🌐 Website • 🤗 Hugging Face • ⏬ Data • 📃 Paper
SafetyBench is a comprehensive benchmark for evaluating the safety of LLMs, which comprises 11,435 diverse multiple choice questions spanning across 7 distinct categories of safety concerns. SafetyBench also incorporates both Chinese and English data, facilitating the evaluation in both languages. Please visit our website or check our paper for more details.
🎉 2024/06/24: SafetyBench has been accepted by the ACL 2024 Main Conference. See our camera-ready version of the paper at Arxiv. SafetyBench is also integrated into SuperBench.
The up-to-date leaderboards are on our website. We have three leaderboards for Chinese, English and Chinese subset respectively. We remove questions with highly sensitive keywords and downsample 300 questions for each category to construct the Chinese subset. Summarized evaluation results of some representative LLMs are shown below:
We put our data on the Hugging Face website.
You can download the test questions and few-shot examples through wget directly. Just run the script download_data.sh
Alternatively, you can download the test questions and few-shot examples through the datasets library. Just run the code download_data.py
test_zh, test_en and test_zh_subset contain test questions for Chinese, English and Chinese subset respectively. dev_zh and dev_en contain 5 examples for each safety category, which can be used as few-shot demonstrations.
Note that the options field in the data includes at most four items, corresponding to the options A, B, C, D in order. For the answer field in the dev data, the mapping rule is: 0->A, 1->B, 2->C, 3->D.
In our paper, we conduct experiments in both zero-shot and five-shot settings. And we extract the predicted answers from models' responses. An example of evaluation code could be found at code. We don’t include CoT-based evaluation because SafetyBench is less reasoning-intensive than benchmarks testing the model’s
general capabilities such as MMLU. But feel free to submit your results based on CoT. The default prompt for zero-shot and five-shot evaluation is shown below:
To enable more accurate extraction of the predicted answers, we made minor changes to the prompts for some models, which is shown below:
You need to first prepare a UTF-8 encoded JSON file with the following format, please refer to submission_example.json for details.
## key is the "id" field of the test questions
## value is the predicted answer: 0->A, 1->B, 2->C, 3->D
{
"0": 0,
"1": 1,
"2": 3,
"3": 2
}
Then you can submit the JSON file to our website.
@article{zhang2023safetybench,
title={SafetyBench: Evaluating the Safety of Large Language Models with Multiple Choice Questions},
author={Zhexin Zhang and Leqi Lei and Lindong Wu and Rui Sun and Yongkang Huang and Chong Long and Xiao Liu and Xuanyu Lei and Jie Tang and Minlie Huang},
journal={arXiv preprint arXiv:2309.07045},
year={2023}
}