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

1 line
11 KiB
JSON
Raw Normal View History

{"_runtime":5136.926281788,"val-aux/sciknoweval/acc/std@16":0.16309517319018776,"val-aux/sciknoweval/reward/worst@4/mean":0.6136428571428572,"val-aux/sciknoweval/acc/worst@8/mean":0.5821809523809525,"critic/rewards/min":0,"val-aux/sciknoweval/acc/worst@8/std":0.04669533954140184,"val-aux/sciknoweval/incorrect_format/best@2/std":0,"val-aux/sciknoweval/incorrect_format/maj@2/std":0,"rollout_corr/rollout_is_seq_max":1.0135035514831543,"val-aux/sciknoweval/reward/worst@16/mean":0.5673476190476191,"val-aux/sciknoweval/score/maj@2/std":0.16273431336078428,"val-aux/sciknoweval/incorrect_format/best@2/mean":1,"timing_s/start_profile":7.35800713300705e-05,"val-aux/sciknoweval/acc/worst@16/mean":0.5673476190476191,"global_seqlen/max":18719,"timing_s/agent_loop/slowest/tool_calls":0,"perf/mfu/actor_infer":0,"val-aux/sciknoweval/score/best@8/mean":0.9082190476190476,"timing_s/agent_loop/generate_sequences/max":6.079234275966883,"val-aux/sciknoweval/acc/maj@2/std":0.16273431336078428,"val-aux/sciknoweval/score/best@16/mean":0.9349142857142857,"self_distillation/dynamic_weighting_enabled":1,"self_distillation/success_group_fraction":0.9375,"rollout_corr/log_ppl_abs_diff":0.001808215893106535,"rollout_corr/rollout_is_mean":1.0000580549240112,"response_length/min":101,"rollout_corr/rollout_is_seq_mean":1.0000791549682617,"val-aux/num_turns/max":2,"training/rollout_probs_diff_valid":1,"num_turns/mean":2,"response_length_non_aborted/clip_ratio":0,"timing_s/gen":12.888554267585278,"rollout_corr/rollout_is_eff_sample_size":0.9991349319207025,"rollout_corr/chi2_token":0.001111235935240984,"val-aux/sciknoweval/score/maj@16/std":0.08119347263943295,"rollout_corr/kl":0.00039874245554472054,"val-core/sciknoweval/acc/maj@16/std":0.08119347263943295,"val-aux/sciknoweval/score/best@4/mean":0.8652047619047618,"srpo/routed_token_count":220.96875,"val-aux/sciknoweval/incorrect_format/best@4/mean":1,"val-aux/sciknoweval/score/worst@4/std":0.08715504524810262,"val-aux/sciknoweval/incorrect_format/best@4/std":0,"prompt_length/clip_ratio":0,"val-aux/sciknoweval/reward/maj@4/std":0.13821444887569323,"self_distillation/empty_target_batch":0.0703125,"critic/returns/mean":-0.006017094478011131,"rollout_corr/log_ppl_diff":0.0003987445088569075,"timing_s/adv":0.4795156195759773,"val-aux/sciknoweval/incorrect_format/maj@16/std":0,"val-aux/sciknoweval/reward/best@8/mean":0.9082190476190476,"timing_s/agent_loop/generate_sequences/mean":2.3858929471098236,"val-aux/sciknoweval/incorrect_format/worst@2/std":0,"critic/rewards/max":1,"val-aux/sciknoweval/reward/worst@16/std":0.02126535217732578,"srpo/grpo_actor/pg_clipfrac_lower":0,"timing_per_token_ms/adv":0.00362215690397614,"prompt_length/mean":296.15625,"rollout_corr/ppl_ratio":1.0004019325133413,"val-core/sciknoweval/acc/best@16/mean":0.9349142857142857,"rollout_corr/k3_kl":0.0004779839419200016,"rollout_corr/rollout_is_ratio_fraction_low":1.767783942341339e-05,"val-aux/num_turns/min":2,"timing_s/agent_loop/slowest/generate_sequences":6.079234275966883,"srpo/grpo_actor/ppo_kl":0,"val-aux/sciknoweval/score/best@2/mean":0.8065190476190476,"timing_s/save_checkpoint":12.84816437587142,"val-aux/sciknoweval/incorrect_format/worst@16/std":0,"srpo/sdpo_loss":0.0033741780134732835,"val-aux/sciknoweval/reward/maj@2/mean":0.7362714285714287,"rollout_corr/log_ppl_diff_min":-0.0124589204788208,"val-aux/sciknoweval/acc/worst@4/mean":0.6136428571428572,"val-aux/sciknoweval/incorrect_format/worst@16/mean":1,"rollout_corr/rollout_is_std":0.029428821057081223,"critic/advantages/max":0.875,"val-aux/sciknoweval/reward/worst@2/mean":0.6664857142857143,"val-aux/sciknoweval/acc/best@8/mean":0.9082190476190476,"val-aux/sciknoweval/score/maj@16/mean":0.7606666666666667,"val-aux/sciknoweval/score/worst@8/mean":0.5821809523809525,"self_distillation/feedback_available_fraction":0,"val-aux/sciknoweval/reward/maj@2/std":0.16273431336078428,"val-aux/sciknoweval/acc/best@4/std":0.11009573982420097,"srpo/correct_sample_fraction":0.75390625,"val-aux/sciknoweval/acc/best@2/std":0.14511294435717983,"val-aux/sciknoweval/score/worst@16/std":0.0212653