Files
Qwen3-4B-Chemistry-RLRT-TR/artifacts/wandb-summary.json

1 line
12 KiB
JSON
Raw Permalink Normal View History

{"val-aux/sciknoweval/reward/worst@16/mean":0.2696952380952381,"val-aux/sciknoweval/score/best@8/mean":0.8311666666666666,"val-aux/sciknoweval/score/maj@8/std":0.1528837415995549,"val-aux/sciknoweval/incorrect_format/worst@16/std":0.04613710025134583,"val-aux/sciknoweval/acc/worst@4/mean":0.39438095238095233,"val-aux/sciknoweval/acc/best@8/mean":0.8311666666666666,"perf/max_memory_reserved_gb":137.505859375,"critic/score/min":0,"self_distillation/empty_target_batch":0.23046875,"val-aux/sciknoweval/incorrect_format/best@2/std":0.012277432195181853,"training/global_step":100,"training/rollout_probs_diff_valid":1,"val-aux/sciknoweval/acc/worst@4/std":0.182545872924148,"val-aux/sciknoweval/acc/best@2/mean":0.6968857142857143,"global_seqlen/mean":85677.25,"critic/returns/mean":-0.025867901742458344,"self_distillation/token_reweight_is_rlrt":1,"val-aux/sciknoweval/reward/worst@16/std":0.08420508892562015,"val-aux/sciknoweval/score/maj@16/std":0.11089304329120943,"timing_s/save_checkpoint":6.956284845247865,"rollout_corr/rollout_is_eff_sample_size":0.9949946665841123,"val-aux/sciknoweval/score/worst@4/mean":0.39438095238095233,"response_length_non_aborted/max":8192,"rollout_corr/rollout_is_std":0.07089795172214508,"num_turns/mean":2,"val-aux/sciknoweval/incorrect_format/maj@2/std":0.03180376195882722,"val-aux/sciknoweval/score/best@4/std":0.1424436179078602,"timing_per_token_ms/update_actor":0.0411384974974895,"val-aux/sciknoweval/incorrect_format/std@16":0.03187742018839054,"val-aux/sciknoweval/acc/best@4/mean":0.7765571428571428,"val-aux/sciknoweval/reward/best@2/std":0.20627882758717012,"val-aux/sciknoweval/incorrect_format/worst@8/mean":0.9419428571428571,"val-aux/sciknoweval/reward/maj@16/std":0.11089304329120943,"val-aux/sciknoweval/reward/best@4/mean":0.7765571428571428,"training/rollout_probs_diff_std":0.03298613056540489,"val-core/sciknoweval/acc/mean@16":0.5916666666666667,"prompt_length/clip_ratio":0,"_wandb":{"runtime":8084},"val-aux/sciknoweval/reward/maj@2/mean":0.5911,"self_distillation/success_group_fraction":0.78125,"self_distillation/token_reweight_context_sample_fraction":0.76953125,"val-aux/sciknoweval/score/best@2/mean":0.6968857142857143,"perf/total_num_tokens":685418,"self_distillation/token_reweight_correct_sample_fraction":0.58203125,"response_length_non_aborted/mean":2381.2578125,"rollout_corr/rollout_is_seq_std":0.0032133334316313267,"val-aux/sciknoweval/score/worst@2/std":0.23108433720880842,"val-aux/sciknoweval/reward/worst@2/std":0.23108433720880842,"critic/returns/max":0.875,"timing_s/start_profile":8.443742990493774e-05,"rollout_corr/rollout_is_seq_fraction_low":0,"training/rollout_probs_diff_mean":0.0073742191307246685,"rollout_corr/rollout_is_seq_min":0.991420328617096,"critic/returns/min":-0.875,"val-aux/num_turns/min":2,"val-aux/sciknoweval/score/best@4/mean":0.7765571428571428,"val-aux/sciknoweval/acc/maj@2/std":0.25693213561174044,"perf/throughput":779.2287777281216,"self_distillation/token_reweight_lambda":0.5,"rollout_corr/rollout_ppl":1.0910621960647404,"val-aux/sciknoweval/reward/best@2/mean":0.6968857142857143,"timing_s/agent_loop/generate_sequences/max":69.66160243563354,"val-aux/sciknoweval/acc/maj@8/mean":0.6276761904761905,"rollout_corr/log_ppl_abs_diff":0.006301159230133635,"val-aux/sciknoweval/acc/worst@16/std":0.08420508892562015,"val-aux/sciknoweval/acc/worst@8/mean":0.32224761904761906,"rollout_corr/training_log_ppl":0.0930957942691748,"val-aux/sciknoweval/score/worst@8/std":0.13204132054783738,"val-aux/sciknoweval/reward/best@16/mean":0.8632047619047617,"timing_s/step":109.9513422101736,"val-core/sciknoweval/acc/maj@16/mean":0.6312809523809525,"rollout_corr/rollout_is_min":8.033764970605262e-06,"self_distillation/token_reweight_w_std":1.7234428022899628e+06,"training/rollout_actor_probs_pearson_corr":0.9742870330810547,"val-aux/sciknoweval/incorrect_format/maj@16/mean":0.9972333333333334,"critic/advantages/mean":-0.025867901742458344,"val-aux/num_turns/mean":2,"rollout_corr/log_ppl_diff":0.0062478416239173384,"val-aux/sciknoweval/acc/worst@8/std":0.13204132054783738,"critic/