BuDDIE

BuDDIE是由J.P. Morgan AI Research创建的一个包含1665份真实世界商业文档的多任务数据集。该数据集特别之处在于它支持文档分类、关键实体提取和视觉问答等多个任务。文档来源于美国州政府网站，涵盖多种类型如表格、证书、报告等，具有丰富的格式和布局变化。数据集的创建过程涉及文档的收集、OCR文本提取和详细的标注工作，旨在为视觉丰富的文档理解领域提供一个全面的研究平台，推动模型在处理多样化和复杂文档任务上的性能提升。

BuDDIE is a multi-task dataset comprising 1,665 real-world business documents, developed by J.P. Morgan AI Research. A key feature of this dataset is its support for multiple downstream tasks including document classification, key entity extraction, and visual question answering (VQA). All documents are sourced from U.S. state government websites, covering diverse document types such as tables, certificates, reports and more, and exhibit rich variations in format and layout. The dataset construction process involves document collection, OCR text extraction, and meticulous manual annotation. It is intended to serve as a comprehensive research platform for the field of visually-rich document understanding (VRDU), and advance the performance of models when tackling diverse and complex document-related tasks.