Overview

Heretic is a groundbreaking tool that automates the removal of censorship, often referred to as "safety alignment," from transformer-based language models. It achieves this without the need for costly post-training processes. By ingeniously combining advanced directional ablation techniques with an efficient parameter optimization system, Heretic delivers powerful, uncensored models that preserve their original intelligence.

Key Features

• Fully Automatic Censorship Removal: Heretic operates entirely without manual intervention, finding optimal "abliteration" parameters through co-minimizing refusals and KL divergence from the original model.
• Preserves Model Intelligence: The sophisticated optimization process ensures that decensored models retain a high degree of their original capabilities, avoiding significant performance degradation.
• Broad Model Compatibility: Supports a wide array of dense models, including many multimodal variants, various MoE architectures, and hybrid models like Qwen3.5.
• User-Friendly CLI: Designed for ease of use, requiring only basic command-line proficiency to decensor language models.
• Advanced Research Capabilities: Includes features for generating visual plots of residual vectors and providing quantitative analysis of residual geometry, aiding interpretability research.
• Community Driven: Boasts a vibrant community with over 5000 models published using Heretic on Hugging Face.

Typical Use Cases

Heretic is ideal for developers, researchers, and AI enthusiasts who require language models free from artificial safety constraints. This includes:

• Unrestricted AI Development: Creating AI models for applications where unbiased and uncensored responses are paramount.
• Research and Experimentation: Facilitating studies into the internal workings and emergent behaviors of large language models.
• Enhanced Creative Tools: Powering generative AI applications that benefit from a wider range of expressive output.
• Personalized AI Experiences: Tailoring LLMs to specific needs or philosophical viewpoints by removing pre-imposed limitations.