Oriel — AI & technology worth noticing, every day.

tencent/DiagSafe

Dataset · https://huggingface.co/datasets/tencent/DiagSafe

DiagSafe Beyond Binary Labels: Diagnostic Evaluation of LLM Safety with Prompt-Level Rubrics Paper (arXiv) · GitHub · Evaluation code · Dashboard Content warning. This dataset contains harmful requests (e.g. violence, sexual content, self-harm, child safety, CBRN) written to test LLM safety. It is intended for safety evaluation and research only. DiagSafe is a diagnostic safety benchmark for large language models. It pairs 3,489 bilingual prompts, grounded in an 84-label… See the full description on the dataset page: https://huggingface.co/datasets/tencent/DiagSafe.

First seen by Oriel on .

AI summary

DiagSafe is a bilingual benchmark for diagnosing LLM safety with prompt-specific rubrics that require item-level judgments and evidence instead of binary labels.

  • The dataset contains 3,489 prompts: 1,749 English and 1,740 Chinese, including 2,560 single-turn and 929 multi-turn examples.
  • Prompts span 11 first-level, 34 second-level, and 84 third-level safety categories, covering areas including cybersecurity, CBRN, child safety, and misinformation.
  • Each rubric includes 2–6 prompt-specific safety items with exemptions, eight usefulness items, a 0–3 usefulness score, and a JSON output contract.
  • DiagAttack embeds harmful intent in benign-looking tasks using role, format, and tone constraints, while DiagBase varies language, dialogue length, and attack form.
  • Rubric-based judging achieved 89.41%–95.34% agreement with human annotations across judges, but results depend on the judge model.

#llm safety #benchmarking #safety evaluation #llm judges

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Datasets: 9, latest observation ; 0 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Datasets: 221, latest observation ; 0 over observed period since

Aliases

SourceExternal IDURL
Hugging Face Datasetstencent/DiagSafehttps://huggingface.co/datasets/tencent/DiagSafe Read discussion