MR2-Bench
收藏资源简介:
MR2-Bench是一个用于多模态检索的推理密集型基准数据集,包含1309个精心设计的查询,这些查询是通过手动收集和标注或从公共数据集中选择性整合而来。该数据集涵盖了多种多模态数据类型,如自然图像、图表和视觉谜题,并支持复杂查询和包含多个图像的文档,更准确地反映了现实世界的应用场景。数据集将推理能力作为关键评估标准,旨在推动多模态检索领域的发展。
MR2-Bench is a reasoning-intensive benchmark dataset for multimodal retrieval. It contains 1,309 meticulously designed queries, which are either manually collected and annotated or selectively integrated from public datasets. The dataset covers a variety of multimodal data types such as natural images, charts and visual puzzles, and supports complex queries and multi-image documents, which more accurately reflects real-world application scenarios. The dataset takes reasoning ability as a core evaluation criterion, aiming to promote the development of the multimodal retrieval field.




