p-e-w / heretic

TL;DR

Heretic is a tool that automates the removal of censorship from transformer-based language models without costly post-training. It effectively retains the model's intelligence while minimizing harmful prompt refusals.

✦ Why It Matters

Try using Heretic today to enhance your language models by removing unnecessary censorship.

Key Takeaways

Install Heretic with pip to begin decensoring models.
Use the command-line interface for easy operation without deep knowledge.
Evaluate model performance with built-in functionality after processing.
Explore advanced features for research by installing optional components.

How It Works

Heretic employs directional ablation to optimize parameters that control how the model responds to harmful prompts. It minimizes refusals while preserving the model's original capabilities by co-minimizing the number of refusals and the KL divergence from the original model.

This process is fully automated, allowing users to achieve high-quality results without manual intervention.

p-e-w / heretic

BerriAI / litellm

tensorflow / tensorflow

open-webui / open-webui

Agentic data science with the Data Agent Kit

BerriAI / litellm

tensorflow / tensorflow

open-webui / open-webui

Agentic data science with the Data Agent Kit