discovered 03 Aug 2026
heretic
→ View on GitHubHeretic is an automated tool designed to remove censorship from transformer-based language models efficiently, utilizing advanced directional ablation techniques complemented by a TPE-based optimization powered by Optuna. It produces decensored models that maintain the original model's performance while significantly reducing responses to harmful prompts, thus enabling users without deep technical knowledge to effectively use the tool. Supporting a variety of dense models, including multimodal and MoE architectures, Heretic delivers high-quality results comparable to manually crafted abliterations.