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

1 line
11 KiB
JSON
Raw Permalink Normal View History

{"response_length/mean":279.4140625,"critic/advantages/max":0.75,"val-aux/sciknoweval/acc/best@2/std":0.02918137925697454,"rollout_corr/rollout_log_ppl":0.05304120097935083,"val-aux/sciknoweval/incorrect_format/worst@8/mean":1,"rollout_corr/rollout_is_seq_max":1.0095514059066772,"response_length/clip_ratio":0,"training/rollout_probs_diff_mean":0.00194946164265275,"timing_s/agent_loop/generate_sequences/min":1.194736858829856,"rollout_corr/rollout_is_seq_min":0.996039092540741,"_timestamp":1.7830039799930708e+09,"val-aux/sciknoweval/score/maj@16/mean":0.7758829787234042,"val-aux/sciknoweval/score/maj@16/std":0.014657220600786369,"val-aux/sciknoweval/reward/best@2/mean":0.7854148936170213,"val-aux/sciknoweval/reward/best@16/mean":0.8047021276595744,"rollout_corr/log_ppl_diff":0.0006709559474984417,"critic/returns/min":-0.625,"rollout_corr/rollout_is_seq_fraction_high":0,"actor/entropy":0.052000705152750015,"critic/score/max":1,"val-core/sciknoweval/acc/best@16/std":0.005902094678345516,"val-aux/sciknoweval/score/best@16/mean":0.8047021276595744,"val-aux/sciknoweval/incorrect_format/best@4/std":0,"val-aux/sciknoweval/acc/worst@4/mean":0.7343617021276596,"rollout_corr/rollout_is_mean":0.9999290108680725,"timing_s/start_profile":8.36830586194992e-05,"val-aux/sciknoweval/acc/worst@16/std":0.02298913016887447,"rollout_corr/rollout_is_ratio_fraction_high":1.3980147741676774e-05,"val-aux/sciknoweval/reward/best@4/std":0.01639341752615574,"self_distillation/token_reweight_correct_sample_fraction":0.68359375,"rollout_corr/rollout_is_seq_fraction_low":0,"self_distillation/token_reweight_w_std":431654.97193862277,"val-aux/sciknoweval/score/worst@16/std":0.02298913016887447,"val-aux/sciknoweval/reward/worst@8/std":0.03147132698227977,"critic/rewards/max":1,"perf/mfu/actor_infer":0,"val-aux/sciknoweval/score/best@2/std":0.02918137925697454,"val-aux/sciknoweval/score/maj@4/std":0.031426065862592674,"timing_per_token_ms/update_actor":0.08712704845144265,"val-aux/sciknoweval/incorrect_format/worst@16/mean":1,"val-aux/sciknoweval/incorrect_format/worst@2/std":0,"rollout_corr/rollout_is_seq_std":0.0017657832941040397,"self_distillation/token_reweight_w_clip_frac":0.012708691239822656,"num_turns/min":2,"self_distillation/empty_target_batch":0.28125,"response_length_non_aborted/min":91,"val-aux/num_turns/max":2,"val-aux/num_turns/min":2,"training/rollout_actor_probs_pearson_corr":0.9965116381645203,"critic/score/min":0,"val-aux/sciknoweval/incorrect_format/worst@4/mean":1,"perf/mfu/actor":0.03965654144334235,"self_distillation/token_reweight_active_token_fraction":0.71875,"timing_s/agent_loop/generate_sequences/max":4.990295339375734,"global_seqlen/min":18525,"val-aux/sciknoweval/incorrect_format/best@8/std":0,"num_turns/max":2,"prompt_length/mean":391.125,"val-aux/sciknoweval/incorrect_format/best@16/std":0,"val-aux/sciknoweval/score/worst@8/std":0.03147132698227977,"global_seqlen/max":24595,"num_turns/mean":2,"rollout_corr/ppl_ratio":1.0006739916279912,"val-aux/sciknoweval/acc/maj@2/std":0.04273575303351148,"val-aux/sciknoweval/incorrect_format/worst@16/std":0,"actor/pg_clipfrac_lower":0,"global_seqlen/balanced_max":21459,"actor/grad_norm":0.21034935116767883,"response_length_non_aborted/mean":279.4140625,"val-aux/sciknoweval/score/best@16/std":0.005902094678345516,"val-aux/sciknoweval/incorrect_format/worst@8/std":0,"timing_s/old_log_prob":2.2547264005988836,"val-aux/sciknoweval/reward/worst@8/mean":0.7189574468085107,"actor/pg_loss":0.0006925989873707294,"prompt_length/clip_ratio":0,"val-aux/sciknoweval/reward/best@2/std":0.02918137925697454,"val-aux/sciknoweval/score/std@16":0.04264975776630575,"_runtime":3535.705405695,"perf/max_memory_allocated_gb":117.12247371673584,"val-aux/sciknoweval/acc/maj@2/mean":0.7684468085106384,"val-aux/sciknoweval/acc/worst@8/std":0.03147132698227977,"val-aux/sciknoweval/incorrect_format/best@8/mean":1,"rollout_corr/rollout_is_seq_mean":0.9999463558197021,"val-aux/sciknoweval/incorrect_format/maj@2/mean":1,"prompt_length/min":190,"timing_per_token_ms/adv":0.0029145551329557463,"global_seqlen/balanced_min":