Files
Qwen3-8B-Chemistry-SDPO-TR/artifacts/wandb-summary.json

1 line
11 KiB
JSON
Raw Normal View History

{"global_seqlen/balanced_min":20404,"perf/mfu/actor":0.05015731000074418,"rollout_corr/rollout_is_max":1.4396264553070068,"critic/returns/max":0.875,"val-aux/sciknoweval/score/worst@4/std":0.06781975137735766,"val-aux/sciknoweval/score/worst@16/std":0.02453756899133995,"critic/rewards/max":1,"_wandb":{"runtime":5598},"critic/returns/min":-0.875,"timing_s/agent_loop/tool_calls/mean":0,"prompt_length/clip_ratio":0,"val-core/sciknoweval/acc/best@16/std":0.02643611186706914,"val-aux/sciknoweval/reward/worst@4/std":0.06781975137735766,"val-aux/sciknoweval/acc/maj@8/std":0.07759319570379096,"rollout_corr/rollout_is_seq_max":1.005556344985962,"val-aux/sciknoweval/score/maj@16/mean":0.7659095238095238,"val-aux/sciknoweval/incorrect_format/best@2/std":0.0005486204557414609,"self_distillation/correct_sample_fraction":0.74609375,"val-aux/sciknoweval/acc/worst@4/mean":0.6645857142857142,"val-aux/sciknoweval/incorrect_format/worst@16/std":0.004610765367099621,"rollout_corr/rollout_is_mean":1.0000677108764648,"val-aux/num_turns/max":2,"timing_s/agent_loop/generate_sequences/min":1.463550142943859,"critic/returns/mean":-0.1100277230143547,"rollout_corr/ppl_ratio":1.0003064163029194,"timing_s/agent_loop/generate_sequences/max":59.862675316631794,"rollout_corr/kl":0.0003049511138506489,"val-aux/sciknoweval/incorrect_format/maj@4/std":0.0015374570050248517,"val-aux/sciknoweval/acc/maj@8/mean":0.7632142857142856,"response_length_non_aborted/mean":426.9609375,"perf/mfu/actor_infer":0,"response_length/min":111,"rollout_corr/rollout_is_seq_min":0.9953861832618713,"critic/score/max":1,"val-aux/sciknoweval/reward/best@4/std":0.08027809906180897,"val-aux/sciknoweval/incorrect_format/worst@8/std":0.004640859533041082,"training/rollout_actor_probs_pearson_corr":0.9996629357337952,"timing_s/update_actor":24.305801395326853,"val-aux/sciknoweval/incorrect_format/maj@16/std":0.0002604287412166727,"global_seqlen/mean":23139.75,"val-aux/sciknoweval/score/worst@8/std":0.04212482341518254,"val-aux/sciknoweval/acc/worst@8/std":0.04212482341518254,"prompt_length/max":625,"val-aux/sciknoweval/reward/maj@8/mean":0.7632142857142856,"val-aux/sciknoweval/score/best@8/std":0.05177293853368067,"val-aux/sciknoweval/score/worst@8/mean":0.6388952380952381,"self_distillation/success_sample_fraction":0.93359375,"perf/cpu_memory_used_gb":198.26213359832764,"response_length/clip_ratio":0.01171875,"val-aux/sciknoweval/reward/maj@2/mean":0.7537380952380953,"training/rollout_probs_diff_mean":0.002571779303252697,"rollout_corr/rollout_is_ratio_fraction_high":0,"rollout_corr/training_log_ppl":0.3963073937320587,"val-aux/sciknoweval/incorrect_format/maj@8/mean":0.9999476190476191,"response_length_non_aborted/max":8192,"val-aux/sciknoweval/reward/best@4/mean":0.8485619047619049,"val-aux/sciknoweval/reward/worst@4/mean":0.6645857142857142,"val-aux/sciknoweval/incorrect_format/maj@4/mean":0.9997333333333334,"val-aux/sciknoweval/reward/best@8/std":0.05177293853368067,"val-aux/sciknoweval/acc/best@2/std":0.10681452617322888,"rollout_corr/rollout_is_seq_fraction_low":0,"timing_per_token_ms/gen":0.6090940510229154,"self_distillation/success_group_fraction":0.9375,"timing_s/testing":128.6522780675441,"rollout_corr/rollout_is_min":0.7055256962776184,"training/global_step":100,"response_length/max":8192,"val-aux/sciknoweval/acc/best@8/std":0.05177293853368067,"val-aux/sciknoweval/reward/maj@4/std":0.10127914193936502,"val-aux/sciknoweval/incorrect_format/best@4/std":0,"val-aux/sciknoweval/incorrect_format/worst@2/std":0.0030314321460128213,"num_turns/mean":2,"val-aux/sciknoweval/reward/maj@2/std":0.12035272201748512,"val-aux/sciknoweval/incorrect_format/worst@16/mean":0.994052380952381,"rollout_corr/rollout_is_seq_max_deviation":0.005556344985961914,"training/rollout_probs_diff_max":0.12101149559020996,"self_distillation/dynamic_weighting_enabled":0,"timing_s/agent_loop/slowest/tool_calls":0,"val-aux/sciknoweval/score/maj@4/mean":0.7588809523809524,"val-aux/sciknoweval/incorrect_format/worst@4/mean":0.9978380952380952,"rollout_corr/rollout_is_seq_mean":1.0000479221343994,"global_seqlen