MMSEARCH
收藏资源简介:
MMSEARCH数据集由香港中文大学多媒体实验室、字节跳动、上海人工智能实验室等机构联合创建,旨在评估大型多模态模型在多模态搜索中的潜力。该数据集包含300个精心收集的实例,涵盖14个子领域,确保与当前LMMs的训练数据无重叠。数据集内容包括最新的新闻和稀有的知识查询,涉及图像和文本的混合信息。创建过程结合了Google Lens和网页截图技术,确保信息的全面性和准确性。MMSEARCH数据集主要应用于多模态AI搜索引擎的开发和评估,旨在解决传统文本搜索在处理复杂、交错的多模态查询时的局限性。
The MMSEARCH dataset was jointly created by multiple institutions including the Multimedia Laboratory of The Chinese University of Hong Kong, ByteDance, Shanghai AI Laboratory, and others. Its core objective is to evaluate the potential of large multimodal models (LMMs) in multimodal search scenarios. This dataset consists of 300 meticulously curated instances spanning 14 sub-domains, with strict guarantees of no overlap with the training data of existing LMMs. The content of MMSEARCH covers cutting-edge news queries and rare knowledge inquiries, involving hybrid information combining both images and text. During its construction, Google Lens and webpage screenshot technologies were integrated to ensure the comprehensiveness and accuracy of the dataset content. Primarily, the MMSEARCH dataset is utilized for the development and evaluation of multimodal AI search engines, aiming to mitigate the limitations of traditional text-based search when dealing with complex, interleaved multimodal queries.

- 1MMSearch: Benchmarking the Potential of Large Models as Multi-modal Search Engines香港中文大学多媒体实验室、字节跳动、香港中文大学MiuLar实验室、上海人工智能实验室、北京大学、斯坦福大学、商汤科技 · 2024年



