# p\-e\-w/heretic

- HackerByte page: <https://hackerbyte.net/en/projects/p-e-w-heretic-1061041576>
- GitHub: <https://github.com/p-e-w/heretic>
- Programming language: Python

Training\-free language model modification: Remove censorship, suppress slop, reduce sycophancy, and more

## AI-generated summary

Heretic is a novel tool that modifies pre\-trained language models without requiring expensive training. It employs mathematical techniques and example prompts to steer model behavior, allowing users to remove censorship, reduce unwanted outputs like sycophancy or humor, and alter emotional alignment. The tool is designed for accessibility, enabling users to modify models on their personal hardware with relative ease.

## Key points

- Modifies LLMs without full training using mathematical techniques and example prompts.
- Can remove censorship, suppress unwanted outputs \(slop, humor, sycophancy\), and alter emotional alignment.
- Supports on\-the\-fly 4\-bit model quantization for reduced VRAM usage.
- Allows users to create reproducible models with detailed instructions and machine\-readable manifests.
- The 'Heretic Grimoire' provides a distributed library of reproducible models that can be collected locally.