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

1 line
10 KiB
JSON
Raw Normal View History

{"val-aux/sciknoweval/reward/worst@2/mean":0.6398875,"val-aux/sciknoweval/incorrect_format/maj@2/mean":0.9178625,"perf/mfu/actor":0.1938518522795187,"training/rollout_actor_probs_pearson_corr":0.9995552897453308,"timing_s/start_profile":9.544752538204193e-05,"val-aux/sciknoweval/score/best@8/mean":0.94765,"response_length_non_aborted/mean":1854.16015625,"val-aux/sciknoweval/acc/best@4/mean":0.8962625000000001,"_timestamp":1.7830577481316543e+09,"rollout_corr/rollout_is_seq_fraction_high":0,"val-aux/sciknoweval/score/maj@16/std":0.11988946359651878,"timing_s/agent_loop/generate_sequences/min":3.829353081062436,"val-aux/sciknoweval/incorrect_format/best@8/std":0.006519985280372976,"num_turns/mean":2,"val-aux/sciknoweval/acc/worst@8/std":0.07452105831546266,"timing_s/agent_loop/slowest/prompt_length":342,"val-aux/sciknoweval/score/maj@16/mean":0.8134375,"val-aux/sciknoweval/acc/worst@4/mean":0.5704125,"rollout_corr/rollout_is_eff_sample_size":0.9992375231521491,"rollout_corr/rollout_is_seq_max":1.0046027898788452,"rollout_corr/rollout_is_mean":1.0000216960906982,"val-aux/sciknoweval/score/maj@8/mean":0.7921125,"val-aux/sciknoweval/reward/best@8/std":0.07747202634384016,"training/rollout_probs_diff_std":0.007731531746685505,"perf/max_memory_reserved_gb":144.7109375,"rollout_corr/kl":0.0003901403655550695,"val-aux/sciknoweval/reward/maj@4/std":0.18390036772261203,"val-aux/sciknoweval/reward/worst@4/mean":0.5704125,"val-aux/sciknoweval/acc/maj@4/mean":0.76555,"val-aux/sciknoweval/score/maj@2/mean":0.7299875,"actor/lr":1e-06,"val-aux/sciknoweval/reward/best@4/mean":0.8962625000000001,"critic/returns/min":-0.875,"val-aux/sciknoweval/reward/best@4/std":0.1329636020160485,"num_turns/max":2,"val-aux/sciknoweval/score/worst@16/mean":0.49838750000000004,"global_seqlen/balanced_min":67861,"val-aux/sciknoweval/incorrect_format/maj@2/std":0.1290617057662198,"timing_s/agent_loop/tool_calls/min":0,"response/aborted_ratio":0,"timing_s/update_actor":19.037666445598006,"timing_s/agent_loop/generate_sequences/mean":17.904436235730827,"training/rollout_probs_diff_max":0.5070421695709229,"timing_s/gen":74.55190750397742,"timing_s/old_log_prob":4.5896848775446415,"rollout_corr/rollout_is_ratio_fraction_high":0,"val-aux/sciknoweval/score/mean@16":0.7296875,"actor/pg_clipfrac":4.500477859892271e-05,"num_turns/min":2,"timing_s/agent_loop/slowest/generate_sequences":67.18781132809818,"val-aux/sciknoweval/incorrect_format/mean@16":0.91796875,"val-aux/sciknoweval/reward/best@16/std":0.03523755515622687,"timing_s/agent_loop/generate_sequences/max":67.18781132809818,"val-aux/sciknoweval/reward/std@16":0.21113028788197769,"val-aux/sciknoweval/acc/worst@8/mean":0.52575,"val-aux/sciknoweval/incorrect_format/maj@4/std":0.10365239649604263,"timing_per_token_ms/update_actor":0.035021397027595616,"val-aux/sciknoweval/incorrect_format/worst@2/mean":0.8673375,"val-aux/sciknoweval/incorrect_format/worst@4/std":0.13896621816207683,"val-aux/sciknoweval/incorrect_format/maj@8/mean":0.9701000000000001,"val-aux/sciknoweval/acc/best@8/mean":0.94765,"rollout_corr/training_log_ppl":0.35900981855229475,"actor/entropy":0.45818907022476196,"val-aux/sciknoweval/incorrect_format/best@4/std":0.02834928965541194,"val-aux/sciknoweval/incorrect_format/maj@8/std":0.07166964518321907,"rollout_corr/rollout_is_seq_mean":1.000037670135498,"val-aux/sciknoweval/acc/worst@4/std":0.12181204475036858,"val-aux/sciknoweval/incorrect_format/std@16":0.12921982287160344,"val-aux/sciknoweval/incorrect_format/best@4/mean":0.99245,"rollout_corr/log_ppl_diff_min":-0.003899216651916504,"critic/advantages/max":0.875,"val-core/sciknoweval/acc/maj@16/mean":0.8134375,"rollout_corr/log_ppl_abs_diff":0.0008009577286429703,"rollout_corr/rollout_is_seq_fraction_low":0,"timing_s/reward":0.11913642473518848,"rollout_corr/rollout_is_seq_max_deviation":0.004602789878845215,"val-aux/sciknoweval/incorrect_format/best@2/mean":0.969875,"_wandb":{"runtime":8906},"timing_per_token_ms/adv":0.000223674584299434,"val-aux/sciknoweval/reward/maj@16/mean":0.8134375,"val-aux/sciknoweval/reward/best@8/mean":0.94765,"perf/mf