heretic

Project Overview

Heretic is a fully automatic tool that removes censorship (safety alignment) from transformer-based language models using directional ablation (abliteration) combined with Optuna TPE optimization — no manual tuning required, preserving model intelligence while reducing refusals.

Key Features

• Fully Automatic: One command, zero configuration
• Abliteration: Directional ablation preserving model semantics
• KL Divergence Optimization: Minimizes damage to original model capabilities
• Wide Model Support: Dense, multimodal, and MoE architectures
• Quantization: bitsandbytes 4-bit support for low-VRAM deployment

Target Audience

• Local LLM researchers and enthusiasts
• Users wanting uncensored AI responses
• Interpretability researchers

Quick Start

1. Set up Python 3.10+ with PyTorch 2.2+
2. Install: pip install -U heretic-llm
3. Run: heretic Qwen/Qwen3-4B-Instruct-2507
4. After completion: save model, test chat, upload to HuggingFace, or run benchmarks

Core Advantages

• Completely automated — no transformer internals knowledge needed
• Outperforms manual abliterations in KL divergence
• 4-bit quantization runs 4B models on 16GB VRAM
• Community has generated 5000+ Heretic models

System Requirements

• OS: Linux (recommended), macOS, Windows
• Architecture: x64 (CUDA GPU recommended)
• Additional dependencies: Python 3.10+, PyTorch 2.2+, optional bitsandbytes

Notes

• Some models require newer PyTorch (e.g., MXFP4 needs 2.6+)
• Recommended to use uv: uv run heretic
• For legitimate research and educational purposes only