Files
Qwen3-8B-Biology-GRPO-TR/artifacts/wandb-summary.json

1 line
10 KiB
JSON
Raw Normal View History

{"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/