HEROS-RL · deterministic reward
Migration Reward Lab
Score a completion with the browser port of the format + SQL safety reward. It rejects destructive first steps, rewards a multi-phase expand-and-contract plan and a rollback. The forge risk gate runs only in the Python package.
—
- Format reward
- —
- Accuracy reward
- —
- First statement destructive
- —
- Error
- —
arXiv:2402.03300arXiv:2501.12948 Source: HEROS-RL · Dataset: heros-migration-500