# Benchmark Gates

No training run should start until the pre-training CyberGym baseline is complete.

Current required gate:

- `cybergym_pretraining_baseline` against the untouched base model.
- First pass: official 10-task subset at `level1`.
- Real comparison pass: frozen `level1` sample of 100-200 tasks, using the same task ids for base, LoRA checkpoints, and merged model.

Artifacts to keep:

- task id list
- model endpoint config
- OpenHands logs
- CyberGym PoC database
- verifier output
- pass/fail table
- cost/runtime notes

The main recipe is `training/recipes/pretraining_cybergym_baseline.md`.