heretic
Training-free language model modification: Remove censorship, suppress slop, reduce sycophancy, and more
What is heretic?
Training-free language model modification: Remove censorship, suppress slop, reduce sycophancy, and more
Overview of heretic
What it is
Heretic is a training-free language model modification tool. It uses mathematical techniques to steer model behavior by contrasting example prompts and matching scorers, without requiring expensive training. It can be run on consumer hardware such as a gaming PC.
Key features
- Suppress refusals and remove censorship (default action)
- Reduce slop, humor, and sycophancy - Alter emotional alignment - Supports most dense models, multimodal models, MoE architectures, and hybrid models - Community has published thousands of models created with Heretic - Independent benchmarks (MMLU, GSM8K) show favorable comparison against competing abliteration tools - Reproducible models with the Heretic Grimoire: a portable library of reproduction manifests - Pluggable scorer and modifier plugins for custom behavior
Use cases
- Removing censorship from language models - Reducing slop (generic, low-quality text)
- Reducing humor - Reducing sycophancy (excessive agreement)
- Altering emotional alignment - Any other behavior modification achievable through contrasting prompts and scorers
Getting started
sh pip install -U heretic-llm heretic Qwen/Qwen3-4B-Instruct-2507 Replace the model identifier with any model you want to process. By default, Heretic performs censorship removal.
License and pricing
License: AGPL-3.0. Pricing: open source.
Based on the project's public README and repository metadata.