Files
qwen3-4b-tooluse-srpo-ema005/artifacts/wandb-summary.json

1 line
11 KiB
JSON
Raw Normal View History

{"perf/throughput":979.2552930338192,"rollout_corr/k3_kl":0.001191004929382622,"actor/pg_loss":-0.018329072030610405,"timing_s/gen":11.10028469748795,"val-aux/tooluse/score/best@2/mean":0.5926029411764705,"val-aux/num_turns/max":2,"rollout_corr/log_ppl_abs_diff":0.003266506794716406,"val-aux/tooluse/acc/maj@4/mean":0.5755735294117647,"val-aux/tooluse/score/maj@4/mean":0.5755735294117647,"val-aux/tooluse/acc/maj@8/mean":0.5769705882352941,"response_length_non_aborted/max":621,"val-aux/tooluse/acc/maj@2/std":0.04960767852121298,"rollout_corr/rollout_is_seq_fraction_low":0,"critic/rewards/min":0,"val-aux/tooluse/score/best@4/mean":0.6068382352941176,"training/rollout_probs_diff_valid":1,"val-aux/tooluse/reward/mean@16":0.5726102941176471,"rollout_corr/rollout_log_ppl":0.08498664719536464,"val-aux/tooluse/score/maj@8/mean":0.5769705882352941,"timing_s/agent_loop/slowest/prompt_length":577,"val-aux/tooluse/incorrect_format/maj@8/std":0.001135690269194578,"val-aux/num_turns/mean":2,"val-aux/tooluse/score/best@4/std":0.0287069371757289,"perf/mfu/actor":0.045203198343943346,"val-aux/tooluse/reward/best@4/mean":0.6068382352941176,"rollout_corr/log_ppl_diff_max":0.035721685737371445,"training/epoch":1,"val-aux/tooluse/incorrect_format/worst@4/std":0,"val-aux/tooluse/acc/worst@4/mean":0.5356617647058824,"val-aux/tooluse/score/std@16":0.04960007010677716,"val-aux/tooluse/reward/worst@8/mean":0.5206617647058823,"val-aux/tooluse/incorrect_format/worst@2/mean":0.00010294117647058824,"perf/max_memory_reserved_gb":152.583984375,"val-aux/tooluse/score/maj@4/std":0.03821000405576706,"timing_s/adv":1.2358614541590214,"rollout_corr/rollout_is_std":0.03342141583561897,"val-aux/tooluse/reward/worst@4/mean":0.5356617647058824,"val-aux/tooluse/acc/worst@4/std":0.0360951572887439,"self_distillation/dynamic_weighting_enabled":1,"val-aux/tooluse/reward/best@4/std":0.0287069371757289,"self_distillation/feedback_used_fraction":0,"val-aux/tooluse/acc/best@8/mean":0.6173235294117647,"val-aux/tooluse/incorrect_format/worst@16/mean":0,"val-aux/tooluse/reward/worst@16/std":0.01768991770749048,"rollout_corr/rollout_is_ratio_fraction_high":0,"val-aux/tooluse/score/maj@16/mean":0.57725,"val-aux/tooluse/incorrect_format/maj@4/mean":0.00042647058823529416,"val-core/tooluse/acc/maj@16/std":0.019721439199833088,"global_seqlen/minmax_diff":12263,"rollout_corr/kl":0.0009711506442058315,"timing_s/agent_loop/tool_calls/min":0,"self_distillation/feedback_available_fraction":0,"val-aux/tooluse/acc/worst@2/std":0.04453765865500009,"timing_s/agent_loop/slowest/response_length":621,"timing_s/agent_loop/generate_sequences/mean":2.140533648460405,"rollout_corr/chi2_seq":0.2624655382242054,"srpo/routed_token_count":116.134765625,"perf/total_num_tokens":357549,"val-aux/tooluse/score/worst@4/std":0.0360951572887439,"timing_s/agent_loop/tool_calls/mean":0,"global_seqlen/balanced_max":44695,"val-aux/tooluse/incorrect_format/maj@2/std":0.0034374036633244977,"val-aux/tooluse/incorrect_format/best@4/std":0.00615057597100357,"timing_s/agent_loop/slowest/tool_calls":0,"timing_s/old_log_prob":4.850968439131975,"val-aux/tooluse/acc/best@2/std":0.03951972692059264,"val-aux/tooluse/incorrect_format/best@8/mean":0.005808823529411765,"critic/score/min":0,"val-aux/tooluse/reward/maj@8/std":0.0264644012974148,"training/rollout_probs_diff_std":0.01246405765414238,"training/rollout_actor_probs_pearson_corr":0.996764600276947,"val-aux/tooluse/reward/best@2/std":0.03951972692059264,"critic/score/max":1,"val-aux/tooluse/incorrect_format/mean@16":0.0009191176470588235,"val-aux/tooluse/acc/best@4/std":0.0287069371757289,"val-aux/tooluse/incorrect_format/best@2/std":0.004880137697302882,"srpo/grpo_actor/pg_clipfrac":0.00016560931271669688,"val-aux/tooluse/acc/best@4/mean":0.6068382352941176,"timing_s/agent_loop/generate_sequences/max":4.9810793828219175,"critic/rewards/mean":0.30859375,"val-aux/tooluse/reward/best@8/std":0.019675115236200497,"val-aux/tooluse/score/worst@8/std":0.02801237697039909,"timing_s/reward":0.12771783582866192,"actor/entropy":0.0914105623960495,"critic/returns/mean":-0.0008