laion/a3-rl-DCAgent_exp_rpt_pymethods2test-large-80-8B

a3-rl-DCAgent_exp_rpt_pymethods2test-large — global_step 80 (EMA winner)

RL (SkyRL) fine-tune of an 8B Qwen3 agent, trained from the a3 GLM-SFT base laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink on the task set DCAgent/exp_rpt_pymethods2test-large.

This is the winner checkpoint at trainer/global_step 80 (max_steps=80), selected by 5-period EMA (α=1/3) of reward/avg_raw_reward over the training chain, constrained to global_step ≤ 80.

• EMA(step 80) = 0.4829
• raw avg_raw_reward at step 80 = 0.5586 (mean over n_samples_per_prompt=8)
• hf_save_interval=5, ckpt_interval=2, n_samples_per_prompt=8

Note: the run trained a few spurious extra steps (81–86) due to a resume-past-max_steps artifact (each resume past the cap trains one extra step). Those steps are not part of the legitimate run and were excluded from checkpoint selection; the legit run ends at global_step 80.

Training Traces

Training-time Daytona/Harbor rollouts for this run are uploaded as a companion dataset: penfever/a3-rl-DCAgent_exp_rpt_pymethods2test-large

The dataset contains the last episode of each trial (per make_and_upload_trace_dataset --episodes last) — the same rollouts the policy was trained on after rollback / truncation.