605957, 184.62127685546875, 187.11749267578125, 2.4975929260253906, 80.2999267578125, 183.72830200195312, 203.70611572265625, 95.74424743652344, 5.603610992431641, 128.99000549316406, 141.58538818359375, -54.783058166503906, -7.2105712890625, -130.6085205078125, -17.147354125976562, 192.7860107421875, -67.03809356689453, -20.152862548828125, 54.50352478027344, 110.35826110839844, -11.492889404296875, 44.602195739746094, -75.72023010253906, 83.7193603515625, 180.44036865234375, 175.42745971679688, 209.94036865234375, 21.856603622436523, -30.694910049438477], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000556.npy"}
{"epoch": 0.8405139833711263, "step": 557, "batch_size": 64, "mean": 62.17573928833008, "std": 83.51129150390625, "min": -88.779296875, "p10": -28.599967193603508, "median": 39.610822677612305, "p90": 191.4623794555664, "max": 204.16685485839844, "pos_frac": 0.734375, "sample": [195.96426391601562, 53.21615982055664, 26.20246124267578, 2.680572509765625, 25.974445343017578, -2.1359596252441406, 193.34756469726562, -2.2909011840820312, 83.52864074707031, 168.10079956054688, -5.3760833740234375, -33.42566680908203, -1.2206859588623047, 48.989105224609375, 10.93124771118164, 29.281238555908203, 72.99703216552734, 3.7357139587402344, 197.57542419433594, 89.69019317626953, 63.35966491699219, 0.1555328369140625, 192.9055938720703, -85.32568359375, -11.949153900146484, -20.700363159179688, 170.71307373046875, 202.62432861328125, -88.779296875, -4.340568542480469, 92.709228515625, 31.615161895751953, 136.4189453125, -3.3021316528320312, -6.03265380859375, 185.8438262939453, 188.09487915039062, 80.10795593261719, 11.469474792480469, 203.15682983398438, 1.737762451171875, 92.90127563476562, -87.96226501464844, -1.2880477905273438, 33.79143524169922, -56.406280517578125, 48.881591796875, 204.16685485839844, 130.80870056152344, 124.46282196044922, 2.3681488037109375, 133.66409301757812, 35.383235931396484, 91.83290100097656, 150.1217803955078, 10.4385986328125, 171.8649444580078, 31.32474136352539, 93.13692474365234, 43.838409423828125, -66.28842163085938, -31.985511779785156, 187.2976531982422, 138.64576721191406], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000557.npy"}
{"epoch": 0.8420256991685563, "step": 558, "batch_size": 64, "mean": 58.581756591796875, "std": 90.31764221191406, "min": -184.99752807617188, "p10": -23.58418045043945, "median": 45.737998962402344, "p90": 181.67862396240236, "max": 207.81488037109375, "pos_frac": 0.734375, "sample": [80.743408203125, 147.9317169189453, 193.50228881835938, -122.33317565917969, 163.21153259277344, 4.483388900756836, 187.17446899414062, 5.539617538452148, -1.7497291564941406, 37.148101806640625, 7.5797119140625, 146.5537109375, -13.609283447265625, 57.06488037109375, 4.264152526855469, 121.37539672851562, 133.84300231933594, 35.36274719238281, 159.27784729003906, 130.80575561523438, 69.14395141601562, 54.32789611816406, 14.565853118896484, 177.40211486816406, 141.6769256591797, -91.18890380859375, 123.90736389160156, -5.187232971191406, 147.4817352294922, 4.9225921630859375, 13.25515365600586, 12.089439392089844, 79.81318664550781, 207.81488037109375, -121.11039733886719, -11.549163818359375, 76.98316955566406, 4.626869201660156, 1.024871826171875, 205.01702880859375, 112.90863037109375, -102.3513412475586, 203.5941162109375, 21.049957275390625, 133.50485229492188, 183.51141357421875, -1.2022705078125, -6.0763092041015625, 199.31300354003906, 118.92073822021484, 89.4363021850586, -184.99752807617188, 159.48191833496094, -24.83106231689453, 17.746505737304688, -15.488365173339844, 104.21820831298828, -9.483253479003906, -32.921897888183594, -8.522842407226562, 152.42147827148438, 57.577449798583984, 18.910659790039062, -20.674789428710938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000558.npy"}
{"epoch": 0.8435374149659864, "step": 559, "batch_size": 64, "mean": 49.02497482299805, "std": 101.84522247314453, "min": -182.70904541015625, "p10": -51.78862838745116, "median": 27.06908416748047, "p90": 187.63218688964844, "max": 266.71453857421875, "pos_frac": 0.71875, "sample": [12.194290161132812, -150.7767333984375, 5.305484771728516, -1.7487564086914062, 85.00601196289062, 266.71453857421875, 121.14098358154297, 187.6678466796875, 0.9747467041015625, -1.157745361328125, -78.61698913574219, 86.23049926757812, -44.883995056152344, 41.39515686035156, 104.52690887451172, 187.54898071289062, 106.99887084960938, 28.596282958984375, 175.96302795410156, 0.174102783203125, 188.13308715820312, 63.983856201171875, 17.863792419433594, 174.06553649902344, 5.3709869384765625, -155.10183715820312, -43.687469482421875, 175.28787231445312, -54.74775695800781, 197.28768920898438, 7.181125640869141, -19.547935485839844, 7.1187744140625, 235.15805053710938, 42.82103729248047, -32.62492752075195, 64.05224609375, -4.436931610107422, 8.067848205566406, -11.244216918945312, 0.04740142822265625, -20.960540771484375, -21.162256240844727, 70.38914489746094, 72.1074447631836, 173.2796173095703, 150.49659729003906, -34.481651306152344, 166.51783752441406, 4.5339813232421875, 62.37248229980469, 204.09425354003906, -182.70904541015625, 25.541885375976562, 130.10464477539062, 192.15374755859375, 2.5360870361328125, 144.58135986328125, 8.213638305664062, 38.84678649902344, 75.41111755371094, 172.71083068847656, -133.73504638671875, -161.54617309570312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000559.npy"}
{"epoch": 0.8450491307634165, "step": 560, "batch_size": 64, "mean": 57.07308578491211, "std": 84.9419174194336, "min": -163.54534912109375, "p10": -23.872319793701163, "median": 33.29174995422363, "p90": 183.16239166259766, "max": 197.8255615234375, "pos_frac": 0.734375, "sample": [25.368865966796875, 101.87825012207031, 182.8250274658203, 113.80101776123047, -14.976730346679688, -3.397815704345703, 167.73483276367188, 4.33245849609375, 195.13079833984375, 1.3821945190429688, 25.928802490234375, 17.462295532226562, -13.095123291015625, 28.9805908203125, 3.1099853515625, -53.85035705566406, -46.15934753417969, 0.23288726806640625, 76.68231964111328, 154.25901794433594, -8.400505065917969, 176.09686279296875, 57.825721740722656, 2.1874866485595703, 96.57301330566406, 82.13579559326172, -27.684715270996094, 184.76327514648438, 183.30697631835938, 61.22187042236328, 16.897476196289062, -4.1966094970703125, -12.956008911132812, 196.28099060058594, 165.91986083984375, -6.9234771728515625, 6.467536926269531, -8.816703796386719, 193.4838104248047, -50.836334228515625, 197.8255615234375, 147.47775268554688, 9.9639892578125, 106.28314208984375, 56.269256591796875, 87.54424285888672, -5.032035827636719, 189.9506378173828, 141.85202026367188, 10.514244079589844, -47.33442687988281, 35.543701171875, -163.54534912109375, -11.856315612792969, 27.97528076171875, 102.89037322998047, -144.8786163330078, 42.036643981933594, 151.09490966796875, 42.402442932128906, 153.268798828125, 147.99984741210938, 31.039798736572266, 72.41532897949219], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000560.npy"}
{"epoch": 0.8465608465608465, "step": 561, "batch_size": 64, "mean": 56.26551818847656, "std": 81.80126190185547, "min": -153.14596557617188, "p10": -22.22985153198242, "median": 23.54886245727539, "p90": 181.9475540161133, "max": 224.17446899414062, "pos_frac": 0.75, "sample": [-3.4117279052734375, 46.51482009887695, 2.46575927734375, 9.960494995117188, 104.02902221679688, 178.6053466796875, 28.165077209472656, 2.6476211547851562, 169.5833282470703, 19.85516357421875, -11.251136779785156, 123.02436828613281, 1.5293769836425781, 36.90070343017578, 11.221263885498047, -3.2101593017578125, 119.8917236328125, 125.92901611328125, 197.4272918701172, 5.2343902587890625, 2.4610633850097656, 103.37130737304688, 126.98321533203125, 22.968650817871094, 94.5943603515625, -5.338451385498047, 6.870338439941406, -36.34925842285156, -153.14596557617188, 224.17446899414062, -38.3490104675293, 92.34886169433594, 85.83767700195312, 175.93310546875, 42.538246154785156, 129.78143310546875, 41.44324493408203, 42.92767333984375, 7.953548431396484, 202.6779327392578, 111.48519897460938, 200.9423370361328, 127.00534057617188, 194.42498779296875, 183.3799285888672, -14.751449584960938, -23.316177368164062, 152.52183532714844, 13.927032470703125, -0.30527496337890625, 7.165069580078125, -13.762359619140625, 5.531196594238281, -63.2376708984375, 131.363525390625, 141.69869995117188, -5.259391784667969, 5.206428527832031, 24.129074096679688, -35.04740905761719, -60.065834045410156, 17.13230323791504, 185.72662353515625, -19.695091247558594], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000561.npy"}
{"epoch": 0.8480725623582767, "step": 562, "batch_size": 64, "mean": 55.7551383972168, "std": 81.73174285888672, "min": -83.52582550048828, "p10": -19.398394393920896, "median": 25.45245361328125, "p90": 180.87720184326173, "max": 223.94320678710938, "pos_frac": 0.6875, "sample": [-7.806728363037109, 178.23782348632812, -66.64436340332031, 123.4046630859375, 36.669708251953125, 62.301551818847656, 8.197364807128906, 223.94320678710938, 185.51821899414062, 7.248786926269531, 115.15156555175781, 10.476791381835938, -83.52582550048828, -78.07304382324219, 21.3048095703125, -13.865901947021484, 76.7391357421875, 29.60009765625, 123.51040649414062, -25.258041381835938, 41.685546875, 190.67784118652344, -0.6425933837890625, -15.948699951171875, 21.065635681152344, 181.41896057128906, -7.741008758544922, -4.667144775390625, -6.4766082763671875, 16.483722686767578, 49.542762756347656, -52.89201354980469, 182.37939453125, 179.61309814453125, 129.6770782470703, 194.34506225585938, 1.9975051879882812, -10.923065185546875, 1.7288284301757812, 152.4362335205078, -5.036308288574219, 5.336307525634766, 9.947898864746094, 135.50233459472656, 50.132293701171875, 189.38815307617188, 119.58584594726562, 139.22412109375, 141.6885986328125, -1.0848846435546875, 174.7785186767578, 5.1483154296875, -20.876834869384766, 53.070281982421875, 36.15484619140625, 38.245567321777344, 155.15582275390625, 162.55307006835938, 101.01164245605469, -7.996700286865234, -2.05743408203125, -80.08685302734375, -6.411998748779297, 4.0656585693359375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000562.npy"}
{"epoch": 0.8495842781557067, "step": 563, "batch_size": 64, "mean": 57.53425216674805, "std": 98.86731719970703, "min": -160.34625244140625, "p10": -77.4218650817871, "median": 35.05681037902832, "p90": 193.22497100830077, "max": 244.1502685546875, "pos_frac": 0.734375, "sample": [46.105751037597656, -0.8669319152832031, 80.48159790039062, -0.5410099029541016, 194.46324157714844, 146.05056762695312, -103.29617309570312, 244.1502685546875, 188.32839965820312, 56.177268981933594, 6.242374420166016, 130.2924346923828, 3.0136489868164062, -72.24441528320312, 177.41705322265625, 60.10015106201172, -160.34625244140625, 198.2497100830078, -78.5101547241211, 203.68185424804688, -74.88252258300781, 204.146484375, -45.155731201171875, 169.73117065429688, 136.68023681640625, -52.692291259765625, 88.04326629638672, 80.44708251953125, 24.021865844726562, -52.842124938964844, -21.759689331054688, -16.513343811035156, 2.6717185974121094, 29.919044494628906, -1.2792911529541016, 192.75253295898438, 12.62091064453125, -107.5586929321289, 12.569355010986328, 24.35622787475586, 191.32986450195312, 196.80584716796875, 147.79342651367188, -98.35293579101562, -79.33071899414062, 66.61304473876953, 21.17780303955078, 16.674598693847656, 152.304931640625, 108.95576477050781, -79.52581787109375, 5.384494781494141, 40.194576263427734, 127.7280502319336, 162.28009033203125, 4.720149993896484, 193.4274444580078, 177.61300659179688, 13.084884643554688, 187.2450408935547, 61.79473876953125, 18.052078247070312, 96.50299072265625, 25.493186950683594], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000563.npy"}
{"epoch": 0.8510959939531368, "step": 564, "batch_size": 64, "mean": 54.58319854736328, "std": 99.52955627441406, "min": -186.68673706054688, "p10": -69.06642456054688, "median": 39.11273765563965, "p90": 182.27588500976563, "max": 205.32334899902344, "pos_frac": 0.78125, "sample": [173.0670166015625, 12.682411193847656, 191.65487670898438, 29.610504150390625, 7.318622589111328, 182.77650451660156, 130.86326599121094, -186.68673706054688, 20.642208099365234, -10.567596435546875, 195.43641662597656, 166.3927459716797, 49.83094787597656, -99.52656555175781, 5.811820983886719, 45.815834045410156, -65.10023498535156, 121.083251953125, 191.4247589111328, 23.627090454101562, -12.226692199707031, 11.348331451416016, 157.48916625976562, 57.491058349609375, 89.89332580566406, 63.025291442871094, 205.32334899902344, 33.06028747558594, 0.6963996887207031, 98.77284240722656, 134.03311157226562, 132.17935180664062, 196.236572265625, 14.363700866699219, -2.0286407470703125, -162.00222778320312, 45.16518783569336, 12.469001770019531, 10.290969848632812, -12.728172302246094, 142.92544555664062, 156.90675354003906, 7.2689056396484375, -70.76622009277344, -101.2854232788086, 7.70806884765625, 92.08476257324219, 120.1348876953125, -58.38848876953125, -169.58004760742188, 179.28729248046875, 164.65328979492188, 50.602752685546875, -20.968046188354492, 173.52325439453125, 10.84893798828125, 165.03407287597656, 191.0098876953125, -148.4666290283203, 181.10777282714844, 6.3390350341796875, 29.07846450805664, 24.705604553222656, 100.55105590820312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000564.npy"}
{"epoch": 0.8526077097505669, "step": 565, "batch_size": 64, "mean": 51.160430908203125, "std": 83.25272369384766, "min": -150.36927795410156, "p10": -25.764447021484372, "median": 29.17276382446289, "p90": 173.42564544677734, "max": 194.5037841796875, "pos_frac": 0.75, "sample": [13.888628005981445, -150.36927795410156, -57.193511962890625, 38.596229553222656, 18.182022094726562, 4.24908447265625, 31.904541015625, 8.5404052734375, -2.7387657165527344, 12.365097045898438, 21.35303497314453, 149.91163635253906, 177.64895629882812, -130.23646545410156, 10.985565185546875, 171.9629364013672, 61.85862350463867, 102.64384460449219, 194.23556518554688, 191.51119995117188, 174.05252075195312, -36.957916259765625, -26.463790893554688, 10.715484619140625, 68.01219177246094, 43.01414489746094, -2.4998855590820312, 188.24427795410156, 104.99232482910156, 86.17012023925781, 151.16732788085938, 2.2025794982910156, 4.525848388671875, -18.45751953125, 68.37137603759766, -16.101089477539062, 106.1318359375, 69.06707000732422, -119.00320434570312, 169.82858276367188, 63.55694580078125, 20.2403564453125, -77.30262756347656, 53.10723114013672, 8.678558349609375, 26.44098663330078, -8.402290344238281, 142.46694946289062, 194.5037841796875, 21.8870849609375, 78.82866668701172, -24.132644653320312, -16.22491455078125, -3.6664581298828125, 150.14871215820312, 16.99266815185547, 36.17765808105469, 137.1592559814453, -11.323921203613281, 1.1020736694335938, 193.82923889160156, 138.40115356445312, 69.33467102050781, 166.1529541015625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000565.npy"}
{"epoch": 0.854119425547997, "step": 566, "batch_size": 64, "mean": 53.09088897705078, "std": 92.97819519042969, "min": -184.88832092285156, "p10": -61.34919509887694, "median": 54.513065338134766, "p90": 181.76873779296878, "max": 214.54660034179688, "pos_frac": 0.703125, "sample": [-3.1228179931640625, -45.21600341796875, -41.90982437133789, 49.09033203125, 59.30580139160156, 183.45010375976562, 186.7785186767578, 139.4986114501953, 88.1769790649414, 191.2378387451172, 147.38922119140625, 53.96153259277344, 75.43809509277344, 19.384483337402344, 13.885955810546875, 87.58451843261719, -184.88832092285156, 88.81442260742188, 55.59650421142578, 51.91944885253906, 122.71490478515625, 3.6539649963378906, -67.79190063476562, 97.78639221191406, 27.13611602783203, 11.3533935546875, 159.96026611328125, 165.05960083007812, 146.25680541992188, -30.927474975585938, 92.26455688476562, 128.85055541992188, 177.84555053710938, -19.361160278320312, -74.61051940917969, -164.23890686035156, 40.96424102783203, 50.598289489746094, 188.7764892578125, -29.6187744140625, 91.61865234375, -46.31621551513672, 214.54660034179688, -41.441226959228516, 189.74095153808594, 19.422199249267578, -18.764739990234375, -116.62696838378906, 74.98524475097656, -0.2661552429199219, 1.4233207702636719, -6.461082458496094, -93.11083984375, 167.78550720214844, -24.561866760253906, 74.20538330078125, 55.064598083496094, 169.33761596679688, 93.39559936523438, -75.34028625488281, 143.49008178710938, 83.369873046875, 2.80010986328125, 196.47286987304688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000566.npy"}
{"epoch": 0.8556311413454271, "step": 567, "batch_size": 64, "mean": 63.92971420288086, "std": 85.13372802734375, "min": -82.48865509033203, "p10": -54.55481529235839, "median": 52.40284538269043, "p90": 181.94881591796874, "max": 223.30142211914062, "pos_frac": 0.75, "sample": [142.3973388671875, -1.7074737548828125, -80.61172485351562, 158.10330200195312, -10.111892700195312, 70.51017761230469, 186.7162322998047, 141.80758666992188, 189.83859252929688, 60.619850158691406, 88.15411376953125, 37.162071228027344, 121.453857421875, -14.833580017089844, 94.9279556274414, -1.7461395263671875, -77.84944915771484, 25.110000610351562, 0.48392486572265625, 146.14175415039062, 28.161468505859375, 0.5358657836914062, -3.3135147094726562, 178.64053344726562, -29.582427978515625, -0.09985733032226562, -75.14105224609375, -57.045101165771484, 38.410736083984375, -48.74414825439453, 129.13711547851562, 98.50090026855469, 192.75765991210938, 223.30142211914062, 151.81231689453125, 33.05192565917969, 90.17052459716797, 28.792465209960938, 182.34353637695312, 208.85345458984375, -69.77677917480469, 90.15997314453125, 19.629417419433594, 158.01220703125, 181.02780151367188, 170.5732421875, 1.8166580200195312, 14.800582885742188, 83.80694580078125, 146.42538452148438, 80.44499206542969, 3.904102325439453, 168.93569946289062, 8.684921264648438, -82.48865509033203, 19.293678283691406, -65.77781677246094, 36.376461029052734, 193.493408203125, -19.804595947265625, 137.63555908203125, 62.412574768066406, 52.545433044433594, 52.260257720947266], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000567.npy"}
{"epoch": 0.8571428571428571, "step": 568, "batch_size": 64, "mean": 62.09715270996094, "std": 86.40604400634766, "min": -94.7709732055664, "p10": -49.98521080017088, "median": 40.4820613861084, "p90": 187.9888107299805, "max": 211.720947265625, "pos_frac": 0.78125, "sample": [66.19686126708984, 42.02950668334961, 14.424697875976562, 50.1301155090332, -14.928882598876953, 66.18111419677734, 65.54693603515625, -69.06825256347656, -1.41326904296875, -55.638023376464844, -36.79531478881836, 194.0514373779297, 32.95783233642578, 175.21774291992188, -33.574851989746094, -62.43876647949219, 29.381912231445312, 1.0636138916015625, 1.7236709594726562, 18.731117248535156, 15.315849304199219, 49.693389892578125, 63.12079620361328, 143.97882080078125, 173.6668243408203, 184.6578826904297, 159.1597137451172, 209.13534545898438, -94.7709732055664, 29.41761016845703, 113.60771179199219, 80.82571411132812, 16.34185791015625, -9.393112182617188, 66.52581787109375, 21.801998138427734, -67.23522186279297, 18.39051055908203, 0.7560081481933594, 111.40998077392578, 2.6884002685546875, 196.24191284179688, 86.55010986328125, 109.04650115966797, 183.06149291992188, 5.47515869140625, 174.16165161132812, 2.6208572387695312, -67.23910522460938, 0.39881134033203125, 117.09573364257812, 147.59246826171875, 38.93461608886719, 115.67926025390625, 198.42971801757812, 13.858657836914062, -1.0050125122070312, 211.720947265625, 172.57745361328125, 183.2181396484375, -66.00820922851562, -20.86992645263672, 189.41635131835938, 210.38592529296875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000568.npy"}
{"epoch": 0.8586545729402872, "step": 569, "batch_size": 64, "mean": 41.79746627807617, "std": 93.21637725830078, "min": -184.81549072265625, "p10": -82.18217315673827, "median": 35.09159851074219, "p90": 179.28986053466798, "max": 197.14840698242188, "pos_frac": 0.75, "sample": [170.83465576171875, -32.20758056640625, 0.7495288848876953, 169.66592407226562, 85.8582763671875, 136.07345581054688, -73.0431137084961, 52.8663330078125, 22.0537109375, 186.78977966308594, 176.31289672851562, 24.278907775878906, 89.97713470458984, 72.03665161132812, -86.48257446289062, 48.51848602294922, 5.020702362060547, 193.6144561767578, 7.5350341796875, 93.92412567138672, 179.07229614257812, 36.955596923828125, 44.118927001953125, 197.14840698242188, 76.36952209472656, 96.57539367675781, 139.498046875, 38.305965423583984, 13.80364990234375, -88.641845703125, -178.7200164794922, 23.198659896850586, -85.05430603027344, 3.6946029663085938, -24.108177185058594, 58.61906433105469, 185.31845092773438, 49.124725341796875, 28.345321655273438, 95.39469909667969, 64.93232727050781, -35.27842712402344, -180.51516723632812, -48.095184326171875, 71.47240447998047, -184.81549072265625, 33.22760009765625, -75.48052978515625, 31.53673553466797, 12.072830200195312, 95.3481216430664, 18.81201171875, 179.3831024169922, -93.44097900390625, 12.92221450805664, 98.1043701171875, 106.6566162109375, 31.33929443359375, 182.6090087890625, -33.049171447753906, -14.44580078125, 5.123573303222656, 185.26344299316406, -22.0408935546875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000569.npy"}
{"epoch": 0.8601662887377173, "step": 570, "batch_size": 64, "mean": 67.33297729492188, "std": 88.75947570800781, "min": -111.14283752441406, "p10": -16.41667938232422, "median": 50.97205352783203, "p90": 186.4275100708008, "max": 264.3060607910156, "pos_frac": 0.8125, "sample": [138.5244903564453, -110.2993392944336, 121.83880615234375, 2.85308837890625, 26.699108123779297, 86.51970672607422, -72.51961517333984, -2.525806427001953, 75.23310852050781, -16.753250122070312, 16.82257843017578, 3.6405677795410156, 11.917945861816406, 151.74505615234375, 31.33755111694336, 165.02276611328125, 0.008960723876953125, -3.581329345703125, -100.90415954589844, 184.73797607421875, 40.17591857910156, -65.04942321777344, 4.326225280761719, 185.36859130859375, 194.93569946289062, -0.6523895263671875, 71.59410858154297, 120.92384338378906, 169.4794921875, 82.91328430175781, 31.07666778564453, 172.63343811035156, 172.36862182617188, -15.63134765625, 6.9631805419921875, 66.96843719482422, 123.28922271728516, 57.09526062011719, 190.65237426757812, 4.039056777954102, 189.9318389892578, 109.76197814941406, 179.41773986816406, 0.0946807861328125, -9.405609130859375, 44.848846435546875, 109.19989013671875, 129.46441650390625, 185.36209106445312, -46.012939453125, 187.19248962402344, -111.14283752441406, 23.69464111328125, 157.31512451171875, 199.05650329589844, 3.5016250610351562, 3.316671371459961, 2.1876869201660156, 61.41908264160156, 102.42705535888672, 264.3060607910156, 1.746368408203125, 186.88133239746094, 10.957550048828125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000570.npy"}
{"epoch": 0.8616780045351474, "step": 571, "batch_size": 64, "mean": 38.9942626953125, "std": 92.57412719726562, "min": -190.0146484375, "p10": -94.87741394042969, "median": 32.29023361206055, "p90": 155.80313110351562, "max": 186.84884643554688, "pos_frac": 0.6875, "sample": [19.85150909423828, 103.69009399414062, 105.03271484375, 33.633514404296875, -14.905342102050781, 137.03749084472656, 26.6092529296875, 128.01467895507812, -0.1178436279296875, 101.69784545898438, 3.5490493774414062, 2.3303451538085938, -52.33140563964844, 31.613929748535156, -2.2895851135253906, 63.129791259765625, 92.56842041015625, 143.1716766357422, -1.7039337158203125, 117.43917846679688, 177.9456329345703, 2.0495376586914062, 21.849945068359375, 104.08308410644531, 151.74729919433594, 90.11872863769531, 153.40670776367188, 48.86456298828125, -190.0146484375, 15.58050537109375, -3.2037696838378906, -144.58872985839844, 33.932044982910156, -3.30267333984375, 20.766311645507812, -3.089092254638672, -90.79330444335938, -1.5215511322021484, -150.4616241455078, -96.62774658203125, 56.06649398803711, 11.15771484375, 32.96653747558594, 76.24793243408203, 88.18572998046875, 186.02987670898438, 184.90179443359375, -129.78701782226562, 88.16828918457031, 4.10760498046875, -12.695587158203125, -174.79226684570312, 44.48283386230469, 129.58851623535156, 163.65972900390625, -31.080154418945312, 176.79843139648438, 156.83016967773438, 148.74830627441406, 89.93885803222656, -52.96638488769531, 14.467954635620117, -117.00420379638672, 186.84884643554688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000571.npy"}
{"epoch": 0.8631897203325775, "step": 572, "batch_size": 64, "mean": 49.476531982421875, "std": 74.42156219482422, "min": -148.20774841308594, "p10": -27.970028686523428, "median": 30.326648712158203, "p90": 174.25189056396493, "max": 196.36097717285156, "pos_frac": 0.75, "sample": [141.32583618164062, -41.18273162841797, 65.92538452148438, 117.60460662841797, 85.3472671508789, 60.486690521240234, 138.63035583496094, 84.47718811035156, -52.31147003173828, 196.36097717285156, 1.7783737182617188, 64.85487365722656, 61.3199462890625, 192.1105194091797, -10.208137512207031, 96.59905242919922, -53.52185821533203, 13.697717666625977, -18.032302856445312, 183.4864501953125, 20.947444915771484, 110.11422729492188, -4.1496124267578125, 184.03456115722656, 95.16227722167969, 52.037864685058594, 0.5024337768554688, 153.54383850097656, -0.10082244873046875, 3.3399658203125, -49.95209503173828, 183.12677001953125, 186.10585021972656, -31.8076171875, 44.62394714355469, 36.850746154785156, 7.292549133300781, -35.325469970703125, 2.3912315368652344, 80.49113464355469, 105.42282104492188, 3.5436859130859375, 79.85166931152344, 132.7255859375, 42.59203338623047, 11.257049560546875, -16.235397338867188, -19.015655517578125, 195.65939331054688, 56.364524841308594, 23.80255126953125, 152.0800018310547, -11.89971923828125, 57.6890869140625, -148.20774841308594, -11.069473266601562, 6.782432556152344, 17.34619903564453, -4.427955627441406, 6.603424072265625, 21.935379028320312, 0.9601974487304688, 8.977596282958984, 85.7823486328125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000572.npy"}
{"epoch": 0.8647014361300076, "step": 573, "batch_size": 64, "mean": 46.549224853515625, "std": 101.63924407958984, "min": -187.44215393066406, "p10": -92.74419174194335, "median": 21.489608764648438, "p90": 193.92747955322267, "max": 252.483642578125, "pos_frac": 0.671875, "sample": [39.15521240234375, 115.54745483398438, -187.44215393066406, 141.38095092773438, -84.78304290771484, 10.22412109375, -8.345829010009766, 34.79147720336914, 15.578720092773438, -105.26187133789062, -169.77195739746094, -105.39186096191406, 12.240493774414062, -2.1266021728515625, 6.9041748046875, 134.5213623046875, -0.8120574951171875, 157.84930419921875, 194.11703491210938, -163.87994384765625, 170.41888427734375, 91.65494537353516, 189.30551147460938, 53.492347717285156, 9.00006103515625, 8.797183990478516, 252.483642578125, 10.167938232421875, 181.1389617919922, -4.7574462890625, 151.0634307861328, 21.910064697265625, 30.44123077392578, -22.330764770507812, 21.06915283203125, 31.89954376220703, 40.936912536621094, 11.297683715820312, 67.49852752685547, 210.59205627441406, 184.3770294189453, 106.61821746826172, 95.38155364990234, -4.603605270385742, -3.5396728515625, 36.114227294921875, 199.04440307617188, -0.53900146484375, 6.742744445800781, 170.6824493408203, 195.91720581054688, 193.4851837158203, -6.78143310546875, 85.4240951538086, -15.019500732421875, -96.15611267089844, 195.951904296875, -99.86383056640625, 35.642704010009766, -2.185178756713867, 6.7255706787109375, -26.212615966796875, -45.28666687011719, 206.656005859375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000573.npy"}
{"epoch": 0.8662131519274376, "step": 574, "batch_size": 64, "mean": 63.921165466308594, "std": 84.43812561035156, "min": -114.50386810302734, "p10": -36.355707931518545, "median": 53.3958740234375, "p90": 185.0748031616211, "max": 200.26055908203125, "pos_frac": 0.75, "sample": [55.49161148071289, -28.879966735839844, -4.386383056640625, -65.66911315917969, -39.94165802001953, 178.7633056640625, 4.167591094970703, -83.69181823730469, -4.506919860839844, -89.31598663330078, 23.946025848388672, 19.041885375976562, -5.8367462158203125, -7.042724609375, -39.55959701538086, 1.5868110656738281, -12.940214157104492, 54.26148986816406, 82.01373291015625, 105.750732421875, 144.95945739746094, 5.873908996582031, -2.168304443359375, 140.6089324951172, 2.3980484008789062, 183.47125244140625, 185.7620391845703, 192.00115966796875, 5.074445724487305, 180.47390747070312, 16.2432861328125, 7.398075103759766, 131.27139282226562, 32.94511413574219, -56.889434814453125, 140.34689331054688, -114.50386810302734, 133.85548400878906, 188.35675048828125, 70.15715789794922, 156.1236572265625, 9.598190307617188, 197.84078979492188, 51.457672119140625, 181.95684814453125, 181.1864776611328, 119.76016235351562, 39.34849548339844, 59.396240234375, 8.375682830810547, -7.4565277099609375, 200.26055908203125, 52.53025817871094, 106.04161834716797, 141.60491943359375, 30.854965209960938, 65.43083190917969, 125.113525390625, 187.82650756835938, 189.62045288085938, -9.916229248046875, 132.84397888183594, 82.80810546875, 57.459259033203125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000574.npy"}
{"epoch": 0.8677248677248677, "step": 575, "batch_size": 64, "mean": 43.40424346923828, "std": 93.58380889892578, "min": -187.80638122558594, "p10": -42.53561935424805, "median": 10.840705871582031, "p90": 193.24349670410157, "max": 240.44149780273438, "pos_frac": 0.625, "sample": [152.30801391601562, -100.52912139892578, -146.004150390625, 69.87638854980469, 240.44149780273438, 11.8983154296875, -12.138999938964844, 108.83270263671875, -2.2649383544921875, -14.022880554199219, 103.52948760986328, -17.60211181640625, 202.72695922851562, -3.28753662109375, -131.75115966796875, 11.57037353515625, 5.081001281738281, 72.71383666992188, 44.71353530883789, 2.0794410705566406, 7.895164489746094, 161.41928100585938, 126.82080841064453, 72.12014770507812, 61.844207763671875, -6.5342864990234375, -52.273109436035156, 57.107398986816406, 200.23577880859375, 69.90705871582031, 60.40199279785156, -43.04426574707031, 162.3600616455078, 209.36392211914062, 162.1063232421875, -41.348777770996094, 195.74569702148438, -3.7725830078125, 197.32275390625, 140.50241088867188, -24.73785400390625, 3.4608078002929688, -2.17425537109375, 3.5826873779296875, 6.54779052734375, -7.6714324951171875, -7.323570251464844, -187.80638122558594, 43.45149230957031, 219.37466430664062, 187.405029296875, 10.111038208007812, -13.99493408203125, 106.3851547241211, -82.34211730957031, 11.802864074707031, 4.911867141723633, -16.8359375, -4.665008544921875, 12.2227783203125, -13.097312927246094, 76.93146514892578, 154.20010375976562, -38.21821212768555], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000575.npy"}
{"epoch": 0.8692365835222978, "step": 576, "batch_size": 64, "mean": 65.78793334960938, "std": 88.57901000976562, "min": -113.58467102050781, "p10": -27.900796508789057, "median": 51.8690071105957, "p90": 184.69744720458985, "max": 204.51515197753906, "pos_frac": 0.78125, "sample": [49.658966064453125, 183.26834106445312, -11.756126403808594, 85.73594665527344, 66.99099731445312, 183.05813598632812, 192.87014770507812, 1.3515167236328125, 29.471878051757812, 92.82438659667969, -110.42974853515625, -13.386917114257812, 147.14041137695312, -109.57234954833984, 133.16647338867188, -23.351768493652344, -86.88158416748047, 68.37955474853516, 181.55430603027344, 6.24169921875, 172.54147338867188, 193.11683654785156, 54.07904815673828, 6.625787734985352, 144.8197479248047, 150.60231018066406, 126.32611083984375, -7.802001953125, 183.02157592773438, 2.6649551391601562, 12.525680541992188, 5.108978271484375, 36.411346435546875, 176.55819702148438, 77.2785873413086, -30.977500915527344, 204.51515197753906, -3.9818801879882812, 4.814727783203125, 25.597270965576172, 67.0177001953125, -61.790252685546875, 142.2481689453125, 3.8211708068847656, 180.6022186279297, 5.039825439453125, 192.0248260498047, 139.00721740722656, -1.0137042999267578, 185.30992126464844, 190.72018432617188, -113.58467102050781, 126.83139038085938, 10.59527587890625, 30.474014282226562, 162.47792053222656, 58.08549118041992, -1.9906692504882812, 194.61325073242188, 94.49994659423828, 8.036056518554688, -29.850379943847656, 15.273605346679688, 11.798480987548828], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000576.npy"}
{"epoch": 0.8707482993197279, "step": 577, "batch_size": 64, "mean": 36.76244354248047, "std": 85.57746124267578, "min": -191.1164093017578, "p10": -44.45385665893552, "median": 20.215428352355957, "p90": 164.39143066406257, "max": 193.7468719482422, "pos_frac": 0.671875, "sample": [66.70075988769531, 33.09858703613281, -144.42022705078125, 7.906394958496094, -16.97112274169922, -21.101882934570312, 3.3030242919921875, -10.05251693725586, 120.44941711425781, 94.94532775878906, -0.6129074096679688, 191.27334594726562, 96.75389862060547, -6.040184020996094, -109.35806274414062, 188.2449951171875, 130.25125122070312, 188.71165466308594, -152.6527862548828, 142.1481475830078, 119.11363983154297, 59.17887878417969, 1.2011051177978516, 170.4129180908203, 7.746036529541016, -54.46184539794922, -66.968505859375, 148.45706176757812, 17.63956642150879, 30.29638671875, -1.0060997009277344, 49.35575485229492, -11.234848022460938, 46.35218811035156, -5.443878173828125, 24.433631896972656, 22.791290283203125, 46.40415954589844, -3.448383331298828, -2.252826690673828, 0.5007896423339844, -1.1318740844726562, 121.08650207519531, -0.18608856201171875, 193.7468719482422, 150.34129333496094, -141.95797729492188, 31.662635803222656, 30.8458251953125, 107.3353271484375, 4.62933349609375, 50.460044860839844, 173.52252197265625, 96.85476684570312, 15.373954772949219, 33.1883544921875, -191.1164093017578, -9.458629608154297, -6.3145599365234375, 95.6452407836914, 5.532978057861328, 177.7550811767578, 9.396697998046875, 3.9405364990234375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000577.npy"}
{"epoch": 0.872260015117158, "step": 578, "batch_size": 64, "mean": 36.504695892333984, "std": 79.34550476074219, "min": -111.54853057861328, "p10": -49.53500289916992, "median": 7.685881614685059, "p90": 172.90128173828126, "max": 208.00152587890625, "pos_frac": 0.625, "sample": [-8.199520111083984, 0.26567840576171875, 12.154670715332031, -7.224449157714844, 10.756759643554688, -59.513572692871094, 11.143531799316406, 27.494338989257812, 143.02639770507812, -37.22380065917969, 5.29876708984375, 48.4151611328125, -11.791976928710938, 111.19892883300781, 135.104736328125, -74.21969604492188, 161.58631896972656, 170.3758087158203, 117.00630187988281, 51.168975830078125, 149.63870239257812, 38.71660614013672, 7.585786819458008, 5.0787506103515625, -0.9690074920654297, 39.987754821777344, 173.98362731933594, -1.9005508422851562, -5.487689971923828, -7.118003845214844, 208.00152587890625, 7.785976409912109, 177.21253967285156, 1.942138671875, -50.040985107421875, 2.6497802734375, -0.8142318725585938, 24.070663452148438, 81.73882293701172, 107.67568969726562, 183.8421173095703, -23.03204345703125, 31.534263610839844, -31.898847579956055, 10.80816650390625, -3.9260292053222656, -17.47991943359375, 178.2698211669922, 69.0640869140625, 157.67816162109375, 13.88165283203125, -62.23987579345703, 202.3868865966797, -13.386837005615234, -10.1602783203125, -69.57872009277344, 5.8967437744140625, -50.89521026611328, 0.64337158203125, -31.636003494262695, -111.54853057861328, 9.271331787109375, -48.35437774658203, 180.59933471679688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000578.npy"}
{"epoch": 0.873771730914588, "step": 579, "batch_size": 64, "mean": 57.39547348022461, "std": 79.9232406616211, "min": -174.68569946289062, "p10": -20.885842704772944, "median": 45.545921325683594, "p90": 166.93317108154298, "max": 197.87677001953125, "pos_frac": 0.796875, "sample": [16.0489501953125, 95.29135131835938, 108.09056091308594, 45.696800231933594, 96.7189712524414, -174.68569946289062, 168.75827026367188, 11.02587890625, 92.69401550292969, 130.31207275390625, 111.26979064941406, 123.04817199707031, 133.0816650390625, 195.89398193359375, 45.77247619628906, 18.562713623046875, 157.3023681640625, 120.3228530883789, -48.48418426513672, 86.59072875976562, 197.87677001953125, 11.26474380493164, 113.95901489257812, -2.101085662841797, -2.2595367431640625, -6.6468658447265625, 193.40306091308594, 43.58845520019531, 175.53372192382812, 80.20945739746094, -45.86143493652344, 0.32837867736816406, 35.23527526855469, 42.652591705322266, 70.5572280883789, 162.6746063232422, 84.5766372680664, 150.04489135742188, -24.675201416015625, 178.21957397460938, -5.029056549072266, 3.799774169921875, -23.273218154907227, -43.267478942871094, 38.81036376953125, 3.4533004760742188, -15.315299987792969, -172.4534454345703, 195.08018493652344, 1.7936058044433594, 8.129653930664062, 94.52937316894531, 7.982879638671875, 133.8212432861328, 15.209312438964844, 9.204706192016602, 113.64254760742188, 122.14114379882812, 4.551292419433594, 82.74137115478516, -14.871353149414062, 18.393524169921875, 56.94886779785156, 45.395042419433594], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000579.npy"}
{"epoch": 0.8752834467120182, "step": 580, "batch_size": 64, "mean": 52.784088134765625, "std": 90.39906311035156, "min": -180.5576934814453, "p10": -31.623375701904287, "median": 40.265092849731445, "p90": 179.76807250976563, "max": 211.47116088867188, "pos_frac": 0.75, "sample": [-20.28900146484375, -19.729507446289062, -36.48096466064453, 34.482173919677734, -7.5660858154296875, 123.55741882324219, 162.612548828125, 45.581695556640625, 40.31755065917969, 15.819820404052734, 182.36422729492188, 40.2126350402832, 22.168670654296875, 11.862396240234375, -44.10490417480469, 49.15400695800781, 139.25303649902344, 106.35350036621094, -5.438999176025391, 21.230247497558594, -0.034015655517578125, 31.717239379882812, 197.30911254882812, 69.24420928955078, 98.11257934570312, -8.075149536132812, 211.23484802246094, 5.299140930175781, -4.195079803466797, 198.3291473388672, 177.34225463867188, 119.98738098144531, 4.302055358886719, 79.47671508789062, -97.70673370361328, 180.80770874023438, -173.45269775390625, -163.49026489257812, 5.229743957519531, 139.28610229492188, -180.5576934814453, 32.753326416015625, -16.282981872558594, 67.89727783203125, 211.47116088867188, 23.455814361572266, -2.8557205200195312, 82.58723449707031, 201.7352294921875, 98.70454406738281, 174.6367645263672, 157.03390502929688, 2.1220245361328125, -76.55929565429688, 42.846187591552734, 9.413656234741211, 131.80288696289062, 136.3662109375, 107.43621826171875, 52.375030517578125, 126.0947036743164, 16.700775146484375, 6.492748260498047, 40.4267578125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000580.npy"}
{"epoch": 0.8767951625094482, "step": 581, "batch_size": 64, "mean": 45.93260955810547, "std": 88.67831420898438, "min": -154.81768798828125, "p10": -59.08143806457519, "median": 45.105709075927734, "p90": 174.79131164550782, "max": 203.60533142089844, "pos_frac": 0.734375, "sample": [143.00401306152344, 88.87686920166016, 71.61247253417969, 23.671531677246094, -70.35769653320312, 58.396820068359375, 45.08967590332031, -24.080642700195312, 190.04763793945312, 131.42489624023438, -53.98194885253906, 9.065120697021484, 25.169830322265625, 68.46363067626953, 4.2238006591796875, -32.70631790161133, 156.2816162109375, 35.908966064453125, 172.40835571289062, 31.03602409362793, 173.18527221679688, 99.52235412597656, 175.4796142578125, -148.69424438476562, 11.868579864501953, 192.07974243164062, 188.234619140625, -43.336997985839844, 45.121742248535156, 5.747535705566406, -83.90095520019531, 68.27861022949219, 27.568439483642578, 46.70320510864258, 55.73405456542969, 91.21531677246094, -23.858383178710938, -154.81768798828125, 203.60533142089844, 169.71853637695312, 138.07095336914062, -38.666900634765625, 63.029632568359375, 57.3338623046875, 37.648231506347656, -61.26693344116211, 40.4244384765625, 67.75309753417969, -46.396400451660156, -145.99563598632812, 177.5716552734375, 85.98971557617188, 181.8068084716797, 23.38977813720703, -0.27680397033691406, 113.04878234863281, -11.203495025634766, -143.275634765625, -20.76507568359375, 66.549072265625, 75.38256072998047, 23.492053985595703, 62.01249313354492, 21.021526336669922], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000581.npy"}
{"epoch": 0.8783068783068783, "step": 582, "batch_size": 64, "mean": 50.751991271972656, "std": 92.29076385498047, "min": -176.6158447265625, "p10": -74.36691055297851, "median": 39.88903999328613, "p90": 173.16316680908207, "max": 209.67547607421875, "pos_frac": 0.71875, "sample": [150.1544189453125, 199.672119140625, 37.49747848510742, 119.84275817871094, 19.816078186035156, -8.518325805664062, 190.0389862060547, -2.6795654296875, 134.17990112304688, 131.84825134277344, 51.89582061767578, 159.95408630371094, 184.4780731201172, 209.67547607421875, 0.6225357055664062, 179.91165161132812, 3.931121826171875, -97.1068115234375, 124.58601379394531, 88.10723114013672, -0.36612701416015625, 42.280601501464844, 8.707462310791016, 144.498046875, 49.791175842285156, 112.01229858398438, 3.594371795654297, -4.112888336181641, -157.6941375732422, -133.84727478027344, -0.46570587158203125, 45.785675048828125, 145.6063232421875, -108.25208282470703, -71.13088989257812, 113.44024658203125, 140.881103515625, 119.14533233642578, 183.46597290039062, -18.451629638671875, 104.86245727539062, 177.65176391601562, 30.650630950927734, 48.92720413208008, -75.75377655029297, 4.4777984619140625, 1.235687255859375, 158.37808227539062, 7.367668151855469, -176.6158447265625, 0.38667869567871094, 18.61742401123047, -2.391630172729492, -0.32492828369140625, 11.858665466308594, 98.02659606933594, 143.12503051757812, -79.16569519042969, 74.61734008789062, -51.032691955566406, 14.411026000976562, -1.4485054016113281, 84.78160858154297, 162.6897735595703], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000582.npy"}
{"epoch": 0.8798185941043084, "step": 583, "batch_size": 64, "mean": 44.44509506225586, "std": 98.64110565185547, "min": -160.2379150390625, "p10": -90.39423522949217, "median": 31.240970611572266, "p90": 185.87639770507815, "max": 198.49359130859375, "pos_frac": 0.625, "sample": [-0.70074462890625, -49.48704528808594, -160.08273315429688, -1.7450485229492188, 198.49359130859375, 110.15824127197266, -160.2379150390625, -11.011665344238281, -147.23764038085938, -75.09964752197266, 94.60850524902344, -96.94905853271484, 30.949832916259766, 150.55836486816406, 92.84577941894531, -10.683357238769531, 187.97610473632812, -36.25137710571289, 190.47869873046875, 94.43489074707031, 174.41665649414062, -114.04048156738281, 31.155120849609375, -3.0416488647460938, 69.03337097167969, 97.39157104492188, 90.03646850585938, -107.37867736816406, 35.99448013305664, 186.90914916992188, -0.0516357421875, -3.6439666748046875, 19.551498413085938, 17.84149169921875, 40.63232421875, 193.231201171875, -62.199867248535156, -67.85427856445312, 31.326820373535156, 188.49363708496094, 13.920448303222656, 168.64520263671875, -152.762451171875, 174.79824829101562, -42.31882858276367, 165.297119140625, 109.91960906982422, 60.36424255371094, 75.9295883178711, -26.919801712036133, 154.78781127929688, 122.84143829345703, 75.2723388671875, -0.36290740966796875, 134.46014404296875, 23.929275512695312, 17.313499450683594, 65.78755187988281, -0.47099876403808594, 108.32113647460938, 193.38525390625, 183.46664428710938, -1.4080810546875, 1.4686393737792969], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000583.npy"}
{"epoch": 0.8813303099017384, "step": 584, "batch_size": 64, "mean": 49.43035888671875, "std": 96.17089080810547, "min": -158.70730590820312, "p10": -81.54541549682615, "median": 31.95223045349121, "p90": 187.24058990478517, "max": 203.62673950195312, "pos_frac": 0.65625, "sample": [196.55364990234375, 106.41427612304688, 78.87422180175781, -4.464807510375977, 80.81036376953125, -87.91812896728516, 191.6345672607422, 188.54713439941406, 164.34886169433594, 109.2745361328125, 13.677085876464844, -6.575037002563477, -16.472686767578125, -109.56865692138672, -93.65110778808594, 12.995872497558594, 1.4065704345703125, 189.37124633789062, 167.43820190429688, -60.730125427246094, 178.13308715820312, 184.19198608398438, -21.04339599609375, 133.73483276367188, 103.04556274414062, 28.961181640625, 50.119815826416016, 34.67597198486328, 0.5319061279296875, 15.158416748046875, 157.6247100830078, -119.34339904785156, 48.44194793701172, -10.139617919921875, -135.55270385742188, 195.80322265625, -8.11724853515625, -0.4119873046875, -31.34142303466797, -9.319908142089844, 132.45718383789062, 60.58985137939453, -66.67575073242188, 76.60015106201172, 63.130393981933594, 166.70037841796875, 29.22848892211914, 157.92010498046875, 87.49832916259766, -16.09539794921875, 97.390380859375, 23.5823974609375, -3.680614471435547, 170.41091918945312, 0.7896308898925781, 203.62673950195312, 200.59539794921875, 37.260948181152344, 20.028778076171875, -1.9675216674804688, -4.2598114013671875, 114.52017974853516, -144.51986694335938, -158.70730590820312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000584.npy"}
{"epoch": 0.8828420256991686, "step": 585, "batch_size": 64, "mean": 34.23884201049805, "std": 84.44027709960938, "min": -176.63015747070312, "p10": -66.02936630249019, "median": 20.95735740661621, "p90": 162.07280426025395, "max": 195.04469299316406, "pos_frac": 0.703125, "sample": [0.5829010009765625, 22.999649047851562, 22.110763549804688, 34.71721649169922, 70.86495971679688, 82.88674926757812, 31.81806182861328, 141.7557373046875, -10.981925964355469, -4.046985626220703, -176.63015747070312, 12.942787170410156, 73.62820434570312, 22.20709991455078, -4.734825134277344, 191.74642944335938, -131.97503662109375, -14.176284790039062, 17.277816772460938, 1.8976364135742188, 149.1273651123047, 29.139575958251953, 0.9648952484130859, -11.404808044433594, 28.936630249023438, 4.303367614746094, -7.094856262207031, 11.721656799316406, 169.90350341796875, 67.9323959350586, -15.348731994628906, 9.879951477050781, 115.7967529296875, -0.11873245239257812, -83.53858184814453, 6.7833404541015625, 45.70337677001953, 19.803951263427734, 195.04469299316406, 41.2984733581543, -25.174530029296875, 119.42893981933594, 72.27179718017578, -120.34584045410156, -166.928466796875, 187.18849182128906, 194.03489685058594, 167.620849609375, -21.937744140625, -86.77037048339844, -95.72028350830078, -15.30755615234375, 33.44633483886719, 41.23090744018555, 139.6948699951172, 12.650955200195312, 11.46533203125, 108.80860900878906, 190.31210327148438, 96.61103820800781, 6.302818298339844, -13.026199340820312, 51.10151672363281, 140.6023406982422], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000585.npy"}
{"epoch": 0.8843537414965986, "step": 586, "batch_size": 64, "mean": 62.39876174926758, "std": 82.4639663696289, "min": -176.5504150390625, "p10": -13.11174850463867, "median": 46.69524002075195, "p90": 183.65963592529297, "max": 206.36297607421875, "pos_frac": 0.75, "sample": [138.467041015625, -3.424365997314453, 135.91488647460938, 47.19401550292969, 72.02203369140625, 157.25439453125, 66.06318664550781, -22.236812591552734, 3.3339614868164062, 27.78133773803711, 184.28689575195312, -11.452468872070312, 46.19646453857422, 191.56768798828125, 195.09014892578125, -6.104824066162109, 113.1788101196289, 104.78164672851562, 68.90052032470703, 185.0916290283203, -9.561027526855469, 172.33187866210938, 24.413105010986328, 104.49356079101562, 76.24333190917969, -0.8956871032714844, 43.58696746826172, 5.4360198974609375, 151.5723114013672, 24.77703094482422, 182.19602966308594, 152.3011474609375, 113.107421875, 106.06416320800781, 186.76168823242188, -23.713478088378906, -145.77117919921875, 206.36297607421875, 1.2773170471191406, 45.614410400390625, -3.479196548461914, -13.822868347167969, 19.438934326171875, 16.513702392578125, 73.40947723388672, -0.3231048583984375, 3.7306060791015625, 8.039302825927734, -1.8862171173095703, 76.12640380859375, 43.20918273925781, -40.886505126953125, 22.045440673828125, -5.1374664306640625, 189.631103515625, 1.336822509765625, -176.5504150390625, 157.1317138671875, -46.51422119140625, 111.46499633789062, 147.71273803710938, 82.69419860839844, 161.87533569335938, 57.256744384765625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000586.npy"}
{"epoch": 0.8858654572940288, "step": 587, "batch_size": 64, "mean": 66.33705139160156, "std": 96.82091522216797, "min": -170.14093017578125, "p10": -54.507155227661116, "median": 65.1988754272461, "p90": 184.7879409790039, "max": 213.75057983398438, "pos_frac": 0.71875, "sample": [-14.882553100585938, -79.34303283691406, 5.140262603759766, 104.54060363769531, 103.92459869384766, -72.42494201660156, 49.300941467285156, -39.861663818359375, 155.49700927734375, 74.76936340332031, 13.307697296142578, 108.71701049804688, -75.6702880859375, -24.464736938476562, 20.739669799804688, 70.67170715332031, 129.04391479492188, -0.9749908447265625, 37.855133056640625, 191.28378295898438, 123.77749633789062, 136.271240234375, -0.8724269866943359, 183.7007598876953, 146.80038452148438, 67.10075378417969, 202.51112365722656, -60.78379440307617, -168.90692138671875, 4.747684478759766, -7.502601623535156, 33.01791763305664, 154.8336639404297, 172.43289184570312, 53.12101745605469, 128.51280212402344, 160.39051818847656, -153.651611328125, -170.14093017578125, 213.75057983398438, 186.1287841796875, -12.545707702636719, 193.49610900878906, 63.2969970703125, 5.038017272949219, 173.80816650390625, 8.555076599121094, 185.25387573242188, 161.6506805419922, 53.42224884033203, 183.67446899414062, -23.225181579589844, 155.3759002685547, 3.241607666015625, 110.1236572265625, 152.41925048828125, 137.31329345703125, -12.940204620361328, -1.1691093444824219, 128.41529846191406, 49.901283264160156, 202.2555694580078, 167.90521240234375, -2.1040496826171875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000587.npy"}
{"epoch": 0.8873771730914588, "step": 588, "batch_size": 64, "mean": 58.586280822753906, "std": 85.37789154052734, "min": -148.41644287109375, "p10": -31.323418045043937, "median": 50.093257904052734, "p90": 167.66156311035155, "max": 227.1214141845703, "pos_frac": 0.8125, "sample": [5.8032073974609375, -55.59751892089844, 7.4295196533203125, 167.89010620117188, -17.789573669433594, 140.5354461669922, 120.73885345458984, 3.60205078125, 9.63641357421875, 79.86093139648438, 186.65267944335938, 121.81069946289062, 22.772361755371094, 161.40245056152344, 61.95613098144531, 122.08458709716797, 190.6561279296875, 227.1214141845703, 102.3023452758789, 132.52279663085938, 28.950775146484375, 94.02976989746094, -22.66217041015625, 175.09710693359375, 3.5359458923339844, 156.16189575195312, 56.24168395996094, 3.620086669921875, 4.399160385131836, -113.84027099609375, 4.194070816040039, 79.49565124511719, -99.27973175048828, 113.67010498046875, 173.84622192382812, 181.30381774902344, 7.781398773193359, 146.44754028320312, 43.94483184814453, -1.7371292114257812, 1.4783477783203125, -133.8282470703125, 7.696846008300781, 12.66558837890625, 122.6931381225586, 78.03202056884766, 167.1282958984375, -2.4870071411132812, 21.23095703125, 164.42672729492188, 8.844451904296875, -40.1654052734375, -35.03538131713867, 25.313812255859375, 134.65631103515625, 114.4913330078125, 0.484405517578125, 88.72895812988281, 5.788673400878906, -18.725723266601562, 59.1302375793457, 125.623046875, 163.17530822753906, -148.41644287109375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000588.npy"}
{"epoch": 0.8888888888888888, "step": 589, "batch_size": 64, "mean": 62.99828338623047, "std": 78.94786071777344, "min": -158.26495361328125, "p10": -30.032155609130847, "median": 56.46449089050293, "p90": 179.1218139648438, "max": 199.66754150390625, "pos_frac": 0.828125, "sample": [41.2159538269043, 29.154800415039062, -111.97358703613281, 184.65048217773438, 97.10503387451172, 35.18816375732422, 24.18292999267578, -0.30838775634765625, 16.135055541992188, 1.366140365600586, 56.73985290527344, -54.3560791015625, 199.66754150390625, 27.393001556396484, 17.141845703125, 2.8609466552734375, 101.65796661376953, 90.88451385498047, 147.49313354492188, 190.132080078125, 60.987640380859375, 27.3931884765625, 152.37258911132812, -59.157989501953125, 90.51118469238281, 129.2261505126953, -0.05890655517578125, 22.91204833984375, 151.67477416992188, 56.18912887573242, 8.5810546875, 76.57503509521484, 149.37648010253906, 137.8684539794922, 34.44591522216797, -4.0578460693359375, 52.41267395019531, 3.306640625, 85.32095336914062, 184.13970947265625, 25.213973999023438, 189.55325317382812, 167.41339111328125, 74.30218505859375, 88.99916076660156, -16.02178955078125, 62.55376434326172, -36.036598205566406, -37.263328552246094, 102.25347900390625, -56.25694274902344, 155.03599548339844, 188.75363159179688, -158.26495361328125, 116.3577880859375, 126.81108093261719, 156.63511657714844, 122.68401336669922, 4.3629150390625, 6.218326568603516, 74.93099975585938, 25.140766143798828, 186.18856811523438, 5.97503662109375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000589.npy"}
{"epoch": 0.890400604686319, "step": 590, "batch_size": 64, "mean": 58.44362258911133, "std": 86.67144775390625, "min": -189.4756622314453, "p10": -16.23951911926269, "median": 28.824039459228516, "p90": 183.32007598876953, "max": 201.0553741455078, "pos_frac": 0.734375, "sample": [109.04690551757812, 33.53594970703125, 156.24761962890625, 16.685970306396484, 88.2373046875, 182.3209228515625, 12.598922729492188, -189.4756622314453, 11.471877098083496, 6.0595550537109375, 22.138931274414062, -12.474666595458984, 8.912139892578125, 77.7664794921875, -171.0906982421875, 24.11212921142578, 164.10903930664062, -17.85302734375, -10.840171813964844, -0.8797836303710938, -33.632286071777344, 74.32081604003906, -34.20468521118164, 123.93841552734375, 191.07289123535156, 1.920196533203125, 23.40685272216797, 66.96771240234375, 201.0553741455078, 17.536636352539062, 76.37644958496094, 139.3575897216797, 12.064573287963867, -3.884521484375, 18.915130615234375, 157.90753173828125, 162.50489807128906, 86.10687255859375, 67.11883544921875, 19.094406127929688, 183.97711181640625, 47.77614974975586, 199.85302734375, 160.82888793945312, 186.53512573242188, -11.25107192993164, -0.866851806640625, 69.77214050292969, 52.42010498046875, 116.501708984375, -32.54882049560547, 171.16018676757812, 15.218021392822266, -4.470054626464844, -3.4778900146484375, 187.16995239257812, 55.62630844116211, 163.46200561523438, 183.36697387695312, -6.029148101806641, 14.888069152832031, -3.678924560546875, 183.2106475830078, -67.62503814697266], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000590.npy"}
{"epoch": 0.891912320483749, "step": 591, "batch_size": 64, "mean": 57.82063293457031, "std": 94.41322326660156, "min": -160.93463134765625, "p10": -47.09610519409179, "median": 33.17125701904297, "p90": 187.087548828125, "max": 211.82595825195312, "pos_frac": 0.71875, "sample": [-0.3696613311767578, 183.3213653564453, 93.93344116210938, 82.94351959228516, 158.426513671875, 129.00296020507812, 186.04965209960938, 19.219970703125, 32.906463623046875, 122.41468811035156, 6.611602783203125, -147.6210174560547, 31.00732421875, -85.43868255615234, 9.632038116455078, -100.90150451660156, 13.582809448242188, 83.57304382324219, 143.51806640625, 12.022134780883789, 181.26657104492188, 48.66376495361328, 168.1048583984375, -38.29181671142578, 164.61842346191406, -94.49507904052734, 190.2445068359375, 35.790706634521484, -5.571666717529297, 23.548248291015625, -160.93463134765625, -39.819976806640625, 211.82595825195312, 197.10415649414062, -0.17489242553710938, 193.18711853027344, -20.79674530029297, 122.212646484375, 153.9048614501953, -0.40871429443359375, 7.548713684082031, -3.5910568237304688, 44.16699981689453, 111.24092102050781, 1.6734485626220703, 188.47911071777344, -2.889505386352539, -66.57034301757812, 8.891433715820312, 168.40985107421875, -43.34797668457031, -48.70244598388672, 197.95640563964844, 58.6006965637207, -3.305450439453125, 187.51454162597656, 172.01834106445312, 38.74678039550781, 26.097274780273438, 33.43605041503906, 186.0912322998047, 10.321136474609375, 1.5702896118164062, 122.35110473632812], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000591.npy"}
{"epoch": 0.8934240362811792, "step": 592, "batch_size": 64, "mean": 52.02532958984375, "std": 83.71086120605469, "min": -154.716064453125, "p10": -21.000161361694335, "median": 35.21669387817383, "p90": 184.67742309570312, "max": 222.77255249023438, "pos_frac": 0.765625, "sample": [0.03163909912109375, 189.52304077148438, 158.32737731933594, 105.68316650390625, 150.54476928710938, 192.4071044921875, 184.2495574951172, -121.1178207397461, 4.73968505859375, 57.00254440307617, -56.3619384765625, 101.8082046508789, 25.791732788085938, 36.241981506347656, -18.819313049316406, 108.72450256347656, 182.15646362304688, 87.88224792480469, -35.0462646484375, 1.5886268615722656, -39.18980407714844, 137.20562744140625, 141.12820434570312, 184.8607940673828, 0.7130889892578125, 101.89614868164062, 45.92987823486328, 197.24850463867188, 14.5291748046875, 34.19140625, 62.33702087402344, 123.57062530517578, -5.777015686035156, 65.13214111328125, 2.794647216796875, 56.179134368896484, -154.716064453125, 17.01622772216797, -127.57791137695312, -6.463104248046875, -6.600091934204102, 40.09046173095703, 2.0966148376464844, -2.878814697265625, -4.354118347167969, 43.22434997558594, 51.523040771484375, 206.130126953125, 202.49859619140625, 24.205276489257812, 222.77255249023438, -0.5242633819580078, 3.7472305297851562, 11.288150787353516, -21.934810638427734, 5.240894317626953, 37.767486572265625, 12.989616394042969, -2.86053466796875, 5.962623596191406, 62.091861724853516, 8.661346435546875, 72.50751495361328, 149.6099090576172], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000592.npy"}
{"epoch": 0.8949357520786092, "step": 593, "batch_size": 64, "mean": 52.26756286621094, "std": 89.58231353759766, "min": -154.25802612304688, "p10": -37.39363975524902, "median": 32.6403751373291, "p90": 184.12857971191409, "max": 202.77867126464844, "pos_frac": 0.734375, "sample": [46.42135238647461, -2.9734420776367188, 127.0255355834961, 69.04643249511719, 164.87216186523438, -37.79615020751953, 26.47296142578125, -28.344207763671875, 195.52316284179688, 136.54075622558594, 141.61123657226562, 36.07516860961914, 24.669464111328125, 191.45291137695312, 6.044227600097656, -1.6824302673339844, 133.21481323242188, 79.5325927734375, 63.0718994140625, 29.205581665039062, 17.023433685302734, 138.9027862548828, 139.0390625, 173.78033447265625, -0.6994132995605469, 12.648258209228516, -17.289836883544922, 170.7767791748047, 43.10938262939453, -109.11788177490234, 134.12393188476562, -68.1209487915039, 0.884124755859375, 99.21031188964844, 61.58541488647461, 162.77330017089844, 3.3720035552978516, -123.18473815917969, 186.08248901367188, -17.342498779296875, -115.28858947753906, -36.45444869995117, 186.56524658203125, -133.11300659179688, 193.72647094726562, 179.5694580078125, 189.08657836914062, 14.402809143066406, -9.340999603271484, 202.77867126464844, 6.021026611328125, -2.9402313232421875, 63.448829650878906, 74.56261444091797, 60.90522766113281, -154.25802612304688, 28.397483825683594, 6.282249450683594, 2.0868988037109375, -13.06597900390625, 20.968017578125, 83.3999252319336, 15.164749145507812, 74.67869567871094], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000593.npy"}
{"epoch": 0.8964474678760394, "step": 594, "batch_size": 64, "mean": 51.68307876586914, "std": 82.78336334228516, "min": -170.30044555664062, "p10": -29.685463714599607, "median": 47.159780502319336, "p90": 177.4438232421875, "max": 210.9537811279297, "pos_frac": 0.703125, "sample": [69.89878845214844, 185.08863830566406, 1.6048660278320312, 52.78651809692383, 141.178955078125, 110.34056091308594, 48.545684814453125, -22.576927185058594, 49.576934814453125, -4.038810729980469, -15.413772583007812, -1.8597869873046875, -3.058565139770508, 15.667709350585938, 5.992473602294922, -0.33808135986328125, -28.336257934570312, -2.5289840698242188, 63.82667541503906, 58.715084075927734, 174.0189971923828, 0.6992721557617188, 1.8806571960449219, 52.19734191894531, 197.4962921142578, 120.67095947265625, 104.86569213867188, -19.050033569335938, 167.8758544921875, -1.1569671630859375, 147.79196166992188, 29.590316772460938, 210.9537811279297, 97.6654281616211, 38.142608642578125, -76.21080780029297, 58.12174987792969, 135.04861450195312, 48.31730270385742, -1.1438140869140625, 184.04428100585938, -30.263694763183594, 164.6009521484375, 181.97500610351562, 4.222217559814453, 46.00225830078125, 16.134658813476562, 178.91160583496094, 23.447097778320312, 165.05429077148438, 103.41201782226562, -170.30044555664062, 35.747039794921875, 203.66641235351562, 48.9742431640625, 54.293968200683594, -18.799949645996094, 12.233123779296875, 76.03353118896484, -40.52674865722656, 93.92926025390625, -36.91449737548828, -56.305206298828125, -144.70126342773438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000594.npy"}
{"epoch": 0.8979591836734694, "step": 595, "batch_size": 64, "mean": 68.9874267578125, "std": 99.11245727539062, "min": -184.6589813232422, "p10": -34.83257522583007, "median": 66.80145263671875, "p90": 182.4975357055664, "max": 202.9038543701172, "pos_frac": 0.765625, "sample": [-0.6728057861328125, 53.14783477783203, 169.75686645507812, 86.56208801269531, 97.80921173095703, 45.098052978515625, 150.08055114746094, 15.723514556884766, 25.91571807861328, 184.0815887451172, 183.31631469726562, 3.1536941528320312, -37.6744499206543, -162.9110870361328, 179.7610321044922, 14.880748748779297, 16.225757598876953, -23.47320556640625, 175.89585876464844, 15.918949127197266, 7.8908538818359375, 156.22593688964844, 176.20306396484375, 10.863811492919922, 154.03451538085938, 65.3408203125, 5.093711853027344, 163.94964599609375, 143.47531127929688, 102.10639953613281, -19.6978759765625, -83.04856872558594, 58.71714401245117, -12.336868286132812, 176.081298828125, 138.62127685546875, 202.9038543701172, -107.94998168945312, -184.6589813232422, 198.70944213867188, 199.57492065429688, 183.23922729492188, 200.7056884765625, 84.50784301757812, -76.11946105957031, 112.65544128417969, 160.17112731933594, 130.5541229248047, -2.6798858642578125, 15.67919921875, 22.821638107299805, 159.60220336914062, 140.90283203125, 29.062332153320312, 68.2620849609375, 6.062053680419922, 180.7669219970703, 144.59056091308594, -28.201534271240234, -1.9088287353515625, -4.793331146240234, -175.37603759765625, 156.53094482421875, 163.4640655517578], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000595.npy"}
{"epoch": 0.8994708994708994, "step": 596, "batch_size": 64, "mean": 19.77166175842285, "std": 84.5800552368164, "min": -178.32559204101562, "p10": -74.46699142456055, "median": 6.310182571411133, "p90": 164.00868377685563, "max": 204.27679443359375, "pos_frac": 0.5625, "sample": [9.837562561035156, 32.7080078125, 182.98187255859375, -15.244491577148438, 37.326107025146484, -103.36355590820312, -41.567352294921875, 27.111793518066406, -8.49005126953125, 88.30657958984375, -83.02774047851562, 43.244789123535156, -1.0396041870117188, -7.4129180908203125, 204.27679443359375, 180.26832580566406, 84.09259033203125, -43.05350875854492, -143.57135009765625, 123.88272094726562, -21.921600341796875, 126.06951904296875, 6.525623321533203, 25.814407348632812, 184.13461303710938, 109.11658477783203, 1.6299057006835938, -3.9023666381835938, -6.262153625488281, -35.532073974609375, -68.51911163330078, -23.24829864501953, 46.329139709472656, 0.4356727600097656, 70.92030334472656, -74.58080291748047, -144.69219970703125, 193.67459106445312, 6.0947418212890625, -98.17975616455078, -28.91085433959961, 0.3768768310546875, -15.972633361816406, -178.32559204101562, -4.9775848388671875, 13.84280776977539, -39.094478607177734, -3.11545467376709, 197.8588409423828, 27.251319885253906, 117.39859008789062, 36.83397674560547, -74.20143127441406, 9.209220886230469, -11.012283325195312, 192.6900634765625, 10.444839477539062, 59.06884765625, 27.48379898071289, 13.684867858886719, 52.551979064941406, 63.36579895019531, -3.2177276611328125, -59.020652770996094], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000596.npy"}
{"epoch": 0.9009826152683296, "step": 597, "batch_size": 64, "mean": 48.03459167480469, "std": 87.84869384765625, "min": -146.843505859375, "p10": -51.14253005981444, "median": 22.543487548828125, "p90": 175.72946777343753, "max": 200.41517639160156, "pos_frac": 0.703125, "sample": [-4.016551971435547, -119.81015014648438, -62.69459533691406, -60.42857360839844, 191.2398223876953, 7.582542419433594, 6.818242073059082, -30.435928344726562, -25.993663787841797, 191.21145629882812, -122.12237548828125, 39.48143768310547, -0.8588848114013672, 48.46356201171875, -146.843505859375, 13.070573806762695, 160.90206909179688, 161.00318908691406, 164.59722900390625, -27.412097930908203, 116.29039001464844, 37.82593536376953, -56.51727294921875, -38.601463317871094, 197.13168334960938, 152.95791625976562, 13.713695526123047, -7.185237884521484, 107.8526611328125, 1.9832210540771484, 69.55023193359375, 63.97307586669922, 188.21591186523438, 62.051780700683594, 5.44891357421875, 145.49530029296875, -4.7740020751953125, 10.984634399414062, 6.505048751831055, 190.54335021972656, 29.78293228149414, 24.154930114746094, 3.5659332275390625, -2.214160919189453, 94.93850708007812, -4.696685791015625, -122.85354614257812, 18.02093505859375, 20.932044982910156, 34.860992431640625, 200.41517639160156, 3.5211029052734375, 122.06676483154297, -21.460205078125, -24.321151733398438, 96.22735595703125, 12.47043228149414, 129.20816040039062, 180.50042724609375, 123.94560241699219, 93.46710205078125, 113.657958984375, 139.70957946777344, 161.1141357421875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000597.npy"}
{"epoch": 0.9024943310657596, "step": 598, "batch_size": 64, "mean": 31.86941909790039, "std": 77.20669555664062, "min": -131.00450134277344, "p10": -50.740763854980464, "median": 9.479072570800781, "p90": 149.20083007812505, "max": 207.69915771484375, "pos_frac": 0.609375, "sample": [3.534534454345703, -39.92884826660156, -14.599212646484375, 10.20416259765625, 64.6869125366211, 1.7905960083007812, 15.107662200927734, 159.55096435546875, 88.03497314453125, -125.4228286743164, 21.352012634277344, 34.42024612426758, 70.37570190429688, -0.5513725280761719, 191.51473999023438, -53.707096099853516, 196.8299102783203, -47.702392578125, 101.3271713256836, 73.31185150146484, -13.78807258605957, -131.00450134277344, -10.254440307617188, 107.58665466308594, 6.381927490234375, 25.518198013305664, 17.96019744873047, 64.07601928710938, -1.3119964599609375, -52.651885986328125, -31.683788299560547, -30.947853088378906, 2.9225311279296875, 89.4837646484375, 127.18586730957031, -3.0104942321777344, 12.654205322265625, -0.34482574462890625, 25.602691650390625, 33.097076416015625, 133.63519287109375, -0.21973419189453125, -0.15217208862304688, 190.35848999023438, -24.434127807617188, 101.00355529785156, 8.753982543945312, 3.5953636169433594, 153.46237182617188, 139.25723266601562, -6.811370849609375, -91.88603210449219, -8.552566528320312, 5.0370330810546875, 20.08147430419922, -109.105224609375, 159.96824645996094, 105.09066772460938, -15.020004272460938, -6.789882659912109, 207.69915771484375, 102.37460327148438, -52.04292297363281, 36.738616943359375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000598.npy"}
{"epoch": 0.9040060468631897, "step": 599, "batch_size": 64, "mean": 44.24391555786133, "std": 86.2352523803711, "min": -154.84832763671875, "p10": -41.39486923217773, "median": 37.36854553222656, "p90": 171.8964630126953, "max": 203.80496215820312, "pos_frac": 0.671875, "sample": [-34.28570556640625, 138.75868225097656, 20.455799102783203, 12.862808227539062, -21.759414672851562, 101.92190551757812, -71.79594421386719, 203.24114990234375, -51.71368408203125, -142.14796447753906, -10.970584869384766, 85.64300537109375, 9.16579818725586, 56.46158218383789, 54.63816833496094, -37.92115020751953, -4.839935302734375, -2.3825836181640625, 1.6972312927246094, 35.0440673828125, 181.56642150878906, 57.13608169555664, 48.44395446777344, -154.84832763671875, -1.3859939575195312, 39.693023681640625, 107.3406982421875, 159.78392028808594, -152.20457458496094, 26.02379608154297, 189.78594970703125, 171.6346893310547, -33.95071792602539, -37.102149963378906, 112.68433380126953, -31.537670135498047, 29.336875915527344, 70.29190063476562, -31.363811492919922, 3.378753662109375, 122.36878967285156, 19.247148513793945, 1.6921329498291016, -42.88360595703125, 180.71783447265625, 34.248252868652344, 89.03367614746094, -97.35877990722656, -36.930389404296875, 116.21171569824219, 197.4803466796875, 107.54443359375, 58.807884216308594, 95.59239196777344, 172.00865173339844, 93.05110168457031, 203.80496215820312, 70.61422729492188, 105.85557556152344, -1.9760971069335938, -7.66632080078125, 61.37531280517578, 147.856689453125, 44.13433837890625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000599.npy"}
{"epoch": 0.9055177626606198, "step": 600, "batch_size": 64, "mean": 41.892059326171875, "std": 79.93952178955078, "min": -129.17626953125, "p10": -34.46447257995605, "median": 16.2016658782959, "p90": 160.42195434570314, "max": 191.3323974609375, "pos_frac": 0.65625, "sample": [88.0888671875, 148.97142028808594, 65.28517150878906, -7.163906097412109, 188.97335815429688, 6.994384765625, 16.83123779296875, -126.50995635986328, 136.2763671875, 184.00323486328125, -11.904788970947266, 55.78586196899414, 116.72541046142578, 75.30097961425781, 0.8231964111328125, -13.27273941040039, -78.55218505859375, -90.66620635986328, -28.941253662109375, 147.515869140625, 10.410179138183594, 97.65457916259766, 160.93930053710938, 168.38185119628906, -15.467803955078125, 174.322998046875, 14.860397338867188, 190.21099853515625, -8.833560943603516, 129.7209930419922, 82.17242431640625, 16.336925506591797, -110.42955017089844, -9.449203491210938, -28.71057891845703, -36.625030517578125, 16.06640625, -7.411949157714844, -0.13798904418945312, 84.3543701171875, -0.9916877746582031, 71.06134033203125, 6.885852813720703, 96.78387451171875, 5.708688735961914, 25.78369903564453, -35.32666015625, -32.452701568603516, -129.17626953125, 137.71478271484375, 3.6297149658203125, 76.308837890625, 49.73976516723633, 20.76995849609375, 42.16203689575195, 2.149383544921875, 9.872467041015625, -13.948707580566406, 79.18179321289062, 191.3323974609375, -4.1273651123046875, -2.13714599609375, 118.02299499511719, 159.21481323242188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000600.npy"}
{"epoch": 0.9070294784580499, "step": 601, "batch_size": 64, "mean": 63.25646209716797, "std": 90.07038116455078, "min": -178.73788452148438, "p10": -28.653581619262685, "median": 41.97982406616211, "p90": 188.73454742431642, "max": 207.73016357421875, "pos_frac": 0.765625, "sample": [194.26168823242188, 54.59905242919922, 188.35150146484375, -178.73788452148438, 188.80714416503906, 189.26275634765625, 203.6044921875, -17.246337890625, 10.393623352050781, -2.9070091247558594, 180.04791259765625, -83.39472198486328, 102.66873168945312, 23.99966049194336, 18.874467849731445, 91.01937866210938, 127.74034118652344, 171.01644897460938, -6.549705505371094, 166.58828735351562, 175.06719970703125, 0.700531005859375, -1.647918701171875, 142.40084838867188, 192.5882110595703, 107.60138702392578, 36.770172119140625, -45.30274963378906, 22.72840118408203, -0.21920013427734375, 8.475704193115234, 160.82705688476562, 62.74896240234375, -66.14244079589844, 14.773096084594727, 77.62648010253906, 4.191383361816406, 47.189476013183594, 36.550743103027344, 4.823389053344727, -5.963714599609375, 51.74983215332031, 6.686107635498047, 98.28349304199219, 67.50135803222656, 105.26351928710938, 17.327110290527344, -33.54240036010742, -45.061973571777344, 10.707366943359375, 23.864147186279297, 202.711181640625, 0.38367462158203125, 2.589040756225586, 60.9867057800293, 107.23448181152344, 207.73016357421875, 188.56515502929688, 175.92340087890625, -4.493095397949219, -6.9132843017578125, 161.87574768066406, 180.3161163330078, -129.46072387695312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000601.npy"}
{"epoch": 0.90854119425548, "step": 602, "batch_size": 64, "mean": 55.60774612426758, "std": 83.89384460449219, "min": -127.50359344482422, "p10": -36.1738037109375, "median": 28.2062349319458, "p90": 187.84632263183596, "max": 201.555419921875, "pos_frac": 0.75, "sample": [-21.739852905273438, 138.12374877929688, 107.85749816894531, 200.00314331054688, 54.60227966308594, 7.5142822265625, 81.76097106933594, 31.962907791137695, 70.47262573242188, 12.373001098632812, 196.13235473632812, -14.144790649414062, 201.555419921875, 180.23291015625, -37.456260681152344, 194.72384643554688, -34.4644775390625, 191.9962158203125, 50.60609436035156, -7.1348114013671875, -29.788528442382812, 1.1729412078857422, 3.184803009033203, -79.89208984375, 41.71942138671875, 155.08448791503906, 16.80801773071289, 119.50684356689453, 68.65441131591797, 142.73294067382812, -7.5808258056640625, -0.3058319091796875, 65.61398315429688, 9.214969635009766, 105.79948425292969, 94.61485290527344, -4.507061004638672, 8.567649841308594, 163.92835998535156, 9.114082336425781, 1.0390853881835938, 106.53131103515625, 50.76030731201172, 176.60557556152344, -36.9063720703125, -127.50359344482422, 73.89547729492188, 91.89492797851562, 9.098819732666016, 0.133819580078125, -99.88114166259766, 24.449562072753906, 189.81634521484375, -10.953834533691406, 16.920639038085938, 132.40158081054688, 17.671630859375, 13.260272979736328, 194.18423461914062, 183.24960327148438, -69.92998504638672, 160.14759826660156, 11.588233947753906, -38.19831085205078], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000602.npy"}
{"epoch": 0.91005291005291, "step": 603, "batch_size": 64, "mean": 49.760406494140625, "std": 91.73974609375, "min": -183.71359252929688, "p10": -33.36254920959472, "median": 27.15921401977539, "p90": 171.63367156982423, "max": 217.52658081054688, "pos_frac": 0.6875, "sample": [126.67449188232422, 175.09222412109375, 179.98214721679688, -4.8780364990234375, -23.89763641357422, -7.66094970703125, 30.533065795898438, 23.785362243652344, -1.1891098022460938, -3.5882568359375, 158.80099487304688, 31.86927032470703, 59.55223083496094, -4.903558731079102, -127.82229614257812, 172.11427307128906, -36.76873779296875, 12.926338195800781, 0.9788131713867188, 217.52658081054688, 75.16848754882812, 59.0009765625, 156.65928649902344, 13.696434020996094, -0.9248123168945312, 45.42639923095703, 120.30142211914062, -12.341827392578125, 44.41978454589844, -157.9180450439453, 8.019119262695312, 15.645477294921875, 135.0209503173828, 96.40388488769531, -25.414775848388672, 131.67739868164062, 5.672206878662109, 3.7749977111816406, -3.210073471069336, 191.2256317138672, -91.58177947998047, 156.18841552734375, 140.8950958251953, 98.45346069335938, 154.7157745361328, 90.03961181640625, -1.9279708862304688, -41.15468215942383, 94.09542083740234, 1.9921493530273438, 1.650390625, -1.9278182983398438, 151.3100128173828, 182.300537109375, 20.94647216796875, -10.260040283203125, 0.13580703735351562, -183.71359252929688, 99.5864486694336, 170.51226806640625, 166.3689727783203, 88.3702392578125, 183.28916931152344, -167.04840087890625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000603.npy"}
{"epoch": 0.9115646258503401, "step": 604, "batch_size": 64, "mean": 47.61776351928711, "std": 95.9340591430664, "min": -193.26641845703125, "p10": -63.12315216064452, "median": 23.738224029541016, "p90": 184.94547424316406, "max": 211.541748046875, "pos_frac": 0.71875, "sample": [19.796600341796875, 109.42357635498047, -2.1114540100097656, -66.7891845703125, 56.97392272949219, 43.483123779296875, 7.461181640625, 96.55947875976562, 189.48184204101562, -54.56907653808594, 60.9615478515625, 12.06241226196289, -0.1610393524169922, 124.70936584472656, 66.87020874023438, 3.5354690551757812, 185.05262756347656, 48.12114715576172, 1.7144317626953125, 179.227783203125, -192.69056701660156, -19.584854125976562, 30.936824798583984, -85.68850708007812, 57.04353713989258, 4.735252380371094, -4.998205184936523, 34.856048583984375, 25.26927947998047, -193.26641845703125, 50.65910720825195, -5.696891784667969, -20.01856231689453, -18.156326293945312, 188.65892028808594, -102.28807830810547, 10.048965454101562, -7.049713134765625, 158.76605224609375, 145.37313842773438, 64.33909606933594, 191.45535278320312, 211.541748046875, 7.174198150634766, 167.48025512695312, 172.43890380859375, 17.278427124023438, 159.68016052246094, 199.41598510742188, -5.096927642822266, 190.05194091796875, 5.821258544921875, 108.03426361083984, 22.087295532226562, 12.764564514160156, 22.207168579101562, 184.69544982910156, -5.1439056396484375, 0.10343170166015625, 123.21868896484375, 148.93649291992188, -142.3270263671875, -107.90099334716797, 160.56793212890625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000604.npy"}
{"epoch": 0.9130763416477702, "step": 605, "batch_size": 64, "mean": 44.499263763427734, "std": 79.50927734375, "min": -182.48843383789062, "p10": -41.338150024414055, "median": 43.4217529296875, "p90": 140.54962158203128, "max": 222.44345092773438, "pos_frac": 0.71875, "sample": [85.06383514404297, 60.36714172363281, -76.88780212402344, 222.44345092773438, 124.56214141845703, 115.20702362060547, 57.75909423828125, 35.91035461425781, 136.6527099609375, -13.102130889892578, 94.55458068847656, -37.41827392578125, 31.467208862304688, -33.83251953125, -89.14118957519531, -69.47100067138672, 5.708953857421875, 17.710113525390625, 142.2197265625, 10.231147766113281, 19.721893310546875, -2.4981536865234375, 71.98274230957031, -77.99381256103516, 8.807991027832031, -29.264869689941406, 178.48504638671875, 63.592864990234375, 37.56086730957031, 68.72314453125, 64.85148620605469, 118.22190856933594, -182.48843383789062, 126.2878189086914, 3.865936279296875, 72.62957763671875, 107.4297103881836, 57.62882995605469, 5.622802734375, 73.05192565917969, -12.1253662109375, 9.276111602783203, 183.8011016845703, 52.690914154052734, 194.58145141601562, 71.9726791381836, 123.29525756835938, 106.39664459228516, 88.2345199584961, -8.58233642578125, 196.5930633544922, 60.74042510986328, 47.991722106933594, -105.82286071777344, -30.736740112304688, 21.177989959716797, 38.851783752441406, -43.018096923828125, 6.259239196777344, -32.81112289428711, -0.14582061767578125, 187.0236053466797, -5.3076324462890625, 91.39260864257812], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000605.npy"}
{"epoch": 0.9145880574452003, "step": 606, "batch_size": 64, "mean": 53.40687561035156, "std": 78.6915283203125, "min": -114.56871032714844, "p10": -30.603417968749994, "median": 18.370742797851562, "p90": 175.3536804199219, "max": 201.96128845214844, "pos_frac": 0.703125, "sample": [-6.601367950439453, 135.378173828125, 114.96302795410156, 194.56129455566406, 4.5692596435546875, 3.8766937255859375, 7.8537750244140625, 11.893814086914062, -0.1693572998046875, 136.3242950439453, 187.44482421875, 3.4290542602539062, 79.16856384277344, 99.40570068359375, 101.4378433227539, -114.56871032714844, 103.51242065429688, -3.435791015625, 50.468814849853516, 152.66036987304688, 126.55548095703125, 36.41168212890625, -44.809776306152344, 102.26483917236328, -26.43848419189453, 1.4709320068359375, 151.20565795898438, -6.256561279296875, 69.73015594482422, 152.55618286132812, 67.23946380615234, 27.77010726928711, -7.706020355224609, 183.3135986328125, 0.8498764038085938, 2.483633041381836, 6.526542663574219, -3.8828067779541016, 96.26359558105469, -4.564659118652344, 183.1761932373047, 175.17056274414062, -2.8227462768554688, 176.3408966064453, -0.7284164428710938, 201.96128845214844, -32.388389587402344, 12.304656982421875, 59.82424545288086, -104.34976959228516, 175.43215942382812, -24.92963409423828, 10.056434631347656, 15.074356079101562, -37.86949920654297, 104.91851806640625, 21.667129516601562, -55.712135314941406, -34.46299743652344, 156.97344970703125, 92.84042358398438, 123.7312240600586, 11.792831420898438, -3.1165809631347656], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000606.npy"}
{"epoch": 0.9160997732426304, "step": 607, "batch_size": 64, "mean": 53.20615768432617, "std": 82.33030700683594, "min": -186.87637329101562, "p10": -13.344554710388183, "median": 32.434871673583984, "p90": 179.2690200805664, "max": 211.31130981445312, "pos_frac": 0.78125, "sample": [1.1999473571777344, -10.000946044921875, -10.846908569335938, 18.981597900390625, 164.26124572753906, 20.947689056396484, 74.77859497070312, 25.7528076171875, -186.87637329101562, 52.15940856933594, 16.30237579345703, 4.926910400390625, 96.08218383789062, 163.51976013183594, 13.543449401855469, -74.878173828125, 35.555885314941406, 64.80815124511719, 103.18635559082031, 182.47264099121094, 39.05522155761719, 182.18798828125, 49.09788513183594, 158.88821411132812, 80.1355972290039, 59.01744079589844, -12.514686584472656, 111.52873229980469, -13.700212478637695, -3.311237335205078, 179.0764923095703, 22.549335479736328, 46.27362060546875, 178.4357452392578, 32.81222152709961, 179.35153198242188, 185.96661376953125, -2.87640380859375, 192.0985565185547, 12.461639404296875, -122.64407348632812, 18.027400970458984, 51.940513610839844, 120.62057495117188, 32.05752182006836, -6.179023742675781, -38.286338806152344, 3.6315994262695312, 180.79135131835938, 160.98854064941406, -29.38213348388672, 9.6480712890625, 108.95198822021484, 211.31130981445312, 13.571853637695312, 136.16517639160156, 19.232929229736328, -4.3997955322265625, 42.034210205078125, 116.72697448730469, 13.769264221191406, -93.41299438476562, 19.43048095703125, 8.18777847290039], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000607.npy"}
{"epoch": 0.9176114890400605, "step": 608, "batch_size": 64, "mean": 40.53437042236328, "std": 81.62832641601562, "min": -140.03195190429688, "p10": -31.130215454101563, "median": 27.073372840881348, "p90": 169.30271759033207, "max": 247.96524047851562, "pos_frac": 0.703125, "sample": [84.46997833251953, -140.03195190429688, 0.85845947265625, 148.38177490234375, -0.24359512329101562, -30.740524291992188, 80.31291961669922, -1.903106689453125, 27.094757080078125, -123.63847351074219, 118.363525390625, 44.01939392089844, 63.78953552246094, 124.38644409179688, 66.81224060058594, 45.024513244628906, 173.58041381835938, -132.00201416015625, 46.06391906738281, 1.1703453063964844, 55.619651794433594, -5.743812561035156, -2.0972442626953125, 190.3125, 77.31040954589844, -31.297225952148438, 4.414909362792969, 8.24445915222168, 89.43841552734375, 182.956787109375, 52.69952392578125, -94.93382263183594, -127.30032348632812, 78.19898986816406, -11.997909545898438, -12.482666015625, 9.8470458984375, -16.669021606445312, 58.6162109375, 3.5258102416992188, 25.186885833740234, 71.58573150634766, 11.380447387695312, -24.80715560913086, 23.961471557617188, 159.32142639160156, -14.480278015136719, 178.77484130859375, 247.96524047851562, 121.75437927246094, 15.579833984375, 1.0692214965820312, 81.04039001464844, 73.02579498291016, 1.9440860748291016, 183.49024963378906, -78.50773620605469, 183.232177734375, -2.060211181640625, -0.14752197265625, 27.05198860168457, 87.65943145751953, 53.16778564453125, 62.580047607421875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000608.npy"}
{"epoch": 0.9191232048374905, "step": 609, "batch_size": 64, "mean": 53.83350372314453, "std": 69.27542114257812, "min": -142.59593200683594, "p10": -10.767171478271482, "median": 54.0305290222168, "p90": 151.86138000488282, "max": 175.2541961669922, "pos_frac": 0.8125, "sample": [175.2541961669922, 118.45048522949219, 78.16876983642578, 2.500782012939453, 16.292144775390625, 0.8514060974121094, 134.0543975830078, 167.21743774414062, 15.096179962158203, 25.687480926513672, -7.757682800292969, 57.564117431640625, -11.979606628417969, 9.798988342285156, 95.4556884765625, 106.64448547363281, 148.92584228515625, 19.785446166992188, 92.11616516113281, 166.54002380371094, -102.55430603027344, 0.84326171875, 39.469940185546875, 53.46626281738281, -2.2743492126464844, 172.74159240722656, 150.7025604248047, 63.754310607910156, -90.84342956542969, 77.66213989257812, 161.5207977294922, -28.43910026550293, 66.40348815917969, -5.584449768066406, 80.62136840820312, 60.84354019165039, 54.59479522705078, -4.898475646972656, 72.5797119140625, 65.42274475097656, -7.9381561279296875, 39.198062896728516, 107.00820922851562, 112.07926940917969, 101.15985107421875, -142.59593200683594, 41.658512115478516, 33.03401184082031, 22.714244842529297, 2.2940731048583984, 14.916299819946289, 10.693477630615234, 115.72315979003906, 100.41484069824219, 9.844493865966797, 5.098941802978516, 53.279876708984375, 164.43328857421875, 96.75752258300781, -40.24504089355469, 152.35801696777344, 133.90469360351562, 66.75645446777344, -43.90306854248047], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000609.npy"}
{"epoch": 0.9206349206349206, "step": 610, "batch_size": 64, "mean": 50.93968963623047, "std": 98.08393096923828, "min": -196.98541259765625, "p10": -65.87448959350586, "median": 51.074073791503906, "p90": 188.7851135253906, "max": 233.48873901367188, "pos_frac": 0.671875, "sample": [87.93101501464844, 140.68389892578125, 11.922294616699219, 199.37632751464844, -16.365921020507812, 128.98670959472656, 52.419525146484375, 29.194229125976562, -196.98541259765625, 140.49627685546875, -8.525863647460938, -123.68766784667969, 99.97361755371094, -1.287017822265625, 2.157135009765625, 61.14717102050781, 171.85679626464844, -54.81884765625, 7.4362335205078125, -10.126152038574219, -30.168609619140625, -125.25222778320312, 49.72862243652344, 22.307456970214844, -68.61677551269531, -10.674819946289062, -24.0440673828125, 196.88551330566406, 62.4464111328125, 191.06298828125, 8.287429809570312, 230.02926635742188, 126.69498443603516, 159.8220672607422, -134.41159057617188, 134.02684020996094, 188.11181640625, 7.900876998901367, 120.6249008178711, 105.64190673828125, 64.6524658203125, 20.774682998657227, 68.71135711669922, -44.0546875, 175.44810485839844, -4.2820281982421875, -82.29907989501953, 63.92910385131836, 177.15794372558594, 189.07366943359375, 20.021316528320312, 72.85295104980469, 120.25616455078125, -29.088577270507812, -70.60786437988281, -59.47582244873047, -53.98609161376953, 138.30743408203125, 233.48873901367188, 61.118141174316406, 70.3410415649414, -40.92918395996094, 199.2054901123047, 37.33734893798828], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000610.npy"}
{"epoch": 0.9221466364323507, "step": 611, "batch_size": 64, "mean": 42.35655212402344, "std": 73.3323745727539, "min": -127.25326538085938, "p10": -27.43397006988525, "median": 12.209429740905762, "p90": 161.39054107666018, "max": 212.84674072265625, "pos_frac": 0.6875, "sample": [93.5951919555664, -0.3982963562011719, -4.654205322265625, -33.756629943847656, 76.93025207519531, 5.106678009033203, 21.7354736328125, 41.958656311035156, -7.0713348388671875, 10.733322143554688, -9.440898895263672, 175.86036682128906, 36.18894577026367, -54.32157897949219, 0.6345596313476562, -13.254981994628906, -127.25326538085938, 190.40838623046875, -27.94287109375, 212.84674072265625, 163.23402404785156, 79.00709533691406, 38.47392272949219, -48.93189239501953, 1.0002059936523438, -54.93953323364258, 72.43296813964844, 62.58595275878906, -6.211273193359375, 180.52249145507812, 0.16732406616210938, 2.614410400390625, 0.49829864501953125, 194.69491577148438, 5.6957855224609375, 3.2472305297851562, -26.24653434753418, 48.13304138183594, 17.363059997558594, 125.21382141113281, 115.95379638671875, -19.91449737548828, -13.22934341430664, 123.24797058105469, -2.42901611328125, 87.28427124023438, -6.9333343505859375, 13.129867553710938, 127.35005187988281, 157.08908081054688, 134.992431640625, 128.7135772705078, -43.62615203857422, 64.20712280273438, 28.681007385253906, 5.839794158935547, 184.0448455810547, -6.659355163574219, 57.50209045410156, -7.9248046875, 104.669189453125, 7.778972625732422, 11.288991928100586, 13.302925109863281], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000611.npy"}
{"epoch": 0.9236583522297808, "step": 612, "batch_size": 64, "mean": 62.06857681274414, "std": 93.59690856933594, "min": -184.57965087890625, "p10": -51.418350219726555, "median": 54.778276443481445, "p90": 189.02663574218752, "max": 218.58401489257812, "pos_frac": 0.765625, "sample": [-2.1368865966796875, -23.946273803710938, 86.98826599121094, 8.538787841796875, 6.4272918701171875, 5.033592224121094, 164.13214111328125, 67.58908081054688, -5.8393402099609375, 26.38921356201172, 172.60989379882812, 41.08464050292969, 45.34248352050781, 57.86408996582031, 93.26216125488281, 114.20506286621094, 76.01057434082031, 15.814346313476562, 21.370132446289062, 106.98069763183594, 97.20952606201172, -96.22494506835938, 106.30879211425781, -43.474029541015625, -121.41590881347656, 16.69481658935547, 190.37042236328125, -83.20072937011719, -9.655563354492188, -7.0039825439453125, 136.36264038085938, 197.81033325195312, -4.165008544921875, -2.0289230346679688, 12.077079772949219, 189.960693359375, 7.38232421875, 160.84243774414062, 186.29791259765625, -159.1805419921875, 201.29635620117188, 39.353599548339844, 145.5005340576172, -75.9076156616211, 195.04898071289062, 85.55532836914062, 176.95172119140625, -54.82305908203125, 24.94232940673828, 198.08123779296875, 51.69246292114258, 9.495582580566406, 218.58401489257812, 12.2515869140625, 144.52223205566406, 186.84716796875, 144.45382690429688, 51.26493835449219, 59.22926330566406, -184.57965087890625, 108.55857849121094, 169.8819580078125, 85.02931213378906, 126.4709243774414], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000612.npy"}
{"epoch": 0.9251700680272109, "step": 613, "batch_size": 64, "mean": 66.87257385253906, "std": 87.59844970703125, "min": -187.30487060546875, "p10": -23.650698852539055, "median": 50.047800064086914, "p90": 190.37029266357422, "max": 201.17459106445312, "pos_frac": 0.796875, "sample": [-5.536613464355469, 107.47147369384766, -0.5388298034667969, 4.859428405761719, 155.7006378173828, 36.968299865722656, 179.3228759765625, 104.39598083496094, -12.264739990234375, 31.932621002197266, 19.969364166259766, 78.29878234863281, -187.30487060546875, 93.72193908691406, 72.41548156738281, 74.17449951171875, 13.624486923217773, 55.03203582763672, 190.43174743652344, 49.719749450683594, 40.98640060424805, 1.0862350463867188, -45.47686767578125, 13.884037017822266, 190.22689819335938, 57.643287658691406, 190.83473205566406, 40.261722564697266, 14.920562744140625, 50.375850677490234, 189.8377685546875, 12.197208404541016, 121.30008697509766, -44.6828498840332, -4.58123779296875, 4.9110260009765625, -108.38812255859375, 195.32852172851562, 101.10882568359375, 201.17459106445312, 111.16554260253906, 178.31365966796875, 26.629024505615234, 147.85928344726562, 196.06494140625, -16.695175170898438, 182.39016723632812, 7.201793670654297, 37.08708190917969, 174.20115661621094, -77.78785705566406, -11.84442138671875, 83.09622955322266, 197.13995361328125, 155.79302978515625, 2.037639617919922, 19.628816604614258, 164.02532958984375, 28.265365600585938, 197.36231994628906, 90.81888580322266, 173.79833984375, -45.41825866699219, -26.631637573242188], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000613.npy"}
{"epoch": 0.926681783824641, "step": 614, "batch_size": 64, "mean": 53.611568450927734, "std": 75.6164321899414, "min": -148.91351318359375, "p10": -36.62869567871093, "median": 50.193254470825195, "p90": 162.08755645751953, "max": 186.9822540283203, "pos_frac": 0.78125, "sample": [152.1188507080078, 76.08280944824219, 25.621353149414062, 165.87515258789062, 126.32112884521484, 66.10789489746094, 6.2313232421875, 0.9589309692382812, 72.21176147460938, 31.069068908691406, 186.9822540283203, 5.8582305908203125, 142.0548095703125, 27.14478302001953, 142.58555603027344, 106.28400421142578, 127.69548034667969, 58.0671272277832, 17.47393798828125, 74.85302734375, -79.86085510253906, 49.97403335571289, 16.725006103515625, -72.741943359375, 177.46603393554688, -17.37969207763672, 66.6615982055664, -78.65692901611328, 65.92173767089844, -44.109344482421875, -1.6339111328125, 3.2725143432617188, 1.224365234375, 179.48593139648438, -22.251903533935547, -1.5255508422851562, 103.04251098632812, 9.693046569824219, -148.91351318359375, 129.38270568847656, 50.4124755859375, 180.2340850830078, 55.434059143066406, 82.60617065429688, 41.23621368408203, 121.45707702636719, 6.420982360839844, 12.9173583984375, 63.8931884765625, -37.50547790527344, 1.8892364501953125, 111.4628677368164, -22.730804443359375, -34.58287048339844, -51.75910186767578, 175.56045532226562, 130.19827270507812, 32.359649658203125, 121.93984985351562, 46.677833557128906, 74.93757629394531, 160.30398559570312, -2.4479217529296875, 162.85194396972656], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000614.npy"}
{"epoch": 0.9281934996220711, "step": 615, "batch_size": 64, "mean": 58.11692428588867, "std": 94.50919342041016, "min": -161.34503173828125, "p10": -44.015478515625, "median": 38.52890396118164, "p90": 187.33630828857423, "max": 198.12892150878906, "pos_frac": 0.796875, "sample": [68.6561279296875, 96.08415985107422, 179.2276611328125, 89.87830352783203, 17.834548950195312, 151.67886352539062, 15.720176696777344, -155.87594604492188, 176.35777282714844, 31.674297332763672, -128.1145782470703, 37.52538299560547, -32.769744873046875, 23.786041259765625, 38.66236114501953, -41.922271728515625, 66.48172760009766, 198.12892150878906, 176.01736450195312, 57.870880126953125, 128.14935302734375, 0.7406768798828125, 82.27169799804688, 0.15644073486328125, 147.8114013671875, -2.52294921875, -130.82733154296875, 40.77104187011719, 7.8165283203125, 192.08242797851562, 163.43606567382812, 9.2012939453125, 5.761131286621094, -9.530647277832031, 158.7149200439453, 138.58514404296875, 89.93536376953125, 38.39544677734375, 17.741683959960938, 188.20933532714844, 191.23583984375, 195.90853881835938, 172.9400634765625, 0.402252197265625, 129.7668914794922, -6.901344299316406, 9.487251281738281, -44.912567138671875, 7.2430419921875, 181.09017944335938, 23.10406494140625, 62.68408203125, -1.0511894226074219, 185.61334228515625, -57.99126434326172, 189.43994140625, 121.16629791259766, 90.62832641601562, 188.07472229003906, 1.8780384063720703, 30.20398712158203, 1.4775848388671875, -161.34503173828125, -124.46112060546875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000615.npy"}
{"epoch": 0.9297052154195011, "step": 616, "batch_size": 64, "mean": 52.18651580810547, "std": 83.01506805419922, "min": -100.14804077148438, "p10": -35.72127075195312, "median": 19.039928436279297, "p90": 178.03963775634767, "max": 209.0535888671875, "pos_frac": 0.703125, "sample": [178.83111572265625, 6.0042877197265625, 2.6954345703125, -10.146766662597656, 71.10205841064453, -44.388851165771484, 209.0535888671875, 141.20162963867188, -26.434951782226562, 86.49088287353516, -28.837738037109375, 1.2891197204589844, 69.90933227539062, -100.14804077148438, -2.9299163818359375, -18.771682739257812, 49.871856689453125, -0.8807754516601562, 32.072425842285156, 200.82760620117188, 16.417510986328125, 183.05673217773438, -83.58901977539062, -10.245780944824219, 176.19285583496094, -2.6870803833007812, -40.05396270751953, 0.9836654663085938, 4.785526275634766, 160.595458984375, 69.4718246459961, 5.083526611328125, 75.0582046508789, 146.53604125976562, 200.3446502685547, 188.75152587890625, -4.219356536865234, 3.4753875732421875, 2.81988525390625, 132.90524291992188, 166.43679809570312, 67.51773071289062, 22.5399112701416, 114.99639892578125, 146.18495178222656, 125.25778198242188, -7.2650909423828125, 21.66234588623047, -38.671356201171875, 0.39527130126953125, -73.05268859863281, -14.159027099609375, -6.7715301513671875, 139.22511291503906, 33.9124755859375, 150.1822509765625, 29.068763732910156, 4.713325500488281, 139.9322509765625, 4.454345703125, 148.22305297851562, 192.16883850097656, 10.699951171875, -80.20828247070312], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000616.npy"}
{"epoch": 0.9312169312169312, "step": 617, "batch_size": 64, "mean": 44.39569854736328, "std": 78.65626525878906, "min": -178.5557403564453, "p10": -23.12635993957519, "median": 33.11262512207031, "p90": 152.0581237792969, "max": 200.5887451171875, "pos_frac": 0.75, "sample": [44.55326843261719, -48.94378662109375, -88.75247192382812, 65.23959350585938, 97.16392517089844, 29.306854248046875, 12.807525634765625, 155.184814453125, 23.524852752685547, -127.03422546386719, 130.62905883789062, -4.183586120605469, 50.598846435546875, 1.5511512756347656, -1.1218528747558594, 65.83191680908203, 9.995552062988281, 23.016250610351562, 76.94856262207031, -1.3160285949707031, 166.40432739257812, 12.18935775756836, -10.190319061279297, 144.76251220703125, -19.05113983154297, 0.018398284912109375, -5.0320281982421875, 102.91572570800781, 23.346229553222656, 20.649860382080078, 45.23699951171875, 0.9936065673828125, 23.879234313964844, 188.47206115722656, 33.88958740234375, 200.5887451171875, 32.335662841796875, -64.86958312988281, 97.92707824707031, 131.6586456298828, 103.11170959472656, 42.01727294921875, -0.7065486907958984, 55.17409896850586, 185.97982788085938, 124.03470611572266, -151.0382080078125, -178.5557403564453, -24.872882843017578, 57.217193603515625, 143.5071563720703, 61.52021026611328, 23.687637329101562, 55.17710876464844, 195.80862426757812, 191.2193145751953, 26.256072998046875, 38.37800216674805, 27.633014678955078, 103.73220825195312, -1.3611068725585938, -11.5113525390625, 95.18717956542969, 38.60413360595703], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000617.npy"}
{"epoch": 0.9327286470143613, "step": 618, "batch_size": 64, "mean": 43.0288200378418, "std": 81.3189697265625, "min": -153.52951049804688, "p10": -43.83451957702636, "median": 15.47697639465332, "p90": 175.9082473754883, "max": 211.3221893310547, "pos_frac": 0.6875, "sample": [179.19207763671875, 190.93527221679688, 48.13874053955078, 38.951805114746094, 3.4331512451171875, -5.5904083251953125, -88.16754913330078, 4.7204132080078125, 140.5458984375, -2.0162601470947266, 3.6566524505615234, 109.3511962890625, 171.864501953125, 96.64553833007812, 192.88827514648438, 28.709136962890625, 69.26024627685547, 14.213872909545898, 211.3221893310547, 132.5375213623047, -38.64375305175781, 206.7393798828125, 6.1279144287109375, 106.03881072998047, -57.204254150390625, 187.6491241455078, 12.611644744873047, 15.122825622558594, 88.67295837402344, 21.35608673095703, -1.2243804931640625, -10.889945983886719, 50.92591857910156, -4.3583984375, 39.266021728515625, -26.758529663085938, 71.53112030029297, -57.42931365966797, 91.9534912109375, 25.529308319091797, -23.445228576660156, -0.8541831970214844, -1.1917686462402344, -153.52951049804688, 81.3785400390625, -44.26756286621094, 6.68817138671875, -136.42234802246094, 15.831127166748047, 134.00917053222656, -47.65966796875, 99.12779235839844, 7.126243591308594, 3.199695587158203, 3.8494415283203125, -42.8240852355957, -3.3117027282714844, 159.02223205566406, 58.846832275390625, 94.69068145751953, 14.398178100585938, 177.6412811279297, -1.636688232421875, 85.56945037841797], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000618.npy"}
{"epoch": 0.9342403628117913, "step": 619, "batch_size": 64, "mean": 59.74002456665039, "std": 90.42899322509766, "min": -190.74563598632812, "p10": -44.023252105712885, "median": 43.2082405090332, "p90": 188.26626281738282, "max": 223.15386962890625, "pos_frac": 0.78125, "sample": [157.52320861816406, 4.8584136962890625, 2.930540084838867, 122.61163330078125, 193.4904022216797, 191.55722045898438, 189.72137451171875, -51.937225341796875, -98.99375915527344, 194.44996643066406, 184.87100219726562, -69.83096313476562, 144.47528076171875, 11.051509857177734, -6.637542724609375, 2.517263412475586, 100.72406005859375, 33.32160949707031, 0.8365116119384766, -6.4162139892578125, 20.088470458984375, 184.73553466796875, 20.694841384887695, -85.26513671875, -0.8386058807373047, 83.48641967773438, 143.3782958984375, 183.16708374023438, 11.315616607666016, -1.0066871643066406, 7.079505920410156, -190.74563598632812, -37.17486572265625, 223.15386962890625, 98.59937286376953, 180.3668975830078, 4.876312255859375, 184.29949951171875, 192.16000366210938, -111.14127349853516, 38.08766555786133, 201.07736206054688, 133.65048217773438, 5.682064056396484, 19.234527587890625, 102.30181884765625, 54.12367248535156, 67.34272766113281, 48.32881546020508, 68.92303466796875, -46.958274841308594, 92.79730987548828, 78.65711975097656, 23.027023315429688, 72.99090576171875, 4.795467376708984, 74.24895477294922, 167.89122009277344, 10.35305404663086, 27.879093170166016, -20.93567657470703, 120.89912414550781, -8.957237243652344, 75.56737518310547], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000619.npy"}
{"epoch": 0.9357520786092215, "step": 620, "batch_size": 64, "mean": 44.221435546875, "std": 83.54029846191406, "min": -169.08291625976562, "p10": -56.48915824890136, "median": 22.910940170288086, "p90": 169.0420166015625, "max": 206.27432250976562, "pos_frac": 0.71875, "sample": [33.62046813964844, -53.09063720703125, -125.04231262207031, 168.15560913085938, -169.08291625976562, -0.8419418334960938, 31.524688720703125, 59.7760009765625, -31.4791259765625, 161.97140502929688, 152.4083709716797, 134.13394165039062, 29.337631225585938, 52.787200927734375, 18.30291748046875, 189.12725830078125, 8.111286163330078, 10.28732681274414, 0.31451416015625, -1.906097412109375, -60.770263671875, 14.51690673828125, -102.1314697265625, 187.0665283203125, -64.50701904296875, 177.39614868164062, -1.6855506896972656, -10.138336181640625, 206.27432250976562, -3.0220184326171875, 170.23052978515625, 160.95858764648438, -2.442270278930664, 21.331222534179688, 137.87429809570312, 92.24685668945312, 5.824798583984375, 85.69073486328125, 132.79190063476562, 1.5804252624511719, 51.17543029785156, 13.177070617675781, 55.799163818359375, 2.411649703979492, -57.9456672668457, -31.579002380371094, 157.9101104736328, 183.49356079101562, -27.34074592590332, 76.4266357421875, 2.8584442138671875, 169.42190551757812, 92.5875244140625, -2.1828155517578125, -72.34388732910156, 12.960819244384766, 33.41822814941406, 95.17953491210938, 118.71128845214844, 19.848953247070312, 43.2535400390625, 12.522178649902344, 24.490657806396484, 38.415557861328125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000620.npy"}
{"epoch": 0.9372637944066515, "step": 621, "batch_size": 64, "mean": 43.45226287841797, "std": 85.12129211425781, "min": -172.4891357421875, "p10": -55.4287368774414, "median": 24.256179809570312, "p90": 171.14294891357426, "max": 195.60125732421875, "pos_frac": 0.6875, "sample": [15.888416290283203, 53.38074493408203, 26.540260314941406, -11.098922729492188, 21.385589599609375, -172.4891357421875, 56.48114013671875, 116.06637573242188, -4.817533493041992, 38.787925720214844, 53.26028060913086, 158.3077392578125, 179.5335693359375, -5.165565490722656, 188.52886962890625, 34.395294189453125, 89.24310302734375, 157.4043426513672, 21.97209930419922, 176.0809326171875, 49.93797302246094, -0.1048431396484375, 149.186767578125, -106.1209945678711, 156.5240936279297, 69.144287109375, -28.328208923339844, -78.03076171875, 90.15655517578125, 38.56411361694336, 90.8812255859375, 17.698707580566406, 8.00482177734375, 195.60125732421875, 124.9970703125, -19.764297485351562, 77.53118896484375, 3.418954849243164, 30.16400146484375, -4.599365234375, -55.773826599121094, 89.70989990234375, 189.46719360351562, -1.606475830078125, 12.912887573242188, -17.293758392333984, 9.797725677490234, -54.62352752685547, -107.31204986572266, -5.72357177734375, 2.5070114135742188, 159.62098693847656, 92.8660659790039, -6.473871231079102, 136.91482543945312, -6.127567291259766, 15.793468475341797, 188.58721923828125, -122.57745361328125, 193.8353271484375, -97.87799072265625, 8.327648162841797, 20.814743041992188, 76.63198852539062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000621.npy"}
{"epoch": 0.9387755102040817, "step": 622, "batch_size": 64, "mean": 47.08264923095703, "std": 79.2149658203125, "min": -161.6846466064453, "p10": -43.157583236694336, "median": 30.70077896118164, "p90": 166.34124908447268, "max": 193.3679656982422, "pos_frac": 0.765625, "sample": [108.79032897949219, 97.24366760253906, -119.03021240234375, -66.64299774169922, 21.06948471069336, 70.29865264892578, -44.18183517456055, 66.73664093017578, -47.26725387573242, 9.31439208984375, 193.3679656982422, 35.133583068847656, 11.809858322143555, -9.690727233886719, 187.54623413085938, 22.90544891357422, 2.3672027587890625, 74.12448120117188, 47.04791259765625, 182.04747009277344, 71.99075317382812, 96.1419906616211, -5.853687286376953, 89.32173156738281, 192.95477294921875, 146.9528350830078, 26.064979553222656, 0.31009674072265625, 27.957462310791016, 87.60546875, 115.82640075683594, -161.6846466064453, 79.42938232421875, -43.516029357910156, 13.8765869140625, 12.252552032470703, 80.11738586425781, 148.5264434814453, 37.216835021972656, -3.9272842407226562, 123.14469146728516, 18.057693481445312, -4.869560241699219, 27.505168914794922, 74.73941040039062, 157.21588134765625, 94.05157470703125, -143.77392578125, -9.0067138671875, 136.0535888671875, 188.21786499023438, 7.616344451904297, -4.090396881103516, -23.725845336914062, 1.1334686279296875, 36.99712371826172, 174.06727600097656, 6.687591552734375, -42.32120895385742, 30.946937561035156, 102.44493103027344, 6.936511993408203, 170.2521209716797, 30.454620361328125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000622.npy"}
{"epoch": 0.9402872260015117, "step": 623, "batch_size": 64, "mean": 49.984649658203125, "std": 78.18663024902344, "min": -159.8165283203125, "p10": -30.66822433471678, "median": 41.09527015686035, "p90": 162.50004425048832, "max": 209.3428955078125, "pos_frac": 0.734375, "sample": [27.871200561523438, 23.348541259765625, -3.758197784423828, -159.8165283203125, 120.0881576538086, 96.251953125, 89.14306640625, 134.61752319335938, 119.09957885742188, 118.08284759521484, -8.919784545898438, 3.4729995727539062, 74.33147430419922, 21.730316162109375, -6.065879821777344, 126.09848022460938, -54.79700469970703, 37.586387634277344, 12.713638305664062, 120.86394500732422, 96.54396057128906, 12.303470611572266, 178.9925537109375, -6.765913009643555, -9.694656372070312, 84.90172576904297, 52.134429931640625, 155.25985717773438, -96.42898559570312, 139.4693145751953, 81.57283782958984, 209.3428955078125, 79.35340881347656, 50.24945068359375, -38.03912353515625, -54.53715515136719, 63.65099334716797, 48.21910858154297, 101.55546569824219, 194.8877716064453, 187.13189697265625, 187.509765625, 1.0202617645263672, 19.454444885253906, -13.469459533691406, -1.8021583557128906, -86.11784362792969, 116.96299743652344, 44.60415267944336, 19.774627685546875, 26.378646850585938, 165.6029815673828, 68.34197998046875, 84.42791748046875, 0.14142608642578125, 18.384605407714844, -2.759002685546875, 2.168834686279297, 54.699134826660156, -8.706207275390625, 181.46206665039062, -4.240081787109375, 11.3642578125, -108.23188018798828], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000623.npy"}
{"epoch": 0.9417989417989417, "step": 624, "batch_size": 64, "mean": 66.91777038574219, "std": 84.38616180419922, "min": -181.79031372070312, "p10": -4.542426300048828, "median": 48.03299331665039, "p90": 184.22162170410158, "max": 226.41769409179688, "pos_frac": 0.8125, "sample": [-95.12831115722656, -2.1424903869628906, 111.1134033203125, 89.82759094238281, 157.4525604248047, 147.14285278320312, 3.5888328552246094, 18.341796875, -4.543243408203125, 16.036727905273438, 68.66881561279297, 13.335533142089844, 2.899951934814453, 26.016712188720703, 132.21792602539062, 87.3154296875, -3.513347625732422, 226.41769409179688, 170.13238525390625, 26.668190002441406, 6.934410095214844, 50.581695556640625, -2.6501808166503906, 166.26722717285156, 9.713165283203125, 103.48401641845703, 45.484291076660156, -33.211334228515625, 106.70751953125, 124.85729217529297, 44.69342041015625, 205.75674438476562, 169.63265991210938, 189.45401000976562, 37.95181655883789, 6.051006317138672, 145.825927734375, 187.10498046875, 209.07940673828125, 166.5708770751953, 59.23891067504883, 107.22396850585938, 76.51664733886719, 92.76460266113281, -92.62602233886719, 182.97410583496094, 200.09811401367188, 6.615642547607422, -2.1655502319335938, 121.29798126220703, 138.69363403320312, 5.12481689453125, -4.540519714355469, 3.0934371948242188, -181.79031372070312, 184.7562713623047, 17.78826141357422, 10.896469116210938, 60.5767822265625, -14.212051391601562, -24.59661102294922, 0.7361335754394531, 158.58534240722656, 43.549354553222656], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000624.npy"}
{"epoch": 0.9433106575963719, "step": 625, "batch_size": 64, "mean": 32.61128234863281, "std": 86.99382019042969, "min": -187.7923583984375, "p10": -73.65824966430664, "median": 21.20009422302246, "p90": 163.47548522949225, "max": 202.57257080078125, "pos_frac": 0.65625, "sample": [95.28475952148438, 110.21326446533203, 27.977561950683594, 102.90261840820312, -9.071653366088867, 168.67660522460938, -39.26627731323242, 6.6876373291015625, 47.82186508178711, 34.6776123046875, 4.164283752441406, -16.939655303955078, 87.49566650390625, 1.826507568359375, 196.80560302734375, 8.74053955078125, 199.39959716796875, 76.85367584228516, -28.12488555908203, 19.513757705688477, 89.77632141113281, 151.33953857421875, 17.501667022705078, -35.309112548828125, -13.560012817382812, 107.81608581542969, 49.290252685546875, 41.35298156738281, -53.98173522949219, 8.023765563964844, -12.460174560546875, 23.05865478515625, 102.5985107421875, -75.29423522949219, 122.88579559326172, -1.338134765625, -81.82696533203125, 202.57257080078125, -24.41070556640625, -45.89057159423828, 109.02853393554688, 201.3631591796875, 115.77798461914062, 8.448741912841797, 15.233596801757812, 61.92236328125, -121.49462890625, -81.06668090820312, -137.57737731933594, 179.458740234375, -59.0899658203125, -187.7923583984375, -69.84095001220703, 172.6073760986328, 145.70814514160156, -1.7851829528808594, 54.3597412109375, 22.188247680664062, -90.47916412353516, -60.7741584777832, 64.8105697631836, 20.21194076538086, 29.78601837158203, 28.334083557128906], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000625.npy"}
{"epoch": 0.9448223733938019, "step": 626, "batch_size": 64, "mean": 59.367549896240234, "std": 98.51020812988281, "min": -144.13800048828125, "p10": -68.7914451599121, "median": 38.47087097167969, "p90": 193.173779296875, "max": 233.04501342773438, "pos_frac": 0.671875, "sample": [170.38294982910156, 39.619049072265625, 8.880348205566406, -30.053016662597656, 233.04501342773438, 175.71310424804688, 199.9164276123047, -4.1746063232421875, 38.08836364746094, 117.908935546875, -31.978939056396484, -13.165996551513672, -66.79594421386719, -0.38259124755859375, -1.4268417358398438, 195.810791015625, 163.6179656982422, -111.41796112060547, 184.56149291992188, -30.692264556884766, -52.03364944458008, 11.7415771484375, 38.85337829589844, 184.26609802246094, 30.60540771484375, -79.56999206542969, 52.212581634521484, 26.884994506835938, -34.302886962890625, 222.66265869140625, 107.11137390136719, -30.934642791748047, 187.020751953125, 167.7801055908203, 23.00336456298828, -69.84721374511719, -25.361557006835938, 136.44473266601562, 31.324581146240234, 83.09693908691406, 4.684356689453125, -7.978660583496094, 203.30667114257812, 1.3674507141113281, -93.57858276367188, 58.81140899658203, 105.22732543945312, 184.0653076171875, 101.29231262207031, 111.39701843261719, 124.97830963134766, 24.120838165283203, -144.13800048828125, 158.6654052734375, -48.86715316772461, 181.73187255859375, 35.03746795654297, 219.7191162109375, 81.16194152832031, 95.34862518310547, 202.95167541503906, -69.64665985107422, 106.05198669433594, -84.60171508789062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000626.npy"}
{"epoch": 0.9463340891912321, "step": 627, "batch_size": 64, "mean": 71.37725067138672, "std": 86.76113891601562, "min": -141.32388305664062, "p10": -17.4724925994873, "median": 61.22080421447754, "p90": 189.9526092529297, "max": 202.85897827148438, "pos_frac": 0.796875, "sample": [73.51180267333984, 1.2406044006347656, 98.16922760009766, 25.376007080078125, 132.78846740722656, 157.07266235351562, 163.81727600097656, 180.0266876220703, 202.85897827148438, 21.818634033203125, -87.66700744628906, 161.02304077148438, 192.76824951171875, 128.883056640625, 22.443756103515625, 1.5611419677734375, -28.809104919433594, 62.494529724121094, 196.72425842285156, 43.70586395263672, 157.6661376953125, 128.15931701660156, 11.119400024414062, 8.201904296875, 28.227603912353516, -32.74521255493164, 180.60525512695312, -7.586265563964844, 169.41448974609375, -1.3257179260253906, -5.271392822265625, 193.49917602539062, 183.38278198242188, 72.78186798095703, 38.66737365722656, 17.632308959960938, -1.4151077270507812, 12.638275146484375, -7.716621398925781, 198.38751220703125, 180.95751953125, 88.76580047607422, -14.081912994384766, -18.92559814453125, 22.078582763671875, 32.845947265625, -24.252357482910156, 121.10287475585938, 15.397071838378906, 153.06393432617188, 76.34820556640625, -140.37062072753906, 17.661575317382812, 91.07034301757812, 165.74139404296875, 84.82839965820312, 58.70283126831055, -141.32388305664062, 62.97526168823242, 3.555389404296875, 59.947078704833984, 175.89234924316406, 200.66017150878906, 201.37258911132812], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000627.npy"}
{"epoch": 0.9478458049886621, "step": 628, "batch_size": 64, "mean": 67.43045806884766, "std": 99.02460479736328, "min": -209.29615783691406, "p10": -39.90553131103515, "median": 78.0006103515625, "p90": 185.99828491210937, "max": 195.3552703857422, "pos_frac": 0.78125, "sample": [195.3552703857422, -21.610763549804688, 176.30624389648438, 143.30038452148438, -101.32596588134766, -43.90582275390625, 186.70159912109375, 0.2535858154296875, 171.26589965820312, 156.62771606445312, 193.06837463378906, 190.93809509277344, 162.03790283203125, 85.20909118652344, 13.253662109375, 6.005462646484375, -209.29615783691406, -169.22100830078125, 35.825775146484375, -11.311851501464844, 190.42433166503906, 21.248687744140625, 115.80271911621094, -151.0976104736328, 104.66809844970703, 119.70771026611328, 80.50511169433594, 136.1887664794922, 29.610519409179688, 175.03701782226562, 143.2603759765625, 178.35678100585938, 83.75161743164062, 186.58837890625, 7.514732360839844, 34.01806640625, 31.956424713134766, 0.37682342529296875, 140.8065185546875, -74.13705444335938, 150.04017639160156, 127.96192932128906, 58.317684173583984, 108.02698516845703, 75.96818542480469, 80.03303527832031, 184.62139892578125, 52.3785400390625, 189.85580444335938, 123.62950134277344, 71.6917953491211, -13.07403564453125, 8.916179656982422, -149.4804229736328, 141.25787353515625, 171.79940795898438, 41.40679931640625, 173.34176635742188, -30.571517944335938, 1.8290252685546875, -9.928815841674805, 70.12301635742188, -9.763618469238281, -16.896835327148438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000628.npy"}
{"epoch": 0.9493575207860923, "step": 629, "batch_size": 64, "mean": 50.71005630493164, "std": 74.09474182128906, "min": -100.81718444824219, "p10": -22.80221843719482, "median": 28.822025299072266, "p90": 155.0877212524414, "max": 226.68121337890625, "pos_frac": 0.71875, "sample": [-13.008865356445312, 22.989173889160156, -100.81718444824219, 103.70436096191406, 40.23082733154297, -1.9135475158691406, 13.322799682617188, 129.7277069091797, -37.85816192626953, 103.01731872558594, 43.71636962890625, 10.300384521484375, 2.651458740234375, 191.48789978027344, -29.786178588867188, -87.56719207763672, 5.437618255615234, 56.850563049316406, -19.399002075195312, 138.57345581054688, -23.942115783691406, 83.19036865234375, -20.142457962036133, 3.6759815216064453, -10.793680191040039, 18.534751892089844, 209.0167236328125, 226.68121337890625, -51.54141616821289, 151.26902770996094, 21.70257568359375, -17.179840087890625, 56.17896270751953, 149.24160766601562, 109.09507751464844, 41.33494567871094, 55.615577697753906, 54.03462219238281, 10.441776275634766, -3.377838134765625, -31.683700561523438, 91.50946807861328, 114.4120101928711, 5.638889312744141, 15.133438110351562, 5.6581878662109375, 27.397621154785156, 150.5091094970703, 96.3717041015625, 171.29852294921875, 96.12561798095703, -0.096405029296875, 156.72430419921875, -4.959754943847656, 30.246429443359375, 57.112823486328125, -15.153709411621094, 173.04718017578125, 52.1348762512207, 188.99737548828125, 17.207382202148438, 135.97518920898438, -13.696983337402344, 90.83836364746094], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000629.npy"}
{"epoch": 0.9508692365835223, "step": 630, "batch_size": 64, "mean": 58.783172607421875, "std": 80.05650329589844, "min": -98.25537109375, "p10": -23.88420028686523, "median": 37.145286560058594, "p90": 182.71543579101564, "max": 218.3644561767578, "pos_frac": 0.71875, "sample": [-18.22576904296875, 39.23658752441406, -3.8218040466308594, 4.198699951171875, 169.72621154785156, 154.64276123046875, 166.83287048339844, 1.6256141662597656, 201.41017150878906, 22.795074462890625, -1.7330474853515625, 91.60595703125, -11.732933044433594, 176.55044555664062, -4.449871063232422, 172.71484375, -88.12026977539062, 70.35658264160156, 88.71638488769531, -35.54461669921875, 35.053985595703125, 3.7257232666015625, 43.656890869140625, 153.37356567382812, 181.99334716796875, 28.44805908203125, -5.4492034912109375, 42.03496170043945, -98.25537109375, 16.001964569091797, -29.97662353515625, -4.2137603759765625, -0.5277786254882812, 200.069580078125, 33.811279296875, 132.63201904296875, 137.32003784179688, 2.0296058654785156, 49.52846908569336, 47.42152404785156, 218.3644561767578, 86.04219818115234, 188.0326385498047, -42.688323974609375, 193.8056640625, 92.06958770751953, 75.37928771972656, 3.5398025512695312, 107.0389404296875, -12.60955810546875, 10.983291625976562, 39.55345916748047, 3.6635284423828125, 187.6026611328125, 132.62002563476562, 29.085464477539062, -44.40937805175781, 33.4605712890625, -17.057838439941406, 183.02490234375, -26.309242248535156, 86.72358703613281, 69.23493957519531, -0.4896202087402344], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000630.npy"}
{"epoch": 0.9523809523809523, "step": 631, "batch_size": 64, "mean": 65.94947052001953, "std": 92.31715393066406, "min": -178.39535522460938, "p10": -56.19505996704102, "median": 51.785240173339844, "p90": 188.11384887695314, "max": 219.46859741210938, "pos_frac": 0.78125, "sample": [113.13032531738281, -7.546852111816406, 15.351715087890625, 182.2681884765625, 66.39529418945312, 135.73638916015625, 211.14141845703125, 50.087249755859375, 7.626129150390625, 26.99334716796875, -60.19158935546875, -63.921260833740234, 86.91279602050781, 61.62190246582031, -55.578346252441406, 53.150238037109375, 2.953601837158203, 9.19784927368164, 105.53561401367188, 206.41629028320312, 164.37939453125, -61.854347229003906, 219.46859741210938, 170.24459838867188, 128.24436950683594, 172.28567504882812, 14.236709594726562, 211.57806396484375, -23.617706298828125, 99.78998565673828, 18.753170013427734, 104.53569030761719, -178.39535522460938, 0.4511070251464844, 17.412559509277344, 152.29855346679688, 184.49708557128906, -5.302001953125, 207.45037841796875, 160.71311950683594, 57.203765869140625, -56.45936584472656, -9.810699462890625, 143.57151794433594, -32.434234619140625, -62.781524658203125, 2.6905441284179688, 8.865547180175781, -79.07221984863281, 86.45156860351562, -20.8818359375, 187.05752563476562, 24.39947509765625, 172.36468505859375, 3.6242446899414062, 162.6390380859375, 4.282905578613281, 50.42024230957031, 22.755569458007812, 189.3925323486328, 72.18934631347656, 173.69064331054688, 27.590164184570312, 188.56655883789062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000631.npy"}
{"epoch": 0.9538926681783825, "step": 632, "batch_size": 64, "mean": 58.3730354309082, "std": 83.9619140625, "min": -161.39666748046875, "p10": -32.701134490966794, "median": 47.597551345825195, "p90": 182.75433959960938, "max": 204.4228515625, "pos_frac": 0.75, "sample": [44.41012954711914, 124.3721923828125, 96.75726318359375, 2.6997528076171875, 28.06946563720703, 132.88287353515625, -33.169960021972656, 111.98751831054688, 2.2797012329101562, 75.60975646972656, -6.3979949951171875, -11.605575561523438, 88.40501403808594, 38.71799087524414, -35.89068603515625, 183.32064819335938, 119.16288757324219, 83.31539916992188, 80.92863464355469, -161.39666748046875, 14.717529296875, -2.5899581909179688, -71.95486450195312, 92.37821197509766, 1.7610015869140625, 104.11958312988281, 28.4815673828125, 144.194091796875, 0.03896331787109375, 31.21106719970703, 5.375038146972656, 202.99916076660156, 124.5195541381836, -8.020675659179688, 181.43295288085938, 36.520973205566406, 52.88420104980469, 78.69551086425781, 26.644195556640625, 86.60565185546875, 161.58267211914062, 168.790771484375, -11.807518005371094, 11.474782943725586, -2.8841094970703125, 184.84100341796875, 107.76485443115234, 0.4915447235107422, 204.4228515625, 184.9192657470703, 191.0653076171875, -149.11392211914062, 12.037460327148438, 136.47402954101562, -13.633056640625, -31.607208251953125, 194.24459838867188, 91.2181396484375, 168.4292755126953, 105.12603759765625, 50.78497314453125, -0.7677860260009766, -59.794883728027344, -62.65680694580078], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000632.npy"}
{"epoch": 0.9554043839758125, "step": 633, "batch_size": 64, "mean": 61.82661437988281, "std": 85.0855484008789, "min": -147.04725646972656, "p10": -19.058532714843746, "median": 38.1275577545166, "p90": 192.90731811523438, "max": 222.45022583007812, "pos_frac": 0.71875, "sample": [-3.02239990234375, 13.280231475830078, 68.75511169433594, 7.1436004638671875, 33.044334411621094, 126.0697021484375, 26.846298217773438, 168.96786499023438, 87.61810302734375, 7.855865478515625, -2.1046981811523438, 7.998411178588867, 199.95384216308594, 22.66433334350586, -1.5803947448730469, 148.34902954101562, -4.980865478515625, 28.59991455078125, 133.28195190429688, -19.951950073242188, 1.9204578399658203, -147.04725646972656, 169.8973388671875, 15.971511840820312, 110.1710433959961, 64.79336547851562, -34.876625061035156, 97.62879943847656, 43.21078109741211, 124.92771911621094, 157.27566528320312, 63.087196350097656, 43.48702621459961, 115.36770629882812, -73.37939453125, 11.342086791992188, -10.925880432128906, -5.855609893798828, -13.199470520019531, -16.973892211914062, -0.0740509033203125, -0.009429931640625, 45.72892379760742, 222.45022583007812, -6.488868713378906, 0.09626197814941406, 100.51737213134766, -105.15489959716797, 61.45042037963867, 193.13272094726562, 193.36871337890625, 166.09963989257812, 178.7730712890625, 201.42669677734375, 24.80773162841797, -31.124156951904297, 65.23790740966797, 1.4693717956542969, 152.09970092773438, 197.20664978027344, 202.61416625976562, -20.780250549316406, 192.38137817382812, 156.06307983398438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000633.npy"}
{"epoch": 0.9569160997732427, "step": 634, "batch_size": 64, "mean": 17.516376495361328, "std": 89.12398529052734, "min": -185.6560516357422, "p10": -85.65383224487304, "median": 3.586155891418457, "p90": 165.39925079345704, "max": 192.8251953125, "pos_frac": 0.609375, "sample": [-2.0328216552734375, -149.41510009765625, -2.0970191955566406, 185.63900756835938, 33.70338439941406, -15.988761901855469, 0.6519851684570312, -6.911962509155273, 17.580726623535156, 2.0042762756347656, 1.8033313751220703, -75.99560546875, -36.34750747680664, 14.440250396728516, -20.527740478515625, 14.971370697021484, 128.7801971435547, -2.975433349609375, -53.086212158203125, 30.286659240722656, -89.23847198486328, 6.888885498046875, 54.265289306640625, -172.8395233154297, 148.12307739257812, 3.5220165252685547, 166.205322265625, 156.01751708984375, 169.3993377685547, -185.6560516357422, 0.508697509765625, 38.93449020385742, 1.0709342956542969, -74.66989135742188, 163.51841735839844, -105.27032470703125, -118.2385482788086, 3.6635284423828125, 192.8251953125, 175.5237579345703, 23.366378784179688, -68.70321655273438, -39.31957244873047, 3.6502952575683594, -77.2896728515625, -122.09085083007812, -21.186199188232422, 18.466957092285156, 185.94015502929688, -3.746013641357422, -22.64609146118164, -0.01825714111328125, 115.56903076171875, 11.560482025146484, 19.946598052978516, -2.4520187377929688, 41.64751434326172, 128.12380981445312, 84.25379943847656, 37.47032165527344, 5.856986999511719, 25.262619018554688, 0.75604248046875, 177.59239196777344], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000634.npy"}
{"epoch": 0.9584278155706727, "step": 635, "batch_size": 64, "mean": 36.72964859008789, "std": 90.65315246582031, "min": -180.46646118164062, "p10": -73.5070831298828, "median": 22.53754425048828, "p90": 178.2061264038086, "max": 208.83355712890625, "pos_frac": 0.65625, "sample": [-23.905838012695312, 41.43109130859375, -51.05110168457031, 1.287078857421875, 55.55693054199219, -6.29510498046875, -78.45824432373047, -180.46646118164062, -166.14544677734375, 179.20492553710938, -1.719003677368164, 124.44314575195312, 18.63123321533203, 20.917221069335938, 150.4775390625, 185.79180908203125, 164.44088745117188, 3.2179031372070312, 90.57342529296875, 182.26760864257812, -3.1793060302734375, 25.12645721435547, -2.8419189453125, 203.55999755859375, 24.33599853515625, -57.98439025878906, 92.71942138671875, 2.1153621673583984, -0.5383453369140625, 59.5919303894043, 59.54742431640625, -69.00875854492188, 184.6453399658203, -25.610389709472656, 19.741134643554688, -75.4349365234375, 167.10586547851562, 24.157867431640625, 93.00838470458984, -16.02074432373047, -4.59979248046875, 175.87559509277344, -31.18637466430664, 195.996337890625, 15.890380859375, -88.77974700927734, 59.59941864013672, 129.9762420654297, 64.90777587890625, 25.31744384765625, 60.487083435058594, 208.83355712890625, -0.6268463134765625, 43.33294677734375, -162.14376831054688, 9.115463256835938, 72.967529296875, 14.810287475585938, 8.914199829101562, 72.95423889160156, -109.99905395507812, 116.21476745605469, 57.625579833984375, -0.021697998046875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000635.npy"}
{"epoch": 0.9599395313681028, "step": 636, "batch_size": 64, "mean": 62.914146423339844, "std": 79.6888656616211, "min": -138.44631958007812, "p10": -15.416300201416016, "median": 42.660240173339844, "p90": 183.33340148925782, "max": 216.89920043945312, "pos_frac": 0.765625, "sample": [184.13446044921875, 1.3691329956054688, -7.7745361328125, 19.24609375, 75.36749267578125, 5.859230041503906, 194.5696563720703, 29.38776397705078, 20.013809204101562, 111.36841583251953, 79.84600067138672, -5.4171600341796875, 181.46426391601562, 138.30284118652344, 13.76190185546875, -47.78643798828125, 30.059219360351562, -15.06158447265625, 41.559654235839844, 142.46990966796875, 59.507232666015625, 59.334266662597656, -0.9698715209960938, 119.32610321044922, 25.847402572631836, 167.89581298828125, 214.93472290039062, -11.370903015136719, 33.05293273925781, 148.14039611816406, 82.06438446044922, 120.90435791015625, 216.89920043945312, 117.30996704101562, -4.883216857910156, 69.71871185302734, 96.6406021118164, 43.51853942871094, -14.583389282226562, 29.664276123046875, 93.33296203613281, 178.24493408203125, -3.100851058959961, 198.8196563720703, 6.1182098388671875, -138.44631958007812, 194.40951538085938, 65.82996368408203, 32.77648162841797, 50.955528259277344, 180.93441772460938, -19.221824645996094, 129.70741271972656, 189.10269165039062, 41.80194091796875, -114.67123413085938, -18.370285034179688, -15.568321228027344, -23.294593811035156, 53.90438461303711, 32.04249572753906, 20.828765869140625, 6.166084289550781, 118.51170349121094], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000636.npy"}
{"epoch": 0.9614512471655329, "step": 637, "batch_size": 64, "mean": 53.618534088134766, "std": 80.47809600830078, "min": -134.46205139160156, "p10": -34.06678161621093, "median": 33.80155372619629, "p90": 187.74112854003906, "max": 221.49044799804688, "pos_frac": 0.734375, "sample": [75.78466796875, 166.38888549804688, 33.95884323120117, 221.49044799804688, -6.1344757080078125, 38.58903503417969, -4.443817138671875, 76.86186981201172, 118.84051513671875, 33.644264221191406, 1.6525650024414062, 92.36498260498047, 9.462303161621094, -16.721031188964844, 62.803077697753906, 58.59214782714844, -47.45943069458008, 161.30416870117188, 182.62376403808594, -38.50444030761719, 46.8619384765625, -30.43291473388672, 47.0643310546875, 35.043861389160156, 2.049985885620117, 29.348419189453125, -6.254158020019531, 180.70840454101562, 206.68455505371094, 204.11151123046875, 188.30712890625, -45.84217071533203, -134.46205139160156, 139.7076416015625, -5.2810516357421875, 92.34048461914062, 21.615394592285156, -2.027484893798828, 34.5906982421875, 98.45185089111328, 36.83952331542969, 188.15640258789062, 0.7140350341796875, 5.97314453125, 111.29165649414062, 2.8621444702148438, -35.62415313720703, 26.075347900390625, -6.247932434082031, 90.71456146240234, -5.2389678955078125, 200.16064453125, 74.78225708007812, 16.328353881835938, 11.4547119140625, -62.040435791015625, 186.77215576171875, 12.358774185180664, -8.512519836425781, -42.9644775390625, 91.32122802734375, 195.91102600097656, 12.680519104003906, 4.133544921875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000637.npy"}
{"epoch": 0.9629629629629629, "step": 638, "batch_size": 64, "mean": 47.013023376464844, "std": 82.53478240966797, "min": -138.8233642578125, "p10": -36.210167694091794, "median": 27.377123832702637, "p90": 172.17381744384767, "max": 197.2318115234375, "pos_frac": 0.734375, "sample": [-36.03450012207031, 60.91527557373047, 158.05738830566406, 44.83932113647461, -138.55255126953125, 5.976203918457031, 191.36680603027344, 128.83291625976562, -12.57009506225586, -51.7618408203125, 9.812084197998047, -16.28588104248047, 1.5671768188476562, -22.209259033203125, -16.26690673828125, 158.86207580566406, 11.849296569824219, -138.8233642578125, 172.49758911132812, -30.34259796142578, 29.61981201171875, 43.1563720703125, 141.29971313476562, 19.15289306640625, -61.47138977050781, 66.51744079589844, 79.57867431640625, 177.2772979736328, -31.27869415283203, 42.1280632019043, 129.0441436767578, 154.29412841796875, 52.681427001953125, 23.728134155273438, 23.208984375, -57.01072692871094, 8.854415893554688, 171.41835021972656, 40.6298828125, 173.24740600585938, 50.38239288330078, 118.56576538085938, 109.06803131103516, 22.6793212890625, 120.11141967773438, 12.760833740234375, 28.320873260498047, 197.2318115234375, -34.137298583984375, 47.67835998535156, 0.9628276824951172, 25.014488220214844, 69.8245849609375, 101.95191955566406, -36.28545379638672, 26.433374404907227, -97.85018920898438, 196.33383178710938, 187.27816772460938, -1.6689643859863281, 5.6163330078125, 22.364242553710938, -26.799453735351562, 155.1907196044922], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000638.npy"}
{"epoch": 0.9644746787603931, "step": 639, "batch_size": 64, "mean": 61.102783203125, "std": 87.1440200805664, "min": -109.46907043457031, "p10": -26.176143455505365, "median": 38.579307556152344, "p90": 186.2684753417969, "max": 209.16453552246094, "pos_frac": 0.703125, "sample": [116.21744537353516, 159.46170043945312, 102.58338165283203, -91.8619384765625, 188.2510223388672, 166.36265563964844, 6.594932556152344, -5.161201477050781, 71.45169067382812, 159.6147003173828, 51.28382110595703, 136.48135375976562, 199.24940490722656, 30.35228729248047, 38.280067443847656, 3.9168701171875, -29.199247360229492, -0.6614608764648438, -104.97123718261719, 38.87854766845703, 80.06849670410156, 48.20503234863281, -1.8663511276245117, 60.795623779296875, -12.41196060180664, 192.6584930419922, 5.0314483642578125, 184.0438690185547, -69.92848205566406, 187.2218780517578, 183.63677978515625, 1.2243499755859375, 0.8105678558349609, 140.9604034423828, -3.4110307693481445, 26.627727508544922, -19.122234344482422, -2.493144989013672, 209.16453552246094, 76.80046081542969, 177.12213134765625, 14.65496826171875, 207.56918334960938, 49.12773895263672, -6.220977783203125, 8.376449584960938, -4.493274688720703, -3.1707191467285156, 33.15691375732422, 7.736721038818359, 161.5045166015625, 118.20758056640625, 96.1839828491211, 90.54718780517578, -9.821685791015625, -4.474056243896484, 9.065399169921875, -109.46907043457031, -49.577476501464844, 192.54867553710938, -69.86378479003906, 173.66952514648438, 144.05767822265625, 158.99957275390625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000639.npy"}
{"epoch": 0.9659863945578231, "step": 640, "batch_size": 64, "mean": 46.48023986816406, "std": 89.78597259521484, "min": -184.1475830078125, "p10": -72.43381576538086, "median": 40.755882263183594, "p90": 161.03768005371094, "max": 202.4161834716797, "pos_frac": 0.75, "sample": [165.32794189453125, 21.791683197021484, 72.36376953125, -103.907470703125, 8.48764419555664, 95.9266357421875, 107.75772857666016, 55.315452575683594, 48.25457763671875, 129.23883056640625, 9.480186462402344, 12.322044372558594, 158.99969482421875, 40.423179626464844, -63.78916931152344, -12.175163269042969, 93.7452392578125, 131.25857543945312, 41.088584899902344, 188.1991729736328, -180.15591430664062, 104.68175506591797, 10.517173767089844, 125.78877258300781, -127.5115966796875, -30.288103103637695, 150.26678466796875, 10.406837463378906, -163.97103881835938, 96.73809051513672, 78.2052001953125, -24.15625, 19.30279541015625, 151.99710083007812, 108.71356964111328, 14.241600036621094, -76.13866424560547, 62.318206787109375, 67.57278442382812, 14.475982666015625, 1.2740249633789062, 112.98670196533203, 4.593666076660156, 50.03056335449219, 148.09213256835938, 176.356689453125, 96.44529724121094, -184.1475830078125, 27.052349090576172, 161.91110229492188, 25.21273422241211, 202.4161834716797, -4.2149505615234375, 179.44970703125, -16.9632568359375, -77.1211166381836, 176.59841918945312, -0.6928253173828125, -4.67401123046875, 119.30403900146484, -1.365835189819336, 39.277000427246094, 6.512065887451172, 123.28817749023438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000640.npy"}
{"epoch": 0.9674981103552532, "step": 641, "batch_size": 64, "mean": 47.14342498779297, "std": 82.23861694335938, "min": -184.79617309570312, "p10": -33.00554656982421, "median": 22.5402889251709, "p90": 166.0672180175781, "max": 229.19664001464844, "pos_frac": 0.75, "sample": [15.136035919189453, -97.27489471435547, 6.339433670043945, 119.61862182617188, 107.56640625, -5.9495391845703125, 166.33187866210938, -0.9844131469726562, -24.380355834960938, 22.71932601928711, -49.82769775390625, 22.38486099243164, 134.6533966064453, 202.68890380859375, 11.7625732421875, 49.903648376464844, 86.34852600097656, 184.39874267578125, 193.59498596191406, 103.65547180175781, 165.44967651367188, 8.43667221069336, -17.891395568847656, 0.4265899658203125, 14.494007110595703, 98.74072265625, -6.858734130859375, 58.86268615722656, 58.96156311035156, 148.084716796875, -16.949188232421875, 33.0963134765625, -14.225601196289062, 85.61833190917969, 3.852222442626953, -55.90601348876953, 148.11058044433594, -62.52020263671875, -36.702056884765625, 8.59576416015625, -184.79617309570312, 120.04052734375, 9.210258483886719, -107.61357116699219, 132.49462890625, 185.98159790039062, 1.7329750061035156, 0.8051013946533203, 47.912620544433594, 154.58172607421875, 229.19664001464844, 38.435340881347656, 179.80494689941406, 73.55785369873047, 18.2637939453125, 22.695716857910156, 14.045974731445312, 10.2218017578125, 30.361053466796875, 68.80165100097656, -24.150833129882812, -1.5692596435546875, 16.270065307617188, 110.53233337402344], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000641.npy"}
{"epoch": 0.9690098261526833, "step": 642, "batch_size": 64, "mean": 38.06249237060547, "std": 83.29491424560547, "min": -158.8187255859375, "p10": -66.17006454467773, "median": 35.88162612915039, "p90": 171.92561035156254, "max": 205.0140380859375, "pos_frac": 0.71875, "sample": [121.34638977050781, 11.40240478515625, 7.206584930419922, -64.88965606689453, 3.3646316528320312, 0.611785888671875, 84.03158569335938, 194.05242919921875, 40.19328308105469, -156.79736328125, 205.0140380859375, -10.165813446044922, 30.413070678710938, -94.14447784423828, 108.3394546508789, -3.8462295532226562, 124.85414123535156, 189.77821350097656, -40.446556091308594, -67.42774963378906, 102.09526062011719, 121.66145324707031, 60.747459411621094, -54.91436004638672, 39.25759506225586, 35.94684600830078, 187.2653350830078, 32.52903747558594, 81.41064453125, 48.53782653808594, 78.2388916015625, 59.761322021484375, 60.34434509277344, 24.82866668701172, -12.411518096923828, 175.99880981445312, -38.160003662109375, 37.880950927734375, 22.102767944335938, -21.084259033203125, 19.20069122314453, 49.49485778808594, 6.057195663452148, -66.71881103515625, 52.781150817871094, 39.78105545043945, -32.90043640136719, 76.00431823730469, 148.65591430664062, 195.09307861328125, 35.81640625, -5.635711669921875, 162.42147827148438, 55.703773498535156, 182.57862854003906, 23.10055160522461, 47.15341567993164, 19.745407104492188, -77.69480895996094, -158.8187255859375, 61.48786926269531, -21.706344604492188, -120.72148132324219, 20.192771911621094], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000642.npy"}
{"epoch": 0.9705215419501134, "step": 643, "batch_size": 64, "mean": 43.91210174560547, "std": 95.46736145019531, "min": -189.40872192382812, "p10": -71.22470321655274, "median": 30.94649600982666, "p90": 179.49569549560553, "max": 268.6382751464844, "pos_frac": 0.734375, "sample": [34.62260437011719, 22.544727325439453, 14.156299591064453, 60.410858154296875, 30.524991989135742, 193.63079833984375, 61.15869140625, 31.368000030517578, 102.3694076538086, -5.1631622314453125, 34.03779602050781, 35.85044860839844, 2.7126083374023438, 21.818923950195312, 134.28717041015625, 29.725799560546875, 15.420112609863281, -16.608997344970703, 187.45211791992188, 196.36940002441406, 85.39007568359375, 2.3923301696777344, 112.12577819824219, 158.01828002929688, -28.73902130126953, 196.73629760742188, 54.59379577636719, 11.257232666015625, 155.95068359375, 112.41567993164062, 111.68718719482422, 13.944599151611328, -30.696121215820312, 142.93350219726562, 88.33876037597656, -70.40220642089844, -71.57720184326172, 6.567169189453125, 195.0148468017578, 159.17230224609375, 53.394981384277344, -164.65786743164062, 55.701019287109375, -74.03313446044922, 4.09014892578125, -130.8597412109375, -140.59532165527344, -189.40872192382812, -41.46275329589844, 10.084915161132812, -107.82791900634766, 142.8045654296875, 268.6382751464844, 1.7775497436523438, 23.8673095703125, -67.85082244873047, -8.10101318359375, 190.284423828125, -11.474857330322266, 43.54289245605469, 139.28909301757812, 73.38055419921875, 160.9307098388672, -12.952163696289062], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000643.npy"}
{"epoch": 0.9720332577475435, "step": 644, "batch_size": 64, "mean": 40.763572692871094, "std": 94.9710464477539, "min": -182.52499389648438, "p10": -82.5366172790527, "median": 31.390304565429688, "p90": 180.0840072631836, "max": 198.94540405273438, "pos_frac": 0.65625, "sample": [177.49151611328125, 30.116592407226562, -178.2174072265625, 183.61399841308594, -11.72739028930664, -106.12874603271484, 3.4246253967285156, 14.867996215820312, 188.6251220703125, 47.04755401611328, 37.46092224121094, -92.05552673339844, 52.88151550292969, 96.67213439941406, 75.60107421875, 32.66401672363281, -2.1266212463378906, 10.112113952636719, 26.463333129882812, 179.670166015625, 23.827356338500977, 163.2660675048828, -154.47482299804688, -13.616928100585938, 12.483558654785156, 134.50648498535156, -4.538303375244141, 154.34109497070312, 9.933158874511719, -60.325828552246094, 64.49754333496094, 45.161376953125, -2.0023727416992188, 76.5512924194336, 180.26136779785156, -25.626182556152344, -13.536407470703125, 183.7144012451172, 116.01084899902344, 5.791164398193359, 167.05264282226562, 72.73851776123047, 87.73385620117188, 37.31932830810547, -27.317604064941406, -182.52499389648438, -24.869461059570312, 197.65005493164062, 130.70126342773438, 133.86312866210938, -22.32830810546875, -0.223358154296875, -8.5462646484375, -101.4659194946289, 51.92975616455078, 99.54563903808594, -148.5679931640625, -28.234508514404297, 95.38751983642578, 29.164623260498047, -36.14744567871094, 187.2432861328125, 37.13764953613281, 198.94540405273438], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000644.npy"}
{"epoch": 0.9735449735449735, "step": 645, "batch_size": 64, "mean": 54.74996566772461, "std": 82.82474517822266, "min": -177.23150634765625, "p10": -25.345166015624994, "median": 44.76000785827637, "p90": 172.6806579589844, "max": 208.17530822753906, "pos_frac": 0.75, "sample": [60.47673797607422, 137.35032653808594, 175.69589233398438, -51.84694290161133, 19.932601928710938, 197.78707885742188, 187.0718536376953, -7.899389266967773, 11.424919128417969, -6.66668701171875, 3.705657958984375, -17.729827880859375, 27.683820724487305, -1.06256103515625, 208.17530822753906, 55.88032531738281, 130.40090942382812, 124.47817993164062, 4.1202850341796875, -89.06413269042969, 188.7669677734375, -35.85545349121094, -11.47296142578125, 161.6138153076172, -4.0891265869140625, -20.541061401367188, 13.917121887207031, 113.1488037109375, 165.64511108398438, 87.53579711914062, 72.92672729492188, 7.6311187744140625, 66.1155776977539, 73.55884552001953, 57.637969970703125, 99.8718490600586, -82.22828674316406, -1.7648086547851562, 51.84300231933594, 185.75914001464844, 154.8804473876953, 147.84042358398438, 191.29644775390625, 31.533859252929688, 27.00281524658203, 105.74620056152344, 128.92745971679688, 41.9291877746582, 20.613128662109375, 11.944412231445312, -116.16883850097656, 154.87725830078125, 24.5582275390625, 66.89837646484375, -27.404067993164062, 10.986133575439453, 82.12073516845703, -13.399482727050781, 9.974639892578125, -177.23150634765625, 2.9390640258789062, 163.23587036132812, 53.371742248535156, 47.59082794189453], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000645.npy"}
{"epoch": 0.9750566893424036, "step": 646, "batch_size": 64, "mean": 57.74522018432617, "std": 80.84638977050781, "min": -177.43609619140625, "p10": -11.18354797363281, "median": 28.98103141784668, "p90": 167.12358245849612, "max": 221.55511474609375, "pos_frac": 0.8125, "sample": [-7.7163848876953125, 0.5869712829589844, -5.1461334228515625, 24.250839233398438, 185.22325134277344, 122.64111328125, 20.785568237304688, 126.46072387695312, 113.67404174804688, 64.5224609375, 15.929206848144531, 2.4630050659179688, 186.9770965576172, 168.3590087890625, 152.46865844726562, -1.484039306640625, 117.11709594726562, 126.42532348632812, 176.51548767089844, 117.02301025390625, 173.10476684570312, 26.965042114257812, -165.45657348632812, 66.9416732788086, -28.72396469116211, 164.2409210205078, 4.311122894287109, 92.99273681640625, 45.10275650024414, 4.8057861328125, 9.333740234375, 1.9706344604492188, 4.5184326171875, 183.1426544189453, 6.866111755371094, 19.19747543334961, 54.42257308959961, 1.3299026489257812, 96.49858856201172, 154.1949005126953, -1.499053955078125, -20.653457641601562, 154.63766479492188, 16.528411865234375, 103.41024780273438, 221.55511474609375, 139.0313262939453, 37.995880126953125, 24.3114013671875, 157.97584533691406, 47.872222900390625, 5.3387908935546875, -11.988983154296875, 11.054237365722656, -177.43609619140625, -9.30419921875, 69.61216735839844, -31.175342559814453, -17.717636108398438, 159.03353881835938, 3.5070762634277344, 27.416580200195312, 30.545482635498047, 132.80734252929688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000646.npy"}
{"epoch": 0.9765684051398337, "step": 647, "batch_size": 64, "mean": 72.1575698852539, "std": 77.93916320800781, "min": -138.43939208984375, "p10": -8.2820068359375, "median": 59.07244873046875, "p90": 179.40820770263673, "max": 221.80345153808594, "pos_frac": 0.84375, "sample": [-7.391021728515625, 99.1760482788086, 159.27828979492188, -0.8678817749023438, 9.481185913085938, 34.0670166015625, 45.55063247680664, 182.35325622558594, 120.64599609375, 169.5722198486328, 99.48052978515625, 141.75180053710938, 31.828598022460938, 31.74829864501953, -11.772209167480469, 19.431640625, 221.80345153808594, 55.68101501464844, 1.8400955200195312, 43.924285888671875, 172.18504333496094, 51.00400924682617, 27.41326141357422, 118.60316467285156, 198.2995147705078, 166.14039611816406, 60.985992431640625, 24.65605926513672, 66.75798034667969, 191.34613037109375, 90.60453033447266, 7.2673797607421875, 68.79154968261719, -36.82621383666992, 96.9711685180664, -36.29663848876953, 34.84954071044922, 181.56875610351562, 8.056617736816406, -126.80921936035156, 72.9892578125, 30.361312866210938, 68.97479248046875, 115.19085693359375, -138.43939208984375, 40.05844497680664, 36.36199188232422, 82.75596618652344, 23.976486206054688, 131.18943786621094, -8.216926574707031, 187.152587890625, 114.12876892089844, -13.227569580078125, 20.843856811523438, 174.36692810058594, 57.158905029296875, 206.72598266601562, 21.896942138671875, 150.20526123046875, 172.2050018310547, -8.309898376464844, 112.44608306884766, 154.13711547851562], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000647.npy"}
{"epoch": 0.9780801209372638, "step": 648, "batch_size": 64, "mean": 47.92961502075195, "std": 78.82329559326172, "min": -173.9031524658203, "p10": -27.47030296325683, "median": 38.80836486816406, "p90": 149.61643676757814, "max": 204.0843505859375, "pos_frac": 0.78125, "sample": [129.87646484375, -0.24725341796875, 134.6390380859375, -173.9031524658203, 185.9235076904297, 3.4562911987304688, -96.67422485351562, 18.177162170410156, 133.2474822998047, 54.68517303466797, -1.2602958679199219, 87.28888702392578, 18.442474365234375, -3.01513671875, -29.624359130859375, -22.444171905517578, 60.588924407958984, 2.6255569458007812, 204.0843505859375, 125.49078369140625, 192.28536987304688, 39.15580368041992, 81.18759155273438, 185.27500915527344, 84.9175796508789, 5.812469482421875, -2.7584228515625, 149.79132080078125, 4.866340637207031, 97.18624114990234, 4.89849853515625, -31.234901428222656, 149.2083740234375, 15.43902587890625, 1.1355533599853516, -119.05669403076172, 141.6028594970703, 44.55685043334961, 16.34088134765625, 85.42533874511719, 47.108642578125, 20.085159301757812, 196.4600067138672, 38.4609260559082, -18.570693969726562, 11.231849670410156, 167.30722045898438, 43.886810302734375, 59.935638427734375, 101.00234985351562, 61.393653869628906, 55.54583740234375, 34.164031982421875, -58.141265869140625, -119.25625610351562, 119.18927001953125, 81.39165496826172, 11.258228302001953, 71.90872192382812, -18.496932983398438, 8.420574188232422, 18.51653289794922, 22.200469970703125, 135.10020446777344], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000648.npy"}
{"epoch": 0.9795918367346939, "step": 649, "batch_size": 64, "mean": 53.71076202392578, "std": 91.3177719116211, "min": -132.58163452148438, "p10": -71.45447463989258, "median": 47.53889465332031, "p90": 180.42583312988282, "max": 208.2930908203125, "pos_frac": 0.734375, "sample": [114.78593444824219, 112.4963607788086, -3.9857845306396484, 27.108901977539062, 139.45852661132812, 163.25067138671875, 18.75775909423828, 145.94854736328125, 3.536062240600586, 183.2551727294922, 7.334373474121094, 188.57191467285156, 18.735214233398438, -22.89130401611328, 147.73658752441406, 61.298667907714844, 0.42182159423828125, 10.04566764831543, 167.72471618652344, 63.19053649902344, 175.61192321777344, -89.38447570800781, -132.58163452148438, -10.871795654296875, -49.793670654296875, 125.78764343261719, 33.77912139892578, 17.797958374023438, -72.00664520263672, -70.16607666015625, 81.90506744384766, 127.47611999511719, 131.52455139160156, 81.27432250976562, 122.01409912109375, 122.36276245117188, 86.7403793334961, 129.2567596435547, 78.06210327148438, 80.83741760253906, 25.92169189453125, 208.2930908203125, 189.48452758789062, -14.18289566040039, -32.20368957519531, 182.69920349121094, 10.26473617553711, 0.69061279296875, -33.4832763671875, 12.120882034301758, 0.9075469970703125, 200.34219360351562, -107.14691162109375, -82.14801025390625, 182.4889373779297, -116.84133911132812, -61.418922424316406, 25.14202880859375, 88.33584594726562, 154.33612060546875, 62.08648681640625, 146.74952697753906, -5.612998962402344, -115.74310302734375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000649.npy"}
{"epoch": 0.981103552532124, "step": 650, "batch_size": 64, "mean": 56.28281784057617, "std": 88.85587310791016, "min": -139.11891174316406, "p10": -43.07516632080078, "median": 40.930551528930664, "p90": 186.89339447021484, "max": 203.98635864257812, "pos_frac": 0.75, "sample": [-5.743797302246094, 111.02964782714844, 55.944793701171875, 191.64883422851562, 3.4179611206054688, 183.4113311767578, 62.082069396972656, 90.67312622070312, 54.423377990722656, 8.53742790222168, 149.97000122070312, 174.7911834716797, -61.93642044067383, 3.3764266967773438, 7.844638824462891, 96.98649597167969, 112.59033203125, 6.1566314697265625, 141.74807739257812, 3.132427215576172, 160.9422607421875, 42.00236892700195, 10.637092590332031, -46.163482666015625, 187.8112030029297, 187.78805541992188, 194.83953857421875, 162.70301818847656, 116.08229064941406, -1.2760391235351562, 8.913955688476562, -133.44189453125, -33.091636657714844, 153.7494354248047, -85.815673828125, 10.09908676147461, 185.7271728515625, 49.853641510009766, 101.0972900390625, -10.471237182617188, 24.0626220703125, 152.0460205078125, -139.11891174316406, 18.47101593017578, 16.617721557617188, -34.79763412475586, 187.39320373535156, 197.04043579101562, 27.448150634765625, -15.323013305664062, 65.87458801269531, -2.627838134765625, 72.72418975830078, 39.858734130859375, -23.150611877441406, 121.29467010498047, -42.67640686035156, -43.246063232421875, 157.46287536621094, 203.98635864257812, -105.70272827148438, 3.439960479736328, 67.54454040527344, 1.4074249267578125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000650.npy"}
{"epoch": 0.982615268329554, "step": 651, "batch_size": 64, "mean": 45.47364044189453, "std": 82.39358520507812, "min": -119.91339874267578, "p10": -34.61907424926758, "median": 23.35807991027832, "p90": 177.08576202392578, "max": 198.62969970703125, "pos_frac": 0.671875, "sample": [3.0992584228515625, -1.0835037231445312, 98.6776123046875, -26.4315185546875, 10.057304382324219, -81.23030853271484, -26.324256896972656, 62.25572204589844, -2.4061508178710938, 177.47483825683594, -0.28326416015625, -15.981582641601562, -80.0289306640625, 38.86975860595703, 160.83206176757812, -100.47904968261719, -32.066017150878906, 120.08018493652344, 80.91934204101562, 171.58299255371094, 142.2279510498047, 3.5614280700683594, 37.16326904296875, 117.92489624023438, -119.91339874267578, 84.24044799804688, 19.75726318359375, 106.38737487792969, -11.601806640625, 131.13381958007812, 26.95889663696289, 189.7060546875, 31.795997619628906, -20.624847412109375, 12.7532958984375, -1.9391937255859375, -18.675582885742188, 30.59685516357422, -11.086761474609375, 42.560386657714844, 187.65696716308594, 18.692676544189453, 188.0674591064453, -0.2558860778808594, 3.0181007385253906, 189.0343475341797, 196.78399658203125, 176.17791748046875, -6.6589202880859375, 172.03158569335938, 38.696197509765625, 0.7908248901367188, 12.317741394042969, 38.09612274169922, 101.78768920898438, 70.69540405273438, 97.93075561523438, 2.6153907775878906, -35.71324157714844, -89.80078125, 8.685562133789062, 56.11909103393555, 198.62969970703125, -65.54657745361328], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000651.npy"}
{"epoch": 0.9841269841269841, "step": 652, "batch_size": 64, "mean": 54.54006576538086, "std": 87.92820739746094, "min": -116.95125579833984, "p10": -51.91326370239257, "median": 27.526451110839844, "p90": 184.35225982666017, "max": 225.97518920898438, "pos_frac": 0.71875, "sample": [3.2060317993164062, 76.39289093017578, 160.50865173339844, 17.347030639648438, 151.16668701171875, 192.35525512695312, 0.5652427673339844, 92.6663589477539, 160.88644409179688, 0.4493370056152344, 162.60919189453125, -43.00764465332031, 3.254789352416992, 28.096744537353516, 93.70703125, 107.59455871582031, 57.14044189453125, 182.14125061035156, 113.2010726928711, -7.843116760253906, -10.203659057617188, 188.61212158203125, 40.9259033203125, 12.27865982055664, 7.637115478515625, 4.188751220703125, -88.25971984863281, -8.566909790039062, 175.44268798828125, 104.26344299316406, 125.41201782226562, -18.743053436279297, 26.956157684326172, -2.8431167602539062, 26.29034423828125, -9.100234985351562, 185.29983520507812, -91.26998138427734, -76.82677459716797, 18.062210083007812, 41.651668548583984, 196.45956420898438, -1.6261825561523438, 96.94491577148438, 12.738426208496094, 7.030811309814453, 176.9461212158203, 192.4898681640625, -5.15056037902832, 157.20181274414062, 225.97518920898438, -55.729957580566406, 132.81382751464844, 7.857643127441406, -36.66468048095703, 62.11357116699219, -89.19775390625, 68.6637954711914, -66.00469207763672, 191.4364013671875, 34.75926208496094, -34.5419921875, -116.95125579833984, 129.35423278808594], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000652.npy"}
{"epoch": 0.9856386999244142, "step": 653, "batch_size": 64, "mean": 64.99937438964844, "std": 85.83916473388672, "min": -119.64254760742188, "p10": -32.64669799804687, "median": 54.19172477722168, "p90": 183.70914306640626, "max": 207.50576782226562, "pos_frac": 0.765625, "sample": [47.716552734375, 135.28228759765625, 4.163688659667969, -81.02597045898438, 102.12252807617188, 149.6658935546875, -0.7303466796875, 102.32206726074219, 207.50576782226562, -5.425628662109375, -36.48102951049805, 36.108741760253906, 149.63330078125, 44.09161376953125, 190.63113403320312, 136.89382934570312, 35.752037048339844, 179.47470092773438, 14.055824279785156, 190.95169067382812, 105.19703674316406, 185.08006286621094, -7.62921142578125, -7.148113250732422, 157.5262908935547, 20.026166915893555, 146.589599609375, 57.30115509033203, -4.441432952880859, 69.87889862060547, 198.40118408203125, 3.796051025390625, 6.863372802734375, 18.855995178222656, 59.629295349121094, -94.06398010253906, 152.75802612304688, 2.1595077514648438, 180.5103302001953, -9.850967407226562, 3.4840011596679688, -51.568939208984375, 191.35720825195312, 123.42692565917969, 119.53450012207031, -119.64254760742188, 126.37828063964844, 5.294921875, 73.34396362304688, 72.22445678710938, -100.04529571533203, 191.08172607421875, -99.494384765625, -4.867149353027344, 1.2361335754394531, 51.08229446411133, 148.84368896484375, 96.42450714111328, 43.9627799987793, 152.9696807861328, 141.52951049804688, 29.08362579345703, -23.69992446899414, 143.87225341796875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000653.npy"}
{"epoch": 0.9871504157218443, "step": 654, "batch_size": 64, "mean": 43.905029296875, "std": 86.32135009765625, "min": -106.8981704711914, "p10": -57.78935241699219, "median": 17.09792709350586, "p90": 186.16097106933594, "max": 200.1537322998047, "pos_frac": 0.734375, "sample": [5.236732482910156, 200.14523315429688, -58.07110595703125, 126.14106750488281, -29.551483154296875, -47.2235107421875, 170.78033447265625, 28.55581283569336, 28.388832092285156, 109.52604675292969, -29.603286743164062, -100.1162109375, 152.58517456054688, 13.38067626953125, 5.176277160644531, 92.4020004272461, 13.036094665527344, -77.03246307373047, 188.08932495117188, 5.4992523193359375, 1.9097900390625, 3.3203277587890625, -67.05125427246094, 200.1537322998047, -2.1546707153320312, 8.687995910644531, -10.460826873779297, -12.934715270996094, -97.17027282714844, 162.91452026367188, 0.4520530700683594, -57.131927490234375, 197.91778564453125, 85.9248046875, 133.9112548828125, 165.77859497070312, 78.75102233886719, 42.847774505615234, 69.6176528930664, 21.656417846679688, 157.18212890625, 19.268402099609375, 15.124053955078125, -50.443511962890625, 45.01791000366211, -13.61627197265625, 186.7296142578125, 184.83413696289062, 75.64347839355469, 59.863014221191406, 32.823402404785156, 198.47393798828125, 2.5237770080566406, 11.520278930664062, 196.24716186523438, -106.8981704711914, 95.60739135742188, 3.6554298400878906, 9.29071044921875, 51.83684158325195, 19.071800231933594, -11.445953369140625, 1.772674560546875, -98.4451904296875], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000654.npy"}
{"epoch": 0.9886621315192744, "step": 655, "batch_size": 64, "mean": 55.55066680908203, "std": 78.08816528320312, "min": -76.21284484863281, "p10": -24.285083007812496, "median": 35.58900260925293, "p90": 180.16594085693362, "max": 208.9170379638672, "pos_frac": 0.75, "sample": [176.89566040039062, 9.528629302978516, -5.856605529785156, 2.2724838256835938, -2.5837783813476562, 48.47193145751953, 142.7758331298828, -25.402328491210938, -21.678176879882812, 7.2467498779296875, 0.9141998291015625, 96.69535827636719, 4.277435302734375, -14.766471862792969, -7.597251892089844, 2.1944732666015625, 35.46832275390625, 72.14968872070312, 108.2923355102539, 3.3336944580078125, 193.18392944335938, -3.1411285400390625, 54.58012390136719, 205.115478515625, 172.3952178955078, 0.6097679138183594, 92.15947723388672, -42.467369079589844, 181.56748962402344, -76.21284484863281, 133.5413818359375, 1.79522705078125, 144.0685272216797, 27.584632873535156, 55.618499755859375, 4.101776123046875, 11.399581909179688, 163.947021484375, 125.43824768066406, 208.9170379638672, -2.5591583251953125, -63.501914978027344, 29.522140502929688, 105.25929260253906, 35.70968246459961, -75.89961242675781, 95.81166076660156, 53.74186706542969, 132.1015625, 192.25770568847656, -9.867752075195312, 40.18370819091797, 196.45672607421875, 68.8216552734375, -3.545909881591797, 38.98910140991211, 185.22915649414062, 6.367073059082031, 91.15998840332031, -33.38763427734375, -56.503318786621094, 169.9031219482422, 39.26727294921875, 32.891876220703125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000655.npy"}
{"epoch": 0.9901738473167044, "step": 656, "batch_size": 64, "mean": 53.92597961425781, "std": 86.01506805419922, "min": -164.82102966308594, "p10": -33.48205528259277, "median": 44.672706604003906, "p90": 189.98748321533205, "max": 205.19287109375, "pos_frac": 0.78125, "sample": [69.20533752441406, 19.881088256835938, 11.0694580078125, 121.82855987548828, -99.89959716796875, -25.302757263183594, 8.853836059570312, 16.980690002441406, 42.01063537597656, -1.4115524291992188, 11.523612976074219, 190.7581787109375, -6.316322326660156, 19.229164123535156, -0.37683677673339844, 47.81968688964844, -161.89906311035156, -7.166206359863281, 145.04905700683594, 150.1597900390625, 195.30807495117188, 59.75972366333008, -164.82102966308594, 97.31086730957031, 174.64968872070312, -36.98746871948242, 58.86155700683594, 109.753662109375, 1.8512725830078125, 81.20620727539062, 188.18919372558594, 34.10429382324219, 76.10385131835938, 122.98831176757812, 31.279544830322266, 205.19287109375, 23.751419067382812, 7.235729217529297, 194.90623474121094, 34.877716064453125, -18.058547973632812, -90.34820556640625, 47.33477783203125, 195.83888244628906, 83.8395004272461, 196.07952880859375, 13.318389892578125, 123.37290954589844, 54.289161682128906, 116.86392211914062, 123.70899963378906, -55.46903610229492, -1.2381591796875, 194.35252380371094, 64.27864837646484, 4.4765625, 62.762664794921875, -101.74322509765625, 5.242706298828125, 18.44039535522461, 11.809089660644531, 89.149169921875, 126.52464294433594, 138.9189453125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000656.npy"}
{"epoch": 0.9916855631141346, "step": 657, "batch_size": 64, "mean": 73.339111328125, "std": 89.30073547363281, "min": -177.65370178222656, "p10": -13.960156249999999, "median": 57.5522575378418, "p90": 191.4303756713867, "max": 216.7664794921875, "pos_frac": 0.8125, "sample": [191.7195587158203, 10.1375732421875, 1.5233268737792969, 193.36712646484375, 141.35804748535156, 144.36114501953125, 211.36737060546875, 193.16409301757812, 82.5047607421875, -49.80760192871094, 0.33840179443359375, -33.6734733581543, 174.19317626953125, 39.568756103515625, 136.09603881835938, 122.2922134399414, 22.73486328125, 190.755615234375, -48.280616760253906, 180.87664794921875, -14.693695068359375, 37.903892517089844, -147.5467529296875, -14.784616470336914, 75.89802551269531, 146.44818115234375, 28.631195068359375, 38.48152160644531, 3.424114227294922, 7.8721160888671875, 22.840362548828125, 152.13296508789062, 103.93600463867188, 181.4661865234375, 67.04610443115234, 53.395843505859375, 74.31379699707031, 175.48355102539062, 185.63494873046875, 49.188720703125, 216.7664794921875, 3.954803466796875, 14.619539260864258, 167.23768615722656, 1.789276123046875, 18.645538330078125, 200.49819946289062, 9.719078063964844, -8.487953186035156, 47.80799865722656, -12.248565673828125, -177.65370178222656, 155.08221435546875, -1.257120132446289, 61.70867156982422, 139.35723876953125, -11.018173217773438, 144.55946350097656, 137.2708282470703, 194.95175170898438, -0.7865066528320312, 117.62945556640625, 1.4153900146484375, 140.471923828125], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000657.npy"}
{"epoch": 0.9931972789115646, "step": 658, "batch_size": 64, "mean": 56.699527740478516, "std": 97.45510864257812, "min": -197.3545684814453, "p10": -41.79279232025145, "median": 26.394962310791016, "p90": 182.63592834472655, "max": 212.47467041015625, "pos_frac": 0.71875, "sample": [109.69808959960938, 152.1626739501953, 182.17881774902344, -46.69915771484375, 101.15428161621094, 27.22069549560547, -162.744384765625, 38.4598503112793, 140.8419952392578, 160.67652893066406, -28.031997680664062, 116.95913696289062, -22.451553344726562, 124.10678100585938, 165.8704833984375, 54.3172492980957, -78.34243774414062, 31.131736755371094, 212.47467041015625, 24.660873413085938, -30.344606399536133, 25.569229125976562, -24.329010009765625, 146.31085205078125, 15.5245361328125, 177.22976684570312, -112.57971954345703, 113.72931671142578, 156.82041931152344, 4.627784729003906, 19.194438934326172, 2.3911399841308594, 0.273773193359375, 182.8318328857422, -0.0091552734375, 55.9653434753418, 161.46295166015625, -25.45269203186035, -16.276214599609375, 7.20379638671875, 193.96670532226562, 2.803255081176758, 15.099319458007812, 195.03709411621094, -4.504190444946289, 149.8043212890625, 16.992050170898438, 205.75283813476562, 177.21592712402344, 19.4364013671875, -72.2686538696289, -26.0447998046875, -20.62889862060547, -78.31443786621094, 124.31153869628906, 135.76199340820312, 0.03052520751953125, -197.3545684814453, 209.06350708007812, -4.771419525146484, 210.39649963378906, 49.11305236816406, 163.89169311523438, 0.19198226928710938], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000658.npy"}
{"epoch": 0.9947089947089947, "step": 659, "batch_size": 64, "mean": 29.31315803527832, "std": 91.34268188476562, "min": -179.4541015625, "p10": -69.44087028503418, "median": 18.560646057128906, "p90": 176.2874923706055, "max": 205.6896209716797, "pos_frac": 0.703125, "sample": [74.49166107177734, 90.55622100830078, 192.15933227539062, 85.55925750732422, -58.27522659301758, 205.6896209716797, 6.801353454589844, 24.40907096862793, 26.723102569580078, 78.69667053222656, 117.70657348632812, -2.745574951171875, 2.184581756591797, -23.610748291015625, 10.777427673339844, 10.639476776123047, 121.13180541992188, -151.76675415039062, 2.235279083251953, -62.23163986206055, 171.65899658203125, -53.02116775512695, -5.9311065673828125, 18.552146911621094, 35.98210906982422, -22.30316162109375, 9.016250610351562, 5.9770097732543945, 2.518199920654297, 11.00567626953125, 22.300308227539062, -16.173553466796875, 197.38870239257812, 46.8583984375, -1.3908920288085938, 6.9283294677734375, 198.84494018554688, 105.4511947631836, 183.7413787841797, 18.56914520263672, 39.124603271484375, -179.4541015625, -116.88922119140625, 163.78701782226562, -72.5305404663086, 20.89865493774414, 5.980434417724609, 33.55731201171875, 165.11468505859375, -44.437339782714844, 25.554351806640625, 185.7891387939453, -170.8401336669922, 26.07318115234375, 49.83265686035156, 102.529541015625, -11.448959350585938, -164.9505615234375, 178.27113342285156, 12.692474365234375, -20.572566986083984, 39.26079559326172, 21.587158203125, -99.991943359375], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000659.npy"}
{"epoch": 0.9962207105064248, "step": 660, "batch_size": 64, "mean": 62.558837890625, "std": 79.1045150756836, "min": -102.30238342285156, "p10": -14.744217109680175, "median": 31.133195877075195, "p90": 171.70264892578126, "max": 221.872802734375, "pos_frac": 0.765625, "sample": [171.87066650390625, 137.86843872070312, 88.0414047241211, -11.551963806152344, 94.745849609375, -102.30238342285156, -15.82745361328125, 163.70960998535156, -99.2295150756836, -47.839332580566406, 45.82838439941406, 140.74945068359375, 26.117389678955078, 171.05160522460938, 30.37448501586914, 148.67491149902344, 194.43658447265625, 154.35470581054688, 16.068466186523438, -15.036781311035156, 77.41446685791016, 18.257965087890625, 20.315284729003906, 24.4443359375, -37.37371826171875, 173.76370239257812, -8.835777282714844, 165.04278564453125, 30.89710235595703, -4.767486572265625, 16.161178588867188, 30.579544067382812, 221.872802734375, 72.73646545410156, -1.5494155883789062, -55.79072570800781, 4.627250671386719, 49.85327911376953, 125.46002197265625, 6.780479431152344, -1.6317596435546875, 2.537111282348633, 69.01666259765625, 28.49164581298828, 161.8482666015625, 43.48717498779297, 126.68463897705078, 1.9290771484375, 63.316505432128906, 179.6097412109375, 138.27813720703125, 182.6769256591797, -10.598098754882812, 25.669700622558594, 166.718994140625, -0.31522369384765625, 2.454233169555664, 122.52024841308594, 178.4359893798828, -14.061567306518555, 99.85066986083984, 31.36928939819336, 12.172531127929688, 171.31060791015625], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000660.npy"}
{"epoch": 0.9977324263038548, "step": 661, "batch_size": 64, "mean": 36.04889678955078, "std": 93.91353607177734, "min": -194.880859375, "p10": -73.9889450073242, "median": 11.668575286865234, "p90": 175.5260177612305, "max": 230.05621337890625, "pos_frac": 0.71875, "sample": [-21.736953735351562, -5.904731750488281, -110.26528930664062, -115.44480895996094, 89.90462493896484, 108.92073059082031, 176.91612243652344, -1.62054443359375, -194.880859375, 105.57139587402344, -56.46892547607422, -15.94180679321289, -9.79315185546875, 188.79318237304688, 194.73226928710938, 154.11404418945312, 196.34445190429688, 130.6417236328125, 230.05621337890625, -77.78932189941406, -183.44107055664062, 7.0156402587890625, 180.0023651123047, 21.917892456054688, 2.199604034423828, 11.924476623535156, -7.784706115722656, -46.435874938964844, 11.412673950195312, 5.2384490966796875, 23.477310180664062, 20.936622619628906, -143.0518035888672, 53.159637451171875, 109.13430786132812, 2.0068588256835938, 32.00425720214844, 85.06829833984375, 1.04693603515625, 125.96084594726562, 90.35284423828125, 0.7556571960449219, -31.890625, 121.03599548339844, 119.63577270507812, 157.3637237548828, 184.32957458496094, 15.186363220214844, 10.137561798095703, 79.4168930053711, -101.83848571777344, 28.28418731689453, -65.12139892578125, -11.407752990722656, 13.152076721191406, 138.98760986328125, 1.6456680297851562, 2.44122314453125, 5.40264892578125, 89.90042114257812, 4.4346771240234375, 0.725006103515625, 3.97637939453125, 172.28244018554688], "npy": "outputs/qwen3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4/margin_logs/step_0000661.npy"}