thu-coai/Safety-Prompts

Contains 100k Chinese safety prompts across 7 typical scenarios (insult, discrimination, crimes, physical harm, mental health, privacy, ethics) and 6 instruction attack types, paired with ChatGPT responses for training safer models. Data is accessible via Hugging Face Datasets and organized in JSON format, designed primarily for fine-tuning rather than evaluation—the project recommends SafetyBench for benchmarking. Complements the broader Safety-Prompts ecosystem including ShieldLM (a customizable safety detector) and provides integration with Tsinghua's Chinese LLM safety evaluation platform.

1,135 stars. No commits in the last 6 months.