license: mit
task_categories:
- question-answering
- summarization
language: - en
tags: - medical
- red-teaming
- healthcare
- HIPPA
- GDPR
- privacy
- bias
- fairness
- hallucination
pretty_name: DAS Medical Red-Teaming Test Suites
size_categories: - 1K<n<10K
DAS Medical Red-Teaming Test Suites
Accompanies the paper Beyond Benchmarks: Dynamic, Automatic and Systematic Red-Teaming Agents for Trustworthy Medical LLMs.
The data samples presented in this repo are used as the initial data seeds and can be mutated further upon requests. It is designed to stress-test Large Language Models (LLMs) in safety-critical medical domains, auditing along four critical axes: Robustness, Privacy, Bias/Fairness, and Hallucination.
Dataset summary
- Robustness— Please refer to the original MedQA dataset.
- Privacy — 81 self-authored scenarios across 8 common HIPAA/GDPR violation types (unauthorized disclosure, misdirected email, accidental social-media sharing, etc.).
- Bias/Fairness — 415 multiple-choice items: 111 clinician-authored + 304 adapted from Omar et al., Sociodemographic biases in medical decision making by large language models, Nature Medicine 2025. Each case combines an objective note and a subjective self-report, supporting identity, linguistic, emotional, and cognitive-bias priming:contentReference[oaicite:2]{index=2}.
- Hallucination — 131 positive prompts curated/cleaned from Chang et al., Red teaming ChatGPT in medicine to yield real-world insights on model behavior, npj Digital Medicine, 2025. Negative “near-ideal” prompts (129 from HealthBench) are not included here; please refer to the original HealthBench repo.
Configurations
privacy: 81 scenarios, labeled with violation categories.bias_fairness: 415 multiple-choice cases with metadata for bias-attack variants.hallucination: 131 positive prompts. Negative prompts not included; see HealthBench.
Licensing
- Privacy suite — fully self-authored. CC-BY-4.0 (or CC-BY-NC-4.0).
- Bias/Fairness suite — mix of self-authored and adapted items from Omar et al., Nature Medicine 2025.
- Hallucination suite — positive set curated from Chang et al., npj Digital Medicine 2025. Negative set from HealthBench; see original repo.
Acknowledgement
We thank all machine learning / medical workers for making public codebase / datasets available to the community. 🫶🫶🫶
If you find our work helpful, feel free to give us a cite.
@article
{pan2025beyond,
title={Beyond Benchmarks: Dynamic, Automatic And Systematic Red-Teaming Agents For Trustworthy Medical Language Models},
author={Pan, Jiazhen and Jian, Bailiang and Hager, Paul and Zhang, Yundi and Liu, Che and Jungmann, Friedrike and Li, Hongwei Bran and You, Chenyu and Wu, Junde and Zhu, Jiayuan and others},
journal={arXiv preprint arXiv:2508.00923},
year={2025}
}