view article Article SafetyPulse: Catching Safety Drift During Fine-Tuning by Reading the Model's Refusal Circuit purvansh01 • 1 day ago • 5
SafetyPulse - data Collection SafetyPulse (Lexsi hackathon, Track 2b): eval splits, fine-tune data, contrast pairs, probe bundle, monitoring and audit logs. • 5 items • Updated 1 day ago • 2