ablation-pymethods2test-shaped-45-8B
RL (SkyRL GRPO) checkpoint from the shaped-reward ablation of the a3-successor
study. Reward = shaped pass-ratio (fraction of tests passing, reward_shaper=pass_ratio),
as opposed to the binary all-tests-pass reward of the a3 series.
- Base model: laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink (a Qwen3-8B SFT)
- Training dataset: DCAgent/exp_rpt_pymethods2test-large
- Checkpoint:
global_step_45, selected as the best checkpoint by EMA (alpha=1/3, trailing-5 window) ofreward/avg_raw_rewardcomputed across the full 80-step training chain (EMA = 0.4712 at step 45). - Training: 80 steps total,
hf_save_interval=5, 14x GH200 nodes on JSC Jupiter.
The rl_config.json in this repo is the exact launch config used for reproducibility.
Training Traces
Training-time Daytona/Harbor rollouts for this run are uploaded as a companion dataset: open-athena/ablation-pymethods2test-shaped
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.