This repository provides the official implementation. It supports deletion and addition attacks with human-readable rationales, and provides ready-to-run scripts for reproducible experiments. For more details, please refer to our paper.

The formatting is currently messy, and it will be refined after the paper is accepted.

We recommend using Python 3.10+. Higher versions should also be compatible. To install dependencies, run:

pip install -r requirements.txt

The datasets have already been included in kg/. We also provide LoRA weights of two datasets for reproduction.

You can find ready-to-run shell scripts in scripts/.

1. hoa_sft.sh sft for knowledge alignment by triple classfication.
2. llm_del_filter.sh filters the candidate entities.
3. llm_del.sh performs the deletion attack.

1. llm_add_filter.sh filters the candidate entities.
2. llm_add.sh performs the addition attack.

We thank the authors of these open-source projects for their contributions: AttributionAttack, KoPA, KG-LLM