JADE-DB is a curated dataset designed for the targeted safety evaluation of Large Language Models (LLMs), particularly focusing on Chinese open-source and international commercial models. It addresses the need for robust safety testing by transforming low-trigger-rate seed questions into high-risk queries through linguistic variation, covering categories like core values, illegal activities, rights infringement, and discrimination.

How It Works

The project employs linguistic variation techniques to automatically generate a diverse set of high-risk test cases from a smaller set of seed prompts. This approach aims to create natural language datasets that are effective in probing LLM safety alignment across various categories and sub-categories, enhancing the scalability and comprehensiveness of safety evaluations.

Quick Start & Requirements

• Data Access: Datasets are available as CSV files (e.g., jade_benchmark_easy_zh.csv, jade_benchmark_medium_zh.csv, jade_benchmark_zh.csv, jade_benchmark_en.csv).
• Demo: Interactive demos are available at https://whitzard-ai.github.io/jade-demo.html.
• Contact: For politically sensitive test cases, contact mi_zhang@fudan.edu.cn.

Highlighted Details

• JADE 5.0: Text-to-Image model safety evaluation.
• JADE 4.0: Safety regulation RAG.
• JADE 3.0: LLM safety alignment.
• Includes datasets for both Chinese and English LLMs.

Maintenance & Community

• Project is led by Fudan University BaiZe Intelligent.
• Technical reports available in Chinese and English.
• Project website: https://whitzard-ai.github.io/jade.html.

Licensing & Compatibility

• The README does not explicitly state a license for the datasets.
• Users are encouraged to cite the technical report.

Limitations & Caveats

• Publicly available datasets exclude politically sensitive test cases due to regulations.
• Datasets contain examples of harmful and illegal content for evaluation purposes.