p-e-w/heretic
Repository · https://github.com/p-e-w/heretic
Training-free language model modification: Remove censorship, suppress slop, reduce sycophancy, and more
First seen by Oriel on .
AI summary
Heretic is a training-free tool that modifies locally runnable language models by optimizing behavioral scores, enabling censorship removal and other behavior changes on consumer hardware.
- Heretic uses contrasting prompts, configurable scorers, and mathematical modifiers rather than expensive training to steer model behavior.
- Its default Arbitrary-Rank Ablation modifier targets refusal suppression, while directional ablation supports projected abliteration and norm preservation.
- Users can install it with Python 3.10+ and PyTorch 2.6+; processing Qwen3-4B-Instruct-2507 takes about 30–40 minutes on an RTX 3090.
- Scorer and modifier plugins let users optimize refusal rate, KL divergence, or custom objectives without modifying the Heretic package.
- Optional reproduction manifests enable byte-for-byte model recreation when environments match, while bnb_4bit reduces VRAM requirements with a slight accuracy cost.
#language models #abliteration #optuna #model reproducibility
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Stars
Total repository stars at each observation.
- GitHub Trending: 39,147, latest observation ; +1,403 over observed period since
Stars gained (1d)
Daily star gain reported by GitHub Trending. A change compares two daily values; it is not additional stars between observations.
- GitHub Trending: 701, latest observation ; +297 over observed period since
Forks
Total repository forks at each observation.
- GitHub Trending: 4,414, latest observation ; +111 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| GitHub Trending | p-e-w/heretic | https://github.com/p-e-w/heretic Read discussion |