Safety-Prompts

Why we included this project

This is a large Chinese-language collection of about 100k prompts with ChatGPT responses, built to test how safely a large language model behaves. The prompts are labeled by category: seven everyday unsafe scenarios like insults, discrimination, and physical harm, plus six instruction-attack patterns such as goal hijacking and prompt leaking. That labeling makes it easy to score a model's responses and see where it fails, which is handy for anyone red-teaming or aligning a Chinese-facing model. It also works as fine-tuning material for nudging a model toward safer outputs. Note that the authors themselves say the data is best used for training and safety tuning; for formal evaluation they point you to their SafetyBench platform instead.