Files
Qwen3-4B-Chemical-RLSD-TR/artifacts/wandb-summary.json

1 line
11 KiB
JSON
Raw Permalink Normal View History

{"self_distillation/token_reweight_context_sample_fraction":0.71875,"actor/kl_loss":0,"val-aux/sciknoweval/incorrect_format/maj@2/mean":0.9997,"training/global_step":100,"prompt_length/min":165,"num_turns/max":2,"val-aux/sciknoweval/reward/maj@2/std":0.10956288078402857,"rollout_corr/chi2_seq":1.0585334873758256,"response_length/max":616,"val-aux/sciknoweval/incorrect_format/worst@4/std":0.0020220104047945714,"rollout_corr/rollout_is_seq_std":0.002407139865681529,"global_seqlen/balanced_max":19622,"val-aux/sciknoweval/score/std@16":0.10973085235155812,"val-aux/sciknoweval/score/maj@4/mean":0.6991761904761905,"rollout_corr/rollout_ppl":1.2673596912063658,"val-aux/sciknoweval/reward/best@4/std":0.05746985351073585,"rollout_corr/kl":0.0009287315268036167,"val-aux/sciknoweval/reward/std@16":0.10973085235155812,"self_distillation/token_reweight_w_std":1.1881619467263469e+06,"rollout_corr/rollout_is_seq_fraction_low":0,"timing_s/agent_loop/generate_sequences/max":5.5460103284567595,"val-aux/sciknoweval/acc/best@2/std":0.08320072408312382,"self_distillation/token_reweight_active_token_fraction":0.71875,"val-aux/sciknoweval/reward/maj@8/std":0.05779624954713381,"val-aux/sciknoweval/score/worst@2/mean":0.6464904761904762,"critic/rewards/mean":0.6171875,"val-aux/num_turns/max":2,"val-aux/sciknoweval/incorrect_format/best@16/mean":1,"val-aux/sciknoweval/score/maj@2/mean":0.6907285714285715,"val-aux/sciknoweval/incorrect_format/best@4/mean":1,"self_distillation/success_sample_fraction":0.71875,"response_length_non_aborted/max":616,"perf/cpu_memory_used_gb":146.29161834716797,"response_length_non_aborted/mean":317.0234375,"training/rollout_probs_diff_mean":0.004920980893075466,"timing_s/adv":0.4910304043442011,"val-core/sciknoweval/acc/mean@16":0.6895833333333333,"val-aux/sciknoweval/acc/maj@4/mean":0.6991761904761905,"val-aux/sciknoweval/acc/maj@2/std":0.10956288078402857,"val-aux/sciknoweval/incorrect_format/maj@4/mean":0.9999476190476191,"val-aux/sciknoweval/incorrect_format/worst@16/mean":0.9967095238095238,"critic/advantages/mean":-0.0011474530911073089,"val-aux/sciknoweval/incorrect_format/maj@4/std":0.0004966783087710074,"val-core/sciknoweval/acc/best@16/mean":0.8013714285714286,"val-aux/sciknoweval/acc/worst@2/std":0.10098054289944847,"val-aux/sciknoweval/score/worst@16/mean":0.5435285714285715,"val-aux/sciknoweval/acc/worst@2/mean":0.6464904761904762,"val-aux/sciknoweval/incorrect_format/best@16/std":0,"val-aux/num_turns/mean":2,"rollout_corr/log_ppl_diff":0.0009287331777159125,"val-aux/sciknoweval/incorrect_format/maj@2/std":0.0011569664768572288,"val-aux/sciknoweval/acc/maj@8/mean":0.7047571428571429,"val-aux/sciknoweval/reward/maj@16/std":0.03834834240388913,"prompt_length/mean":296.15625,"val-aux/sciknoweval/incorrect_format/best@2/std":0.0002127457895589398,"val-aux/sciknoweval/reward/best@8/mean":0.7858285714285714,"val-aux/sciknoweval/incorrect_format/maj@16/mean":1,"val-aux/sciknoweval/score/maj@8/mean":0.7047571428571429,"critic/advantages/max":0.75,"val-aux/sciknoweval/incorrect_format/worst@2/mean":0.9994476190476191,"val-aux/sciknoweval/score/best@8/mean":0.7858285714285714,"val-aux/sciknoweval/incorrect_format/worst@8/mean":0.9980238095238095,"training/epoch":1,"training/rollout_probs_diff_max":0.30163663625717163,"val-aux/sciknoweval/acc/worst@4/std":0.08537783027231041,"critic/score/mean":0.6171875,"val-aux/sciknoweval/reward/worst@2/std":0.10098054289944847,"val-aux/sciknoweval/score/worst@8/std":0.06319778055664713,"rollout_corr/rollout_is_seq_mean":1.0000463724136353,"val-aux/sciknoweval/score/best@16/std":0.02391792727193744,"timing_s/update_actor":15.488128297030926,"val-aux/sciknoweval/acc/maj@4/std":0.08403267534041585,"rollout_corr/training_log_ppl":0.23310650100756902,"val-aux/sciknoweval/acc/best@4/mean":0.7644238095238095,"rollout_corr/rollout_is_ratio_fraction_high":1.2321644135226961e-05,"val-aux/sciknoweval/reward/maj@8/mean":0.7047571428571429,"val-aux/sciknoweval/acc/worst@8/mean":0.5691952380952381,"num_turns/mean":2,"val-aux/sciknoweval/reward/best@8/std":0.038120394142420314,"val-aux/scikno