# a3-rl-DCAgent_exp_rpt_curriculum-medium-10-8B

RL (SkyRL/terminus-2) fine-tune of `laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink` on `DCAgent/exp_rpt_curriculum-medium` (512-task set, 2 epochs).

Checkpoint = **global_step_10**, selected by EMA-best of `reward/avg_raw_reward` (5-step EMA, α=1/3) across the full chain (job 591150). EMA at step 10 = 0.3106 (reward 0.2812). Data-limited run; training metrics emitted through step 16, checkpoints exported through step 21.

## Training Traces

Training-time Daytona/Harbor rollouts for this run are uploaded as a companion dataset:
**[penfever/a3-rl-DCAgent_exp_rpt_curriculum-medium](https://huggingface.co/datasets/penfever/a3-rl-DCAgent_exp_rpt_curriculum-medium)**

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 per-step metrics CSVs, the metrics report, and the reward-vs-steps plot (Jupiter has no W&B).