Part of the BioRefusalAudit project (GitHub),
submitted to AIxBio Hackathon 2026, Track 3 (Biosecurity Tools). Measures refusal depth —
the divergence between a model's surface behavior ("I refuse") and its internal SAE feature
activations — on bio-safety prompts.
This dataset is gated. To request access, attest that:
You are a biosecurity or AI safety researcher using the data for defensive purposes.
You will not reproduce… See the full description on the dataset page:
https://huggingface.co/datasets/Solshine/biorefusalaudit-gated.