heretic: Remove LLM Censorship Without Manual Tuning, Lowering KL Divergence to 0.16

Combines directional abliteration with Optuna-based parameter optimization to automatically remove safety alignment from language models. It is a fully automated tool that can be run with a single command, requiring no understanding of complex transformer internals.

It works by simultaneously minimizing the number of refusal responses and the KL divergence from the original model. On Gemma-3-12b, while other manually created abliterated models recorded KL divergences of 0.45–1.04, Heretic lowered it to 0.16, minimizing the loss of original intelligence.

It broadly supports models ranging from Dense to MoE and hybrid models like Qwen3.5. Using bitsandbytes quantization reduces VRAM pressure, taking approximately 20–30 minutes to process a 4B model on an RTX 3090.

Enabling research options allows for 2D projection visualization of residual vectors and layer-wise geometric structure analysis. The community has already published over 5,000 models using this tool.