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

1 line
11 KiB
JSON
Raw Permalink Normal View History

{"val-aux/tooluse/score/worst@8/std":0.04295702328078124,"rollout_corr/rollout_log_ppl":0.13506622198246987,"global_seqlen/max":26968,"val-aux/tooluse/incorrect_format/best@16/mean":0,"_wandb":{"runtime":3606},"val-aux/tooluse/acc/best@4/mean":0.6149852941176471,"val-aux/tooluse/score/best@4/std":0.008533867314311838,"rollout_corr/rollout_is_seq_max_deviation":0.03070974349975586,"training/rollout_probs_diff_mean":0.0043131280690431595,"val-aux/tooluse/incorrect_format/maj@16/mean":0,"timing_s/gen":8.898955885320902,"prompt_length/clip_ratio":0,"val-aux/tooluse/acc/std@16":0.0397091660129265,"critic/score/min":0,"self_distillation/empty_target_batch":0.0078125,"val-aux/tooluse/reward/best@16/mean":0.6176323529411765,"self_distillation/reprompt_sample_fraction":0.9921875,"num_turns/mean":2,"rollout_corr/rollout_is_seq_max":1.0307097434997559,"val-core/tooluse/acc/best@16/std":0.00046480825380232674,"self_distillation/token_reweight_eps_w":0.20000000000000004,"val-aux/tooluse/reward/worst@4/std":0.04688580916450452,"val-aux/num_turns/mean":2,"response_length_non_aborted/clip_ratio":0,"val-aux/tooluse/score/maj@8/std":0.016895760456278525,"val-aux/num_turns/min":2,"perf/time_per_step":27.41770182736218,"rollout_corr/rollout_is_ratio_fraction_high":8.322236681124195e-05,"val-aux/tooluse/reward/best@4/std":0.008533867314311838,"val-aux/tooluse/acc/worst@16/std":0.03331555519855917,"val-aux/tooluse/reward/best@2/std":0.021252889471013774,"val-aux/tooluse/acc/worst@2/mean":0.5792499999999999,"val-aux/tooluse/score/maj@4/std":0.028165057960484432,"critic/returns/mean":-0.008946404792368412,"val-aux/tooluse/score/std@16":0.0397091660129265,"critic/rewards/mean":0.43359375,"global_seqlen/balanced_min":21525,"val-aux/tooluse/acc/worst@16/mean":0.5220294117647059,"val-aux/tooluse/acc/best@4/std":0.008533867314311838,"self_distillation/token_reweight_w_clipped_mean":1.0005437803920358,"val-aux/tooluse/incorrect_format/worst@8/mean":0,"val-aux/tooluse/acc/maj@4/std":0.028165057960484432,"timing_s/testing":21.686641938984394,"training/rollout_probs_diff_max":0.41022178530693054,"training/rollout_probs_diff_std":0.013580762781202793,"val-aux/tooluse/reward/mean@16":0.59375,"val-core/tooluse/acc/maj@16/std":0.010892157021494882,"actor/grad_norm":0.2933488469570875,"perf/throughput":785.1861594947975,"val-aux/tooluse/score/worst@2/mean":0.5792499999999999,"rollout_corr/log_ppl_diff_min":-0.01413802057504654,"rollout_corr/rollout_is_max":2.820589303970337,"rollout_corr/log_ppl_abs_diff":0.0032595000756145964,"critic/rewards/max":1,"val-aux/tooluse/acc/maj@8/std":0.016895760456278525,"val-aux/tooluse/incorrect_format/worst@16/std":0,"val-aux/tooluse/reward/maj@8/mean":0.6039705882352941,"timing_s/agent_loop/generate_sequences/min":0.5163722541183233,"val-core/tooluse/acc/best@16/mean":0.6176323529411765,"rollout_corr/chi2_token":0.0012444618623703718,"val-aux/tooluse/incorrect_format/std@16":0,"rollout_corr/rollout_is_seq_fraction_high":0,"prompt_length/max":934,"val-aux/tooluse/incorrect_format/worst@8/std":0,"rollout_corr/training_ppl":1.150276112370193,"rollout_corr/rollout_is_mean":1.0004314184188843,"val-aux/tooluse/acc/worst@4/mean":0.5617647058823529,"timing_per_token_ms/update_actor":0.09024581746989048,"timing_s/agent_loop/tool_calls/mean":0,"critic/advantages/mean":-0.008946404792368412,"val-aux/tooluse/acc/maj@2/mean":0.5937794117647059,"critic/returns/min":-0.875,"timing_s/step":27.41770182736218,"val-aux/tooluse/score/maj@2/mean":0.5937794117647059,"response/aborted_ratio":0,"timing_s/reward":0.0715668685734272,"prompt_length/mean":578.875,"rollout_corr/rollout_is_std":0.03966965898871422,"val-aux/tooluse/incorrect_format/best@8/std":0,"timing_s/agent_loop/tool_calls/max":0,"response_length_non_aborted/mean":93.875,"val-aux/tooluse/score/worst@4/mean":0.5617647058823529,"val-aux/tooluse/incorrect_format/maj@4/mean":0,"val-aux/tooluse/incorrect_format/best@2/std":0,"rollout_corr/rollout_is_min":0.5896446108818054,"_runtime":3606.405012169,"val-aux/tooluse/reward/maj@2/std":0.04051488817390775,"self_distillation/success_sample_f