a3-rl-DCAgent_mix_h4_binary_easy-50-8B

RL (SkyRL) checkpoint trained on the DCAgent/mix_h4_binary_easy task mix, starting from the a3 GLM-SFT base laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink.

• Run: a3-rl-DCAgent_mix_h4_binary_easy (Jupiter chain 573271→573276)
• Training: 2 epochs, completed at global_step 67 (max_steps=80 was a ceiling; 67 = epoch-data completion, a legit deployable endpoint — not collapse).
• Checkpoint selected: global_step 50 by trailing-5 EMA (α=1/3) of reward/avg_raw_reward across the full chain.
  • EMA at step 50 = 0.4991 (highest of all aligned eligible checkpoints)
  • raw avg_raw_reward at step 50 = 0.529
  • pass@8 at step 50 = 0.672
• hf_save_interval=5

Training Traces

Training-time Daytona/Harbor rollouts for this run are uploaded as a companion dataset: penfever/a3-rl-DCAgent_mix_h4_binary_easy

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.

Training Logs

See training_logs/ for the parsed metrics (metrics.csv, report.md, reward_plot.png) and the raw chain .out logs (Jupiter has no W&B network access).