SFTGuard v0.1.0 is the first public alpha release of a fail-closed preflight and regression gate for supervised fine-tuning pipelines.
Highlights:
- Audit bounded chat JSON/JSONL plus trusted exported token-role and loss-mask evidence.
- Detect schema/role-order faults, duplicate or conflicting examples, train/eval leakage, credential-shaped text, EOS mistakes, prompt supervision, and destructive truncation.
- Gate paired base-vs-adapter target and retention metrics with explicit PASS, FAIL, or ABSTAIN outcomes.
- Inspect native artifacts in a browser-local static viewer with no report upload endpoint or analytics.
Frozen internal synthetic evidence:
- 270/270 injected known faults detected across 9 registered fault classes.
- 0/30 clean controls flagged.
- Byte-identical replay from the recorded implementation commit.
Claim boundary: these are preregistered internal regression results for known synthetic fault templates. They are not an external benchmark, model-quality result, safety certification, or proof that unknown failures will be detected. v0.1 trusts exported mask evidence and does not yet ship direct Hugging Face/TRL tokenizer integration.
Validated on Ubuntu and Windows with Python 3.11 and 3.12, plus CodeQL and Python/npm dependency audits.