Release version: v1.2
Canonical dataset version: v1.2
MPIB is a benchmark for evaluating the safety and robustness of medical Large Language Models (LLMs) against prompt injection attacks. It contains 9,697 clinically grounded benchmark instances derived from MedQA and PubMedQA, including benign baselines and adversarial variants.
train (80%): 7,759 samples… See the full description on the dataset page:
https://huggingface.co/datasets/jhlee0619/mpib.
