heretic — Automatic Abliteration & Refusal-Direction Interpretability

Keyword: heretic · abliterate · decensor a model · remove refusals · refusal direction · plot residuals

Responsible use. Heretic is AGPL-3.0 software for research and for customizing models you have the right to modify. Respect each base model's license and every platform's Terms of Service when downloading, modifying, uploading, or serving weights. Abliteration lowers a model's refusal rate on all topics, including genuinely harmful ones — do not use it to produce a model for illegal content (e.g. CSAM, weapons-of-mass-destruction uplift, targeted harassment) or to evade safety controls you are contractually bound by. Legitimate uses include interpretability/alignment research, red-teaming, reducing benign over-refusal, and running uncensored models on your own hardware for lawful work.