| license: apache-2.0 | |
| library_name: transformers | |
| pipeline_tag: text-generation | |
| # Safety Alignment as Continual Learning: Mitigating the Alignment Tax via Orthogonal Gradient Projection | |
| This model is the official implementation of the paper [Safety Alignment as Continual Learning: Mitigating the Alignment Tax via Orthogonal Gradient Projection](https://arxiv.org/abs/2602.07892). | |
| **OGPSA** (**O**rthogonal **G**radient **P**rojection for **S**afety **A**lignment) is a method that preserves general capabilities during safety alignment via an orthogonal gradient projection strategy, balancing safety with general utility. It estimates a low-rank reference subspace from gradients on a small set of general-capability data and removes from each safety gradient the component lying in this subspace. | |
| ## Resources | |
| - **Paper:** [https://arxiv.org/abs/2602.07892](https://arxiv.org/abs/2602.07892) | |
| - **Code:** [https://github.com/SunGL001/OGPSA](https://github.com/SunGL001/OGPSA) | |
| ## Citation | |
| If you find this model or dataset useful in your research, please cite our paper: | |
| ```bibtex | |
| @article{sun2026safety, | |
| title={Safety alignment as continual learning: Mitigating the alignment tax via orthogonal gradient projection}, | |
| author={Sun, Guanglong and Zhang, Siyuan and Wang, Liyuan and Zhu, Jun and Su, Hang and Zhong, Yi}, | |
| journal={arXiv preprint arXiv:2602.07892}, | |
| year={2026} | |
| } | |
| ``` |