Y-J-Ju/MIRe_ViD2R

MIRe多模态检索预训练数据集包含了图像和文本查询-段落对，用于训练不需要在对齐阶段融合文本特征的多模态检索系统。数据集由简洁的问答对转换成扩展段落构成，更好地模拟了现实世界中的检索任务。数据集包含了来自ST-VQA、TextVQA、LLaVAR、Instruct4V和LLaVA-1.5等多个来源的样本，保证了多样性，并排除了WiT数据。The MIRe Pre-training Dataset for Multimodal Query Retrieval consists of image-text query-passage pairs for training multimodal retrieval systems that do not fuse text features during the alignment stage. The dataset is composed of concise question-answer pairs converted into extended passages, better mimicking real-world retrieval tasks. It includes samples from sources like ST-VQA, TextVQA, LLaVAR, Instruct4V, and LLaVA-1.5, ensuring diversity and excluding WiT data.