A parser can perform well on isolated pages and still lose the thread across a hundred-page file. The key challenge here is continuity: text blocks and tables may span pages, so the benchmark tests whether systems preserve structure and relationships over long documents, not just whether they recognize content on a single page.
What Sets It Apart
- The 100-document collection spans 100–300 pages per document, with roughly 18,000 pages and 205,000 annotated layout elements. That scale makes document-level failures visible rather than averaging them away in short samples.
- Annotations capture bounding boxes, block types, hierarchy, reading order, and cross-page continuations. This lets evaluation measure structural fidelity as well as transcription.
- The benchmark reuses MPDocBench-Parse and its eight metrics, including cross-page text and table measures. Results can therefore distinguish a readable transcription from one that also preserves document organization.
- The pages mix Chinese and English, providing a multilingual parsing test rather than a single-language OCR score.
Great Fit If…
Use it to compare end-to-end parsers on long reports, books, or other multi-page documents where tables and layout continuity matter. Look elsewhere if you need a lightweight OCR test, broad language coverage, or unrestricted redistribution: the images are provided for academic and research use, and the evaluation is computationally demanding, especially for very large tables.
