CARES

CARES是由北卡罗来纳大学教堂山分校等机构创建的综合性医疗视觉语言模型信任度评估数据集。该数据集包含约41,000个问题-答案对，覆盖16种医疗图像模式和27个人体解剖区域。数据集通过整合多个开源医疗视觉语言和图像分类数据集构建而成，旨在评估医疗大型视觉语言模型在信任度方面的表现，包括事实性、公平性、安全性、隐私性和鲁棒性。CARES的应用领域主要集中在提高医疗诊断的准确性和可靠性，解决现有模型在实际应用中的信任问题。

CARES is a comprehensive medical vision-language model trustworthiness evaluation dataset developed by the University of North Carolina at Chapel Hill and other institutions. This dataset comprises approximately 41,000 question-answer pairs, covering 16 medical imaging modalities and 27 human anatomical regions. Built by integrating multiple open-source medical vision-language and image classification datasets, it is designed to assess the performance of medical large-scale vision-language models across core trustworthiness dimensions, including factuality, fairness, safety, privacy, and robustness. The primary application scenarios of CARES center on improving the accuracy and reliability of medical diagnosis, and addressing the trust-related challenges faced by existing models in real-world clinical deployments.