heretic: Training-free language model modification.
What problem it solves
Heretic addresses the difficulty of modifying language-model behavior to remove refusals, suppress slop, reduce sycophancy, or apply other behavioral changes. It provides a training-free approach that steers models through mathematical techniques rather than expensive retraining.
Key features
Configurable behavior changes — The default configuration suppresses refusals, while included configurations can reduce slop, humor, or sycophancy, alter emotional alignment, and shorten responses. Users can also supply their own configuration.
Source: README, config.default.toml
Pluggable scorers and modifiers — Behavior is controlled through scorer plugins, and processing uses a selectable modifier. Custom scorers can be loaded by import path or file path without modifying the package.
Source: README, src
Broad model support — The tool supports dense, multimodal, mixture-of-experts, and modern hybrid model architectures that can run on the user's hardware.
Source: README
Reproducible model exports — Processed-model repositories can include human-readable reproduction instructions and a machine-readable manifest. Heretic can use that manifest to guide reproduction and verify exported weight-file hashes.
Source: README
Local model workflow — After processing, users can save the model, upload it, chat with it for testing, run standard benchmarks, or combine these actions.
Source: README
Quick start
Copied verbatim from the project's README.
pip install -U heretic-llm
heretic Qwen/Qwen3-4B-Instruct-2507Who it may suit
Inference: may suit researchers and developers who want to modify locally runnable language models without performing a separate training run.
Topics to discuss
Which scorer combinations should be included as additional standard configurations?
What should the plugin API expose to make custom behavior objectives easier to implement?
How can reproducibility manifests better support varied hardware and software environments?
Written by ReadmeX AI from the project's README, source and public GitHub data; it may contain mistakes.
Read the original README https://github.com/p-e-w/heretic#readme
ReadmeX opened this community for a trending project; it is not run by the project. Maintainers can claim it from the community page.
Comments
…