heretic

Why we included this project

Heretic removes the safety alignment that many transformer models ship with, and it does so without retraining. The tool automates abliteration, the technique of editing model activations to suppress refusal behavior, and uses an optimizer to tune the parameters for you. The result is a decensored model that keeps most of the original's reasoning ability, since the process minimizes both refusals and divergence from the base model. You run it from the command line, point it at a Hugging Face model, and it handles dense, multimodal, MoE, and some hybrid architectures. If you self-host models and the built-in guardrails block responses you actually need, this is a practical way to get an uncensored local model for roleplay, creative writing, or research.