This is the official repo of the paper "Attack Prompt Generation for Red Teaming and Defending Large Language Models" accepted to Findings of EMNLP 2023.
The code runs well under python 3.8.0. The required packages are as follows:
- openai == 0.27.4
- backoff == 2.2.1
- fire == 0.5.0
- transformers == 4.28.1
- peft == 0.3.0
- datasets == 2.11.0
- torch == 1.13.1
We put SAP dataset in ./datasets/.
You can run attack.py to generate your own SAP dataset with Attack Framework. An example command of SAP5 generation can be found in case_generate.sh.
An example script for a fine-tuning iteration can be found in defense_example.sh.
Some parts of this repository are adopted from alpaca-lora, you can find more information in https://github.com/tloen/alpaca-lora. Thanks for the contributions!