Heretic

Heretic is an open-source command-line tool that automatically removes safety guardrails and censorship from large language models through a mathematical process called "direction ablation" (or abliteration). 

A high-level overview of how the tool functions: How Heretic Works

Instead of manually tweaking model weights or using jailbreak prompts, Heretic automates the process. It uses a parameter optimizer powered by Optuna to adjust the model's internal representations. It co-minimizes the model's rate of refusal for "harmful" requests while keeping the KL divergence—a measure of "brain damage" or deviation from the model's original intelligence—extremely low. 

The Llama 3.3 FT Test

A Financial Times investigation highlighted just how quickly the tool works. Researchers were able to strip the safety filters from Meta’s heavily guarded Llama 3.3 model in just under 10 minutes. They did not need expensive or specialist hardware to accomplish this, using only four lines of code. 

How to Use It

Installation: Heretic is hosted on GitHub and can typically be installed via Python with pip install heretic-llm.

Execution: The process is entirely automated. You run a basic command specifying the target model (e.g., Llama 3.3), and the tool mathematically alters the censorship parameters without requiring you to understand complex transformer internals.

Availability: You can find the codebase and instructions on the Heretic GitHub 

Repository. Many users also share their pre-abliterated models on Hugging Face.