OCRFlux Bench Single
A multilingual benchmark of 2,000 PDF pages with Markdown annotation to test document parsing capabilities using OCR.
Description
OCRFlux Bench Single is a high quality benchmark designed to assess the performance of OCR systems in parsing PDF pages. It includes 2,000 pages from internal documents, manually annotated in Markdown, English, and Chinese. The objective is to test the fidelity of OCR systems in converting PDF pages to structured formats.
What is this dataset for?
- Evaluate the accuracy of OCRs on complex documents (titles, tables, structured text)
- Test multilingual OCR models on English and Chinese content
- Improve Markdown generation from visual content (PDF, scan)
Can it be enriched or improved?
Yes, the dataset can be extended to other languages or types of documents (e.g. forms, legal documents). Title levels, section distinctions, or fine annotation of tables (thead, tbody) could also enrich the structure.
🔎 In summary
🧠 Recommended for
- OCR developers
- Text extraction researchers
- PDF-to-Html projects
🔧 Compatible tools
- Tesseract
- LMv3 layout
- PaddleOCR
- PymuPDF
- Markdown-it
💡 Tip
Use the HTML version of tables to avoid information loss in complex structures.
Frequently Asked Questions
Can this dataset be used to train an OCR model?
Rather, it is designed for evaluation, but can be used as a basis for fine-tuning on structured Markdown.
What format is used for table annotations?
Tables are represented in HTML in Markdown in order to manage merged cells (rowspan, colspan).
Does the dataset contain documents in languages other than English and Chinese?
No, the current version is limited to English and Chinese. However, it is possible to extend it to other languages with documents annotated according to the same structure.
