ReGA: Representation-Guided Abstraction for Model-based Safeguarding of LLMs

Usage

1. Download LLMs and datasets from huggingface.co and modify their paths in language_model.py and data.py
2. Create folder logs in current path
3. python main.py --fname EXPERIMENT_NAME --model-name all --abs-model rep --test-loaders all --test-data 1000 --threshold all --pca-dim 8 --state-num 32 --harmful-data 64

You can customize the hyperparameters in the arguments.