ibm-research/REAL-MM-RAG_TechSlides

REAL-MM-RAG-Bench是一个现实世界多模态检索基准，包含了多种模态的文档，包括文本、表格和图像，用于评估模型在处理自然语言查询时的检索能力。数据集中的查询是通过视觉语言模型生成，并经过大型语言模型过滤和重写，以模拟真实世界的检索场景。此外，数据集还采用了多级别查询重写，以测试模型在语义理解方面的鲁棒性。

REAL-MM-RAG-Bench is a real-world multi-modal retrieval benchmark that includes documents with a variety of modalities such as text, tables, and images, designed to evaluate the retrieval capabilities of models when handling natural language queries. The queries in the dataset are generated by a vision-language model and filtered and rephrased by a large language model to simulate real-world retrieval scenarios. Additionally, the dataset employs multi-level query rephrasing to test the robustness of models in semantic understanding.