🪝PISCES - Precise In-parameter Suppression for Concept EraSure

Code for the paper "Precise In-Parameter Concept Erasure in Large Language Models" (link).

Files

The codebase is still being finalized, but for now I uploaded the main files used in the paper. These are:

• evals.py - includes code for running all of our evaluations, as well as wrapper classes for making the API access to transformers and transformer_lens models identical.
• feature_finder.py - used to find the SAE features that will be used to unlearn a concept. Currently there's no demo of this (planning on adding soon), but all of the code is there.
• editor.py - this includes the code used to perform the unlearning. The most important function there is unlearn_concept, which receives a Concept object which includes the relevant features and hyperparameters to unlearn.

We also have a demo for erasing the Harry Potter concept (erasing_harry_potter.ipynb), which shows how to use the framework. Note that the editing code currently only works with the transformer_lens library.

In data/cvs.json you can find all of the data used in all of our evaluations.

Citation

Please cite as:

@misc{gurarieh2025preciseinparameterconcepterasure,
      title={Precise In-Parameter Concept Erasure in Large Language Models}, 
      author={Yoav Gur-Arieh and Clara Suslik and Yihuai Hong and Fazl Barez and Mor Geva},
      year={2025},
      eprint={2505.22586},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.22586}, 
}

Feel free to contact if you have any thoughts, questions or suggestions :)