saillab/vsfmed-v2

VSF-Med v2数据集包含VSF-Med v2（医学视觉语言模型漏洞评分框架）的框架和聚合分析结果。VSF-Med v2是一个经过临床医生验证的对抗性评估框架，专门用于评估医学视觉语言模型（VLMs）的漏洞和安全性。该数据集提供了协议冻结工件（如攻击分类法、评分标准、模型列表等）、试点研究的聚合结果（如统计指标、法官间一致性、严重性分布等）以及数据库模式定义。需要注意的是，数据集本身不包含MIMIC-CXR图像或MIMIC-Diff-VQA的问题/答案对，这些需要从原始来源单独获取。数据集旨在支持研究人员复现和评估医学VLMs在对抗性条件下的脆弱性。VSF-Med v2 dataset contains the framework and aggregated analysis results of the VSF-Med v2 (Medical Visual-Language Model Vulnerability Scoring Framework). VSF-Med v2 is a clinician-validated adversarial evaluation framework specifically designed to assess the vulnerabilities and safety of medical visual-language models (VLMs). This dataset provides protocol-frozen artifacts including attack taxonomies, scoring criteria, model lists and other relevant materials, aggregated pilot study results such as statistical metrics, inter-rater reliability, severity distribution and other related findings, as well as database schema definitions. It should be noted that the dataset itself does not include MIMIC-CXR images or the question-answer pairs from MIMIC-Diff-VQA, which need to be obtained separately from their original sources. The dataset is intended to support researchers in reproducing and evaluating the vulnerability of medical VLMs under adversarial conditions.