rajpurkarlab/ReXInTheWild

--- pretty_name: ReXInTheWild license: cc-by-nc-nd-3.0 task_categories: - visual-question-answering - question-answering configs: - config_name: default data_files: - split: eval path: data/qa.csv --- # ReXInTheWild: A Unified Benchmark for Medical Photograph Understanding ## Overview ReXInTheWild is an expert-verified VQA benchmark for evaluating vision–language models on **medical photographs taken with ordinary cameras**. Unlike traditional medical VQA datasets focused on specialized imaging (e.g., X-rays, pathology), this dataset targets **everyday photographs containing medically relevant content**, requiring both natural image understanding and clinical reasoning. The dataset contains 955 clinician-verified multiple-choice questions about 484 images. Images were selected from the Biomedica dataset, a collection of PubMed Central images. If you use this dataset, please cite the associated paper: *ReXInTheWild: A Unified Benchmark for Medical Photograph Understanding*. ## Dataset Structure - `data/qa.csv`: question-answer pairs and metadata - `data/images/`: corresponding images Each row includes: - `file_name`: relative path to image - `question` - `choice_a`–`choice_e` (3-5 answers per question) - `answer` - `tag`: clinical category (Head & Neck, Trunk & Extremities, etc.) - article metadata (title, authors, link) ## Licensing This dataset is released under **CC BY-NC-ND 3.0**. Images are derived from the noncommercial split of the PubMed Central Open Access Subset and carry individual license restrictions (CC-BY-NC, CC-BY-NC-SA, or CC-BY-NC-ND). We release the dataset under CC-BY-NC-ND, the most restrictive license. ## Bibliography Lozano, A., Sun, M.W., et al.: BIOMEDICA: An open biomedical image-caption archive, dataset, and vision-language models derived from scientific literature (2025)