SoMarkAI/DocParsingBench

DocParsingBench is a document intelligence benchmark of 1,400 images, systematically collected and annotated from real business workflows. It is the first dataset to systematically catalogue the document elements most frequently encountered in enterprise settings, covering five major domains: finance, legal, scientific research, manufacturing, and education.

---

🆕 Latest Updates

[2026.04.17] DocParsingBench evaluation toolkit release. Unified scoring is now available for the three core elements of document parsing — text, formulas, and tables — together with batch CLI evaluation, segment-level matching, visualization analysis, and leaderboard generation. 📊

[2026.03.09] DocParsingBench dataset release. The first document parsing dataset built for real-world industry scenarios, covering finance, legal, scientific research, manufacturing, and education. Now available on Hugging Face and ModelScope! 🔥🔥🔥

---

🎯 Key Highlights

✅ 1. Five Real-World Industry Domains

Samples were drawn from real business workflows, preserving scan noise, stamp occlusion, and blurred characters:

Domain	Document Types	Characteristics
Finance	Brokerage research reports, listed-company annual reports, prospectuses	Multi-column tables, stamped PDF scans, complex tables
Legal	Legal documents, contract clauses, industry standards	Standardized headers and footers, dense footnotes, cross-referenced citations
Scientific Research	Academic papers, programming textbooks, full-text patents	Mixed two- and three-column layouts, formula-heavy text, code blocks, chemical formulas
Manufacturing	Operating SOPs, forms and receipts	Blurry scans, handwritten fill-ins, QR codes / barcodes
Education	Textbooks across English, chemistry, and mathematics	Chemical structures, reaction equations, multiple-choice options, fill-in-the-blanks

✅ 2. Full Coverage of Complex Layouts

• Single-column: textbooks, legal documents
• Two-column: academic papers, some research reports
• Three or more columns: brokerage financial reports, data reports
• Mixed layouts: interleaved text and figures, tables spanning columns, nested sidebars

---

📊 Dataset Composition

Dimension	Category
Total samples	1,400 pages
Languages	Chinese, English, Chinese-English mixed
Industry distribution	Finance / Legal / Scientific Research / Manufacturing / Education
Layout distribution	Single-column / Two-column / Three-column / Mixed
Annotation format	Markdown
Chemistry annotation	Follows the SoMarkdown specification, combining SMILES with LaTeX to fully render chemical structures

---

🏗️ Annotation Philosophy

• Fully human-annotated, without exception — the "ground truth" of document parsing should not be something a model guesses. In practice, we found that errors from model pre-annotation are silently accepted by annotators: when shown an existing markdown draft, annotators tend to tweak rather than redraw, and the model's mistakes get frozen into the dataset. Every bounding box is drawn stroke by stroke; every character is typed by hand. It is not the fastest approach, but it is the one that yields the highest quality.
• This was the most time-consuming part of the project. We wrote detailed boundary rules for every class of document element, then pilot-annotated, iterated, and finalized them class by class.
• Every spec went through 3+ rounds of pilot annotation, team-wide training, and consistency testing, keeping inter-annotator disagreement on the same page below 5%. In total, 6 full batches were rejected and re-annotated, with an annotation-to-review time ratio of 1 : 0.8.

📈 Evaluation Benchmark

The companion industrial document parsing benchmark is released alongside this dataset — community submissions are welcome! 📊

❤️ Acknowledgements

This is work that demands extraordinary patience. A single complex page mixing chemical formulas and tables can take hours; one boundary disagreement can send an entire batch back for re-annotation, costing hundreds of person-hours. Behind every annotated page is a human eye parsing the document's structure and a human hand confirming the markdown. Thanks to every annotator who saw this through.

We also thank the SoMarkdown project for the chemical-structure annotation specification that let us express LLM-readable chemistry precisely.

---

📖 Citation

If you use DocParsingBench in your research, please cite it as follows:

@misc{DocParsingBench-2026,
  title={DocParsingBench},
  author={SoMark},
  year={2026},
  publisher={Hugging Face,ModelScope},
  howpublished={\url{https://modelscope.cn/datasets/SoMark/DocParsingBench}}
}

---

📜 License

This project is released under the ODC-BY (Open Data Commons Attribution License) and is open to both academic research and commercial use.

---

🌟 If you find this dataset useful, please consider giving us a ⭐ on ModelScope / Hugging Face!

How to Download

🤗 Hugging Face

from datasets import load_dataset
ds = load_dataset("SoMarkAI/DocParsingBench")

☁️ ModelScope

from modelscope.msdatasets import MsDataset
ds = MsDataset.load('SoMark/DocParsingBench')

Git Clone

# Hugging Face
git clone https://huggingface.co/datasets/SoMarkAI/DocParsingBench

# ModelScope
git clone https://www.modelscope.cn/datasets/SoMark/DocParsingBench.git