Patcher: Post-Hoc Patching of Backdoored Large Language Models

This project is official implementation from the paper "Patcher: Post-Hoc Patching of Backdoored Large Language Models". A security research framework for localizing and removing backdoor attacks from Large Language Models (LLMs) by patching the models. This project implements the attack, patching, and evaluation pipeline using gradient saliency analysis to identify trigger tokens and patch compromised models. For more details, see the README.md in the files.