MRMR
收藏资源简介:
MRMR是一个专家级的多学科多模态检索基准,包含1502个经过人类专家仔细验证的查询,涵盖了23个领域。与之前的基准相比,MRMR在三个关键方面取得了进步:首先,它挑战了跨多个专业领域的检索系统,能够在不同领域之间进行细粒度的模型比较;其次,查询是推理密集型的,需要更深入地解释图像,例如诊断显微镜幻灯片;此外,还引入了矛盾检索这一新型任务,要求模型识别冲突的概念。与仅限于单个图像或单模态文档的早期基准不同,MRMR提供了一个具有多图像查询和混合模态语料库文档的现实场景。
MRMR is an expert-level, multidisciplinary and multimodal retrieval benchmark, comprising 1502 queries carefully validated by human experts and spanning 23 domains. Compared with prior retrieval benchmarks, MRMR advances in three key aspects: First, it challenges retrieval systems across multiple specialized domains, enabling fine-grained model comparisons between different fields; Second, the queries are reasoning-intensive, requiring in-depth interpretation of images such as diagnostic microscope slides; Furthermore, a novel task of contradiction retrieval is introduced, which requires models to identify conflicting concepts. Unlike earlier benchmarks that are limited to single images or unimodal documents, MRMR presents a realistic scenario with multi-image queries and mixed-modal corpus documents.




