> ## Documentation Index
> Fetch the complete documentation index at: https://mintlify.com/p-e-w/heretic/llms.txt
> Use this file to discover all available pages before exploring further.
# Citation & Acknowledgments
> How to cite Heretic and acknowledgments to prior research
If you use Heretic in your research or projects, please cite it appropriately and acknowledge the foundational work that made it possible.
## Citing Heretic
If you use Heretic for your research, please cite it using the following BibTeX entry:
```bibtex theme={null}
@misc{heretic,
author = {Weidmann, Philipp Emanuel},
title = {Heretic: Fully automatic censorship removal for language models},
year = {2025},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/p-e-w/heretic}}
}
```
### Plain Text Citation
For non-academic contexts:
> Weidmann, P. E. (2025). Heretic: Fully automatic censorship removal for language models. GitHub repository. [https://github.com/p-e-w/heretic](https://github.com/p-e-w/heretic)
## Academic References
Heretic builds upon significant prior research in interpretability and safety alignment removal:
### Foundational Papers
**Refusal in Language Models Is Mediated by a Single Direction**
This paper introduced the concept of "abliteration" - removing refusal behaviors by orthogonalizing model weights with respect to a computed "refusal direction."
**Citation:**
```bibtex theme={null}
@article{arditi2024refusal,
title={Refusal in Language Models Is Mediated by a Single Direction},
author={Arditi, Andy and Obsterheide, Oscar and Lam, Avery and Qi, Jiaxin and
Morrison, Aaron and Safeian, Aaquib and Garriga-Alonso, Adri{\`a} and
Duvenaud, David},
journal={arXiv preprint arXiv:2406.11717},
year={2024}
}
```
**Paper:** [arxiv.org/abs/2406.11717](https://arxiv.org/abs/2406.11717)
**Key contributions:**
* Identified that refusal behaviors in LLMs are mediated by a consistent direction in activation space
* Showed that orthogonalizing weights with respect to this direction removes refusal
* Demonstrated the approach preserves model capabilities on harmless prompts
**Jim Lai's Extensions to Abliteration**
Jim Lai (grimjim) developed two important extensions to the original abliteration technique:
### Projected Abliteration
Adjusts refusal directions to only subtract the component orthogonal to the "good" direction, preserving more of the model's intended behavior.
**Article:** [Projected Abliteration](https://huggingface.co/blog/grimjim/projected-abliteration)
**Implementation in Heretic:**
```python theme={null}
# From main.py:448-457
if settings.orthogonalize_direction:
# Implements projected abliteration
good_directions = F.normalize(good_means, p=2, dim=1)
projection_vector = torch.sum(
refusal_directions * good_directions, dim=1
)
refusal_directions = (
refusal_directions - projection_vector.unsqueeze(1) * good_directions
)
refusal_directions = F.normalize(refusal_directions, p=2, dim=1)
```
### Norm-Preserving Biprojected Abliteration
Further refinement that preserves the norm of activations during abliteration.
**Article:** [Norm-Preserving Biprojected Abliteration](https://huggingface.co/blog/grimjim/norm-preserving-biprojected-abliteration)
**Key contributions:**
* Better preservation of model capabilities
* Reduced side effects from abliteration
* Improved balance between refusal removal and intelligence retention
**Maxime Labonne's Contributions**
Maxime Labonne has been a pioneer in practical abliteration implementations and has shared multiple high-quality abliterated models.
### AutoAbliteration
An implementation exploring automation of the abliteration process.
**Article:** [AutoAbliteration](https://huggingface.co/posts/mlabonne/714992455492422)
### Variable Ablation Weights
Labonne's work on [gemma-3-12b-it-abliterated-v2](https://huggingface.co/mlabonne/gemma-3-12b-it-abliterated-v2) explored using non-constant ablation weights across layers, which inspired Heretic's flexible weight kernel approach.
### Educational Content
**Article:** [Abliteration: Making LLMs Say Anything](https://huggingface.co/blog/mlabonne/abliteration)
This comprehensive guide helped popularize abliteration techniques and provided practical insights that informed Heretic's development.
## Acknowledgments
The development of Heretic was informed by the research and implementations listed above, as well as:
### Prior Implementations
Several publicly available implementations of abliteration techniques provided inspiration and insights:
* **[AutoAbliteration](https://huggingface.co/posts/mlabonne/714992455492422)** by Maxime Labonne
* **[abliterator.py](https://github.com/FailSpy/abliterator)** by FailSpy
* **[wassname's Abliterator](https://github.com/wassname/abliterator)** by wassname
* **[ErisForge](https://github.com/Tsadoq/ErisForge)** by Tsadoq
* **[Removing refusals with HF Transformers](https://github.com/Sumandora/remove-refusals-with-transformers)** by Sumandora
* **[deccp](https://github.com/AUGMXNT/deccp)** by AUGMXNT
**Note:** Heretic was written from scratch and does not reuse code from any of these projects. However, examining these implementations provided valuable insights into practical considerations and edge cases.
### Key Technologies
Heretic leverages several excellent open-source projects:
* **[Optuna](https://optuna.org/)** - Hyperparameter optimization framework with TPE sampler
* **[PyTorch](https://pytorch.org/)** - Deep learning framework
* **[Hugging Face Transformers](https://huggingface.co/transformers)** - Model loading and inference
* **[bitsandbytes](https://github.com/TimDettmers/bitsandbytes)** - Quantization support
* **[Rich](https://rich.readthedocs.io/)** - Terminal formatting and progress display
## License
Heretic is free and open-source software released under the **GNU Affero General Public License v3.0 (AGPL-3.0)**.
### What This Means
**You are free to:**
* Use Heretic for any purpose (personal, commercial, research)
* Study how Heretic works
* Modify Heretic to suit your needs
* Distribute Heretic and your modifications
**Under these conditions:**
* You must license your modifications under AGPL-3.0
* You must provide source code for any modifications you distribute
* If you run a modified version as a network service, you must provide the source code to users
* You must preserve copyright and license notices
### Models Generated by Heretic
**Important:** The AGPL-3.0 license applies to Heretic itself (the software), **not** to models you process with Heretic.
Models you create using Heretic:
* Inherit the license of their **base model**
* Are **not** covered by AGPL-3.0
* Can be shared under the base model's license terms
Always respect the original model's license when sharing processed models.
### Full License Text
```
Copyright © 2025-2026 Philipp Emanuel Weidmann + contributors
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU Affero General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU Affero General Public License for more details.
You should have received a copy of the GNU Affero General Public License
along with this program. If not, see .
```
**Full license text:** [GNU AGPL-3.0](https://www.gnu.org/licenses/agpl-3.0.html)
### Contributing
**By contributing to this project, you agree to release your contributions under the same license (AGPL-3.0).**
Contributions are welcome! See the [GitHub repository](https://github.com/p-e-w/heretic) for:
* Issue tracking
* Pull request guidelines
* Development setup instructions
## Contact
**Author:** Philipp Emanuel Weidmann
**Email:** [pew@worldwidemann.com](mailto:pew@worldwidemann.com)
**Project:** [github.com/p-e-w/heretic](https://github.com/p-e-w/heretic)
**Community:** [Discord](https://discord.gg/gdXc48gSyT)
## Recognition
Heretic has been recognized by the community:
* **#1 Repository of the Day** on TrendShift
* Over 1,000 community-created models on Hugging Face
* Active community on Discord with ongoing development and support
Show your support by starring the repository
Connect with the community and get involved