Files
Qwen3-8B-Biology-GRPO-TR/artifacts/wandb-summary.json
ModelHub XC 67b8171dec 初始化项目,由ModelHub XC社区提供模型
Model: SeongryongJung/Qwen3-8B-Biology-GRPO-TR
Source: Original Platform
2026-08-11 09:55:18 +08:00

1 line
10 KiB
JSON

{"rollout_corr/log_ppl_diff_max":0.007616400718688965,"response_length/min":235,"val-aux/sciknoweval/reward/worst@16/std":0.05929977285001639,"rollout_corr/rollout_is_std":0.0328901931643486,"timing_s/agent_loop/slowest/prompt_length":431,"perf/throughput":514.4348805748427,"val-aux/sciknoweval/reward/best@4/mean":0.6779,"critic/score/min":0,"val-aux/sciknoweval/reward/best@2/std":0.2054002064937382,"val-aux/sciknoweval/incorrect_format/maj@4/std":0.06365340394879623,"actor/pg_clipfrac":0.00012224168438024208,"val-aux/sciknoweval/score/maj@2/mean":0.5014,"critic/advantages/max":0.875,"val-aux/sciknoweval/incorrect_format/worst@8/std":0.13449797212902684,"val-aux/sciknoweval/score/maj@4/mean":0.52822,"val-aux/sciknoweval/score/maj@16/mean":0.5538599999999999,"val-aux/sciknoweval/reward/best@16/std":0.07256344316810731,"timing_s/update_actor":14.441238490864635,"training/rollout_probs_diff_valid":1,"val-core/sciknoweval/acc/best@16/mean":0.7777600000000001,"rollout_corr/rollout_is_seq_mean":0.9997690320014954,"val-aux/sciknoweval/acc/best@2/std":0.2054002064937382,"val-aux/sciknoweval/acc/best@8/mean":0.73414,"timing_s/agent_loop/generate_sequences/min":3.021278627216816,"val-aux/sciknoweval/incorrect_format/worst@8/mean":0.8033600000000001,"val-aux/sciknoweval/incorrect_format/maj@2/std":0.10303116829198576,"timing_s/adv":0.12742269970476627,"critic/advantages/min":-0.875,"timing_s/agent_loop/tool_calls/min":0,"val-aux/sciknoweval/score/mean@16":0.49875,"val-aux/sciknoweval/acc/maj@2/std":0.25438537975653214,"val-aux/sciknoweval/reward/best@8/mean":0.73414,"timing_s/reward":0.0772107895463705,"training/rollout_probs_diff_std":0.008115967735648155,"rollout_corr/k3_kl":0.0006231129761431475,"global_seqlen/balanced_max":37603,"val-aux/sciknoweval/incorrect_format/best@16/mean":1,"val-aux/sciknoweval/acc/worst@2/mean":0.39715999999999996,"response_length_non_aborted/max":6262,"val-aux/sciknoweval/reward/worst@8/std":0.11648076304197817,"val-aux/sciknoweval/incorrect_format/worst@2/mean":0.91962,"val-aux/sciknoweval/incorrect_format/best@8/mean":0.99992,"val-aux/sciknoweval/acc/worst@4/std":0.1768647807422225,"perf/mfu/actor":0.12333177659241132,"rollout_corr/rollout_is_eff_sample_size":0.9989191068729061,"timing_s/save_checkpoint":12.46778230369091,"val-aux/sciknoweval/acc/std@16":0.2541486398580183,"val-aux/sciknoweval/acc/worst@4/mean":0.303,"response_length_non_aborted/clip_ratio":0,"val-aux/sciknoweval/reward/maj@16/std":0.0756443320805554,"val-aux/sciknoweval/score/maj@2/std":0.25438537975653214,"num_turns/min":2,"val-aux/sciknoweval/score/maj@4/std":0.19553605651535125,"actor/grad_norm":0.2531181573867798,"val-aux/num_turns/mean":2,"timing_s/agent_loop/tool_calls/mean":0,"rollout_corr/rollout_ppl":2.0389733486808836,"val-aux/sciknoweval/acc/maj@2/mean":0.5014,"response_length_non_aborted/mean":713.5625,"val-aux/sciknoweval/incorrect_format/worst@2/std":0.12679376175372256,"prompt_length/mean":349.90625,"rollout_corr/rollout_is_seq_max":1.0050616264343262,"val-aux/sciknoweval/incorrect_format/best@4/std":0.009215556459094006,"rollout_corr/rollout_is_seq_fraction_low":0,"val-aux/sciknoweval/incorrect_format/worst@16/std":0.1034399395499081,"val-aux/sciknoweval/score/maj@8/std":0.1254995217888188,"perf/time_per_step":66.15220173634589,"val-aux/sciknoweval/acc/maj@4/mean":0.52822,"val-aux/sciknoweval/score/worst@4/std":0.1768647807422225,"perf/max_memory_reserved_gb":144.787109375,"val-aux/sciknoweval/incorrect_format/best@4/mean":0.9988800000000002,"perf/mfu/actor_infer":0,"val-aux/sciknoweval/reward/maj@8/mean":0.5476200000000001,"training/rollout_probs_diff_mean":0.003967562224715948,"_step":100,"critic/rewards/mean":0.52734375,"critic/returns/max":0.875,"val-aux/sciknoweval/reward/maj@16/mean":0.5538599999999999,"val-aux/sciknoweval/score/best@16/std":0.07256344316810731,"timing_per_token_ms/adv":0.0004680390662365427,"rollout_corr/ppl_ratio":1.0009363645222038,"val-aux/sciknoweval/score/best@8/mean":0.73414,"val-aux/sciknoweval/acc/maj@4/std":0.19553605651535125,"global_seqlen/balanced_min":33498,"val-aux/sciknoweval/incorrect_format/worst@4/std":0.1409918624624078,"val-aux/sciknoweval/score/worst@2/mean":0.39715999999999996,"actor/kl_loss":0,"perf/total_num_tokens":272248,"rollout_corr/rollout_is_seq_max_deviation":0.00594097375869751,"val-aux/sciknoweval/incorrect_format/maj@16/std":0.006510218388623519,"training/rollout_probs_diff_max":0.27054858207702637,"val-aux/sciknoweval/acc/best@4/mean":0.6779,"timing_s/step":66.15220173634589,"val-aux/sciknoweval/reward/maj@4/std":0.19553605651535125,"prompt_length/clip_ratio":0,"rollout_corr/chi2_token":0.000629746587947011,"perf/cpu_memory_used_gb":104.6416015625,"critic/rewards/min":0,"rollout_corr/log_ppl_abs_diff":0.0013740459107793868,"response_length/clip_ratio":0,"rollout_corr/rollout_is_mean":0.999800443649292,"val-aux/sciknoweval/score/worst@8/mean":0.2291,"_wandb":{"runtime":6317},"timing_s/gen":47.96704002469778,"critic/rewards/max":1,"timing_s/start_profile":0.00017099827527999878,"timing_s/testing":115.95291896350682,"val-aux/sciknoweval/incorrect_format/mean@16":0.95375,"val-aux/sciknoweval/reward/best@4/std":0.14785794921891468,"training/rollout_actor_probs_pearson_corr":0.9996527433395386,"response_length_non_aborted/min":235,"response/aborted_ratio":0,"response_length/mean":713.5625,"val-aux/sciknoweval/incorrect_format/maj@16/mean":0.99924,"val-aux/sciknoweval/score/best@16/mean":0.7777600000000001,"response_length/max":6262,"val-aux/sciknoweval/incorrect_format/best@16/std":0,"global_seqlen/max":43400,"val-aux/num_turns/max":2,"val-aux/sciknoweval/reward/worst@2/mean":0.39715999999999996,"num_turns/mean":2,"val-aux/sciknoweval/score/maj@16/std":0.0756443320805554,"timing_s/agent_loop/slowest/response_length":6262,"val-aux/sciknoweval/acc/maj@8/mean":0.5476200000000001,"val-aux/sciknoweval/score/best@8/std":0.10570202036547473,"val-aux/sciknoweval/score/worst@8/std":0.11648076304197817,"val-aux/sciknoweval/acc/worst@2/std":0.22536220172213436,"val-aux/sciknoweval/score/worst@16/std":0.05929977285001639,"val-core/sciknoweval/acc/mean@16":0.49875,"val-core/sciknoweval/acc/maj@16/mean":0.5538599999999999,"timing_s/stop_profile":3.900192677974701e-05,"val-aux/sciknoweval/incorrect_format/std@16":0.10381127722390864,"val-aux/sciknoweval/acc/worst@8/std":0.11648076304197817,"val-aux/sciknoweval/incorrect_format/worst@4/mean":0.8675400000000001,"val-aux/sciknoweval/reward/worst@4/std":0.1768647807422225,"val-aux/sciknoweval/reward/std@16":0.2541486398580183,"val-aux/sciknoweval/reward/best@2/mean":0.60202,"val-aux/sciknoweval/incorrect_format/best@2/std":0.04408740320835782,"prompt_length/min":180,"val-aux/sciknoweval/score/worst@4/mean":0.303,"training/global_step":100,"val-core/sciknoweval/acc/maj@16/std":0.0756443320805554,"timing_s/old_log_prob":3.4488426838070154,"val-aux/sciknoweval/incorrect_format/maj@2/mean":0.95504,"val-aux/sciknoweval/reward/worst@8/mean":0.2291,"val-aux/sciknoweval/acc/best@4/std":0.14785794921891468,"val-aux/sciknoweval/acc/best@8/std":0.10570202036547473,"actor/pg_clipfrac_lower":0,"val-aux/sciknoweval/reward/maj@4/mean":0.52822,"val-aux/sciknoweval/incorrect_format/worst@16/mean":0.74356,"val-aux/sciknoweval/reward/maj@2/std":0.25438537975653214,"val-aux/sciknoweval/reward/best@16/mean":0.7777600000000001,"val-aux/sciknoweval/incorrect_format/maj@8/std":0.0280523469618475,"val-aux/sciknoweval/reward/mean@16":0.49875,"rollout_corr/log_ppl_diff":0.0009347237064503133,"val-aux/sciknoweval/reward/worst@2/std":0.22536220172213436,"global_seqlen/mean":34031,"val-core/sciknoweval/acc/best@16/std":0.07256344316810731,"perf/max_memory_allocated_gb":132.03091621398926,"val-aux/sciknoweval/reward/best@8/std":0.10570202036547473,"global_seqlen/min":25660,"rollout_corr/rollout_is_max":1.875651478767395,"val-aux/sciknoweval/acc/worst@8/mean":0.2291,"val-aux/sciknoweval/reward/worst@4/mean":0.303,"timing_s/agent_loop/generate_sequences/max":40.42630261927843,"critic/score/mean":0.52734375,"timing_s/agent_loop/slowest/generate_sequences":40.42630261927843,"val-aux/sciknoweval/incorrect_format/maj@8/mean":0.9949799999999999,"rollout_corr/rollout_is_ratio_fraction_high":0,"rollout_corr/rollout_is_seq_fraction_high":0,"actor/ppo_kl":0.00011320239463330495,"val-aux/sciknoweval/reward/worst@16/mean":0.18696,"val-aux/sciknoweval/acc/worst@16/mean":0.18696,"actor/pg_loss":0.00012731854803860188,"global_seqlen/minmax_diff":17740,"timing_s/agent_loop/tool_calls/max":0,"rollout_corr/rollout_is_min":0.3743396997451782,"rollout_corr/rollout_is_seq_std":0.001492936396971345,"val-aux/sciknoweval/score/worst@2/std":0.22536220172213436,"_runtime":6317.872940246,"rollout_corr/training_ppl":2.040964142885059,"timing_s/agent_loop/slowest/tool_calls":0,"actor/entropy":0.6494476199150085,"timing_per_token_ms/update_actor":0.05304442453522022,"val-aux/sciknoweval/acc/best@2/mean":0.60202,"prompt_length/max":499,"val-aux/sciknoweval/acc/maj@8/std":0.1254995217888188,"num_turns/max":2,"val-aux/sciknoweval/score/best@4/mean":0.6779,"val-aux/sciknoweval/reward/maj@8/std":0.1254995217888188,"val-aux/sciknoweval/score/std@16":0.2541486398580183,"val-aux/sciknoweval/reward/maj@2/mean":0.5014,"val-aux/sciknoweval/incorrect_format/best@2/mean":0.9902200000000001,"val-aux/sciknoweval/score/best@4/std":0.14785794921891468,"actor/lr":1e-06,"training/epoch":7,"critic/returns/min":-0.875,"rollout_corr/rollout_log_ppl":0.6952399488363881,"critic/score/max":1,"_timestamp":1.7830701084260151e+09,"val-aux/sciknoweval/score/best@2/mean":0.60202,"val-aux/sciknoweval/incorrect_format/maj@4/mean":0.98158,"val-aux/sciknoweval/acc/worst@16/std":0.05929977285001639,"rollout_corr/rollout_is_seq_min":0.9940590262413025,"timing_s/agent_loop/generate_sequences/mean":8.260560911876382,"rollout_corr/training_log_ppl":0.6961746725428384,"val-aux/sciknoweval/score/worst@16/mean":0.18696,"rollout_corr/rollout_is_ratio_fraction_low":5.474292720464291e-06,"rollout_corr/chi2_seq":0.7056945031508803,"rollout_corr/log_ppl_diff_min":-0.0043163299560546875,"val-aux/sciknoweval/incorrect_format/best@8/std":0.0012623787070447602,"timing_per_token_ms/gen":0.2625856180733653,"rollout_corr/kl":0.0009347146087534952,"val-aux/sciknoweval/score/best@2/std":0.2054002064937382,"val-aux/sciknoweval/score/maj@8/mean":0.5476200000000001,"critic/advantages/mean":-0.02030278369784355,"val-aux/num_turns/min":2,"critic/returns/mean":-0.02030278369784355}