RealDocBench
收藏资源简介:
RealDocBench是由Extend AI团队构建的一个面向真实受监管文档的双轨基准数据集,旨在评估文档解析系统在字段级问答和布局理解方面的性能。该数据集包含1356个字段级问题和1500个经过人工验证的页面标注,覆盖抵押贷款、金融、供应链和医疗四大领域,数据来源于真实受监管文档,并采用隐私保护合成技术处理敏感内容。数据集通过严格的问答跟踪和布局跟踪设计,支持类型化字段值提取和九类公共分类法下的边界框标注,主要应用于企业级文档解析系统的性能评估,解决现有基准在文档真实性、评估指标和操作成本等方面的不足。
RealDocBench is a dual-track benchmark dataset for real regulated documents developed by the Extend AI team, which aims to evaluate the performance of document parsing systems in terms of field-level question answering and layout understanding. The dataset contains 1356 field-level questions and 1500 manually verified page annotations, covering four major domains: mortgage, finance, supply chain, and healthcare. It is sourced from real regulated documents, with sensitive content processed using privacy-preserving synthesis techniques. Designed with rigorous question answering tracking and layout tracking mechanisms, the dataset supports typed field value extraction and bounding box annotation under a nine-category public taxonomy. It is primarily used for performance evaluation of enterprise-grade document parsing systems, addressing the limitations of existing benchmarks in terms of document authenticity, evaluation metrics, and operational costs.




