Practical overview
When it fits
Use Heretic when you have a compatible local model environment and want an automated transformation and comparison workflow rather than manually selecting abliteration parameters.
Who it is for
- Developers running language models locally
- Researchers comparing model behavior before and after abliteration
- Users investigating refusal reduction in transformer-based models
Documented uses
- Modify a supported language model through an automated abliteration process
- Search for parameters that reduce refusals while minimizing KL divergence from the original model
- Compare refusal and divergence metrics between source and transformed models
- Experiment with transformer architectures in a self-hosted workflow
Requirements & limits
The project is public and licensed under AGPL-3.0. The documented workflow uses Python 3.10+, PyTorch 2.2+ and a command-line installation of the heretic-llm package.
- Model architecture matters: the project documents support for most dense models, many multimodal models, several MoE architectures and some hybrid models, while pure state-space and certain research architectures are not
- PyTorch version can affect model coverage; the repository identifies PyTorch 2.2 as the minimum and notes later requirements for some configurations.
- The workflow is self-hosted and depends on hardware suitable for the selected model.
- Evaluation values can vary by platform and hardware.
- The documented optimizer uses Optuna and seeks lower refusals alongside lower KL divergence from the original model.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Install Heretic and prepare a model transformation
Set up the documented Python and PyTorch environment, install Heretic, and provide a Hugging Face model identifier.
Before you start
- Python 3.10 or newer
- PyTorch 2.2 or newer, with later versions potentially required by particular models or configurations
- Hardware appropriate for the selected model
Steps
Prepare a Python 3.10+ environment with an appropriate PyTorch installation.
Install the package with the command below.
pip install -U heretic-llmRun the documented command-line form with a model identifier, such as
heretic Qwen/Qwen3-4B-Instruct-2507.Replace the example identifier with the model intended for transformation.
Expected outcome
Heretic begins its automatic model-modification workflow using the supplied model identifier.
Guide limitations
- Pure state-space models and certain other research architectures are not supported out of the box.
- The supplied documentation does not establish a universal hardware requirement or guarantee successful execution for every model.
- Some models and configurations may require a later PyTorch version than the documented minimum.
Run the documented model comparison evaluation
Use Heretic’s documented evaluation form to compare a source model with a Heretic-generated model.
Before you start
- An installed Heretic environment
- A source model identifier
- A generated or published comparison model identifier
Steps
Provide the source model with the
--modeloption.Provide the comparison model with the
--evaluate-modeloption.Run the documented form, such as
heretic --model google/gemma-3-12b-it --evaluate-model p-e-w/gemma-3-12b-it-heretic.Review refusal and KL-divergence results, taking platform and hardware dependence into account.
Expected outcome
The evaluation reports the documented behavioral and divergence measures for the model pair.
Guide limitations
- The repository says exact values may depend on platform and hardware.
- Automated benchmarks and mathematical metrics are not substitutes for human evaluation.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 2 October 2026
- Original website title
- p-e-w/heretic · Official homepage title · retained in its source language