stockmark/OmniDocBench-JASyn

OmniDocBench-JASyn是一个用于评估日语文档解析性能的数据集。它继承了OmniDocBench的格式，并评估视觉语言模型在日语文档上的端到端性能，涵盖OCR、布局分析、表格分析、公式OCR和阅读顺序识别。该数据集包含518张文档图像，覆盖14种文档类型，包括日语特有的垂直文本和混合垂直/水平文本格式。数据通过基于大语言模型的代码生成完全合成，并应用了打印污渍、倾斜、褶皱和JPEG退化等机械噪声以增加视觉多样性。注释格式与OmniDocBench兼容，提供五种类型的边界框（标题、文本块、孤立方程、图表、表格）及其内容，以及阅读顺序定义。数据集专注于端到端评估，仅包含块级注释和图像，不提供PDF数据。

OmniDocBench-JASyn is a dataset for evaluating Japanese Document Parsing performance. It inherits the format of OmniDocBench and evaluates the end-to-end performance of VLMs on Japanese documents, covering OCR, layout analysis, table analysis, formula OCR, and reading order recognition. This is an evaluation dataset of 518 document images covering 14 document types, including Japanese-specific formats such as vertical-text documents and mixed vertical/horizontal text documents. The data is fully synthetic, generated via LLM-based code rendering with applied mechanical noise like print stains, skew, wrinkles, and JPEG degradation to enhance visual diversity. The annotation format is compatible with OmniDocBench, providing five types of bounding boxes (title, text_block, equation_isolated, figure, table) and their contents, along with reading order definitions. The dataset focuses on end-to-end evaluation, featuring only block-level annotations and images, without PDF data.