# a3-rl-DCAgent_selfinstruct-naive-sandboxes-2-verified-70-8B

## Training Traces

Training-time Daytona/Harbor rollouts for this run are uploaded as
a companion dataset:
**[penfever/a3-rl-DCAgent_selfinstruct-naive-sandboxes-2-verified](https://huggingface.co/datasets/penfever/a3-rl-DCAgent_selfinstruct-naive-sandboxes-2-verified)**

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.

## Provenance

- Base (SFT) model: `laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink`
- RL dataset: `DCAgent/selfinstruct-naive-sandboxes-2-verified`
- Checkpoint: `global_step_70` (EMA-best over steps 1-80, 5-period EMA α=1/3)
- Training: SkyRL GRPO, 56 GPU / 14 nodes, fan-out K=4; chain 567548→567550→567551 (terminated early at step ~81 by user request)