The jailbreak-evaluation is an easy-to-use Python package for language model jailbreak evaluation. The jailbreak-evaluation is designed for comprehensive and accurate evaluation of language model jailbreak attempts. Currently, jailbreak-evaluation support evaluating a language model jailbreak attempt on multiple metrics: Safeguard Violation and Relative Truthfulness.
This is the official package repository for "Rethinking How to Evaluate Language Model Jailbreak", by Hongyu Cai, Arjun Arunasalam, Leo Y. Lin, Antonio Bianchi, and Z. Berkay Celik.
Note on PyTorch and FastChat Installation: The jailbreak-evaluation depends on PyTorch and FastChat. To streamline the installation, jailbreak-evaluation does NOT install these libraries for you.
It is recommended to use pip for installation. Please make sure the latest version is installed, as jailbreak-evaluation is updated frequently:
python -m pip install --upgrade pip
# Please install the PyTorch: https://pytorch.org/get-started/locally/
# Please install the FastChat: https://github.com/lm-sys/FastChat?tab=readme-ov-file#install
# Install the jailbreak-evaluation
python -m pip install -U jailbreak-evaluation
# Initialize the Natural Language Toolkit (NLTK)
python -c "import nltk; nltk.download('punkt')"
Evaluate a Language Model Jailbreak Attempt Using jailbreak-evaluation. You should have an OpenAI API key to use the jailbreak-evaluation, which can be obtained from OpenAI Platform.
from jailbreak_evaluation import MultifacetedEvaluation
# Initialize the jailbreak evaluation
openai_api_key = "your openai_api_key"
multifaceted_evaluation = MultifacetedEvaluation(openai_api_key)
# Construct a jailbreak attempt example
intent = "intent"
response = "response"
# Evaluate the above jailbreak attempt example
result = multifaceted_evaluation.evaluate(intent, response)
print("The evaluation result for all levels:")
print(result)
print("The evaluation result for combination level, which is our best accurate result:")
print(result.combination_level)
Please let us know if you encounter a bug or have any suggestions by filing an issue.
We welcome all contributions from bug fixes to new features and extensions.
We expect all contributions discussed in the issue tracker and going through PRs.
If you use jailbreak-evaluation in a scientific publication, we would appreciate citations to the following paper:
@article{cai2024rethinking,
title={Rethinking How to Evaluate Language Model Jailbreak},
author={Hongyu Cai and Arjun Arunasalam and Leo Y. Lin and Antonio Bianchi and Z. Berkay Celik},
year={2024},
journal={arXiv}
}
The jailbreak-evaluation is developed and maintained by PurSec Lab.
The jailbreak-evaluation uses Apache License 2.0.