obliteratus

Open-source toolkit identifying and surgically removing refusal directions from models, offering multiple extraction strategies and inference-time interventions for researchers studying alignment and safety.