tekosML's picture
Publish sanitized benchmark dataset and final dataset card (part 3)
e9b1b92 verified
|
Raw
History Blame Contribute Delete
641 Bytes
# Deterministic coding fixtures
`cases.json` contains three atomic Python tasks and hidden executable assertions. Prompts cap answers below 768 tokens. Do not include the tests in requests.
```bash
python scripts/validate_stage2.py coding results.jsonl
python scripts/validate_stage2.py coding /dev/null --emit-prompts
```
Input JSONL accepts `{"case_id": ..., "output_text": ...}` or native harness rows whose exact `prompt_text` identifies a case. Code is run with the current Python in isolated mode, a five-second timeout, and no network-dependent packages. Run validators only on trusted benchmark output; this is not an OS sandbox.