MMhops
收藏资源简介:
MMhops是由中科院自动化所与腾讯联合构建的大规模多模态多跳推理基准数据集,涵盖31,117条样本,包含28,256张图像和8,832个实体。该数据集通过维基百科知识库构建,采用GPT-4o等模型进行自动化标注,包含桥接推理和对比推理两种任务类型,支持3-4跳动态推理链。其创新性在于首次系统整合视觉与文本模态的多跳推理,要求模型通过跨模态知识检索和动态路径规划解决复杂问题,为多模态大语言模型的深度推理能力评估提供了重要基准。
MMhops is a large-scale multimodal multi-hop reasoning benchmark dataset jointly developed by the Institute of Automation of the Chinese Academy of Sciences and Tencent. It contains 31,117 samples, including 28,256 images and 8,832 entities. Constructed based on the Wikipedia knowledge base, this dataset adopts models such as GPT-4o for automated annotation, and features two task types: bridging reasoning and comparative reasoning, supporting 3- to 4-hop dynamic reasoning chains. Its core innovation lies in the first systematic integration of multi-hop reasoning across visual and textual modalities, requiring models to solve complex problems through cross-modal knowledge retrieval and dynamic path planning, thus providing a critical benchmark for evaluating the deep reasoning capabilities of multimodal large language models.
MMhops-R1 数据集概述
数据集名称
MMhops-R1
核心描述
MMhops-R1 是一个用于多模态多跳推理的数据集。
当前状态
数据集与相关代码即将发布。
备注
根据提供的信息,该数据集详情页面目前未包含更详细的技术规格、数据构成、规模或获取方式。




