⬢github Python · 39.7K ★ +2K since we first saw it · pushed 13 h ago · AGPL-3.0
p-e-w/heretic
Training-free language model modification: Remove censorship, suppress slop, reduce sycophancy, and more
Heretic is a Python tool that modifies language models without training, using mathematical techniques and example prompts to change behavior like removing refusals/censorship, reducing slop, sycophancy, or humor. It runs on consumer hardware, supports quantization, works with most dense, MoE, and hybrid models, and lets users publish modified models to Hugging Face.
Why now: People are discussing it because it went viral (trending repository) and independent community benchmarks on Reddit suggest its abliterated models compare favorably to competing abliteration tools.
Who it is for: Local LLM enthusiasts who run open models on their own GPUs and want to customize model behavior without training.
abliterationanti-censorshipanticensorshipcensorshipcensorship-resistancellmtransformer
Stars over our 306 snapshots: 37.7K to 39.7K, since 4 d ago.
Where people talked about it
- ⬢github p-e-w/heretic 2 d ago
API: https://socialmediatrends-api.osmike.com/v1/repos/p-e-w/heretic