smileysmiley9990/SCORE-Bench-Copy

SCORE-Bench是一个精心策划的集合，包含224份由专家手动注释的多样化、真实世界的文档。它旨在通过真实的生产级挑战来基准测试文档解析系统。与传统学术数据集通常由干净的数字原生PDF组成不同，该基准特别针对企业工作流中发现的复杂性。数据集允许研究人员和开发者超越“干净”的评估，测试系统如何处理现实世界的不规则性。它包括：

* **复杂布局**：具有深度嵌套表格的财务报告、多栏密集文本的技术手册以及以空白（而非线条）定义结构的文章。
* **视觉噪声和多样性**：带有倾斜的扫描表格、带有伪影的复印文档以及包含混合打印和手写文本的表格。
* **语义模糊性**：选择用于打破脆弱系统的文档，要求解析器区分不同的结构解释（例如，识别两栏文章与键值对列表）。

SCORE-Bench中的每份文档都由领域专家手动注释，而非从元数据算法生成。

SCORE-Bench is a curated collection of 224 diverse, real-world documents manually annotated by experts. It is designed to benchmark document parsing systems against true production-grade challenges. Unlike traditional academic datasets often composed of clean, digital-native PDFs, this benchmark specifically targets the complexity found in actual enterprise workflows. The dataset allows researchers and developers to move beyond "clean" evaluation to test how systems handle the irregularities of the real world. It includes:

* **Complex Layouts:** Financial reports with deeply nested tables, technical manuals with multi-column dense text, and articles where whitespace (rather than lines) defines structure.
* **Visual Noise & Variety:** Scanned forms with skew, photocopied documents with artifacts, and forms containing mixed printed and handwritten text.
* **Semantic Ambiguity:** Documents selected to break brittle systems, requiring parsers to distinguish between varying structural interpretations (e.g., identifying a two-column article versus a list of key-value pairs).

Every document in SCORE-Bench has been manually annotated by domain experts, not algorithmically generated from metadata.