DocAtlas-Bench

DocAtlas-Bench is the evaluation benchmark of DocAtlas: Multilingual Document Understanding Across 80+ Languages. It contains the 5,575 difficulty-stratified page-level samples of the benchmark, covering 82 languages, with ground truth for end-to-end page parsing: text, tables, and reading order.

Benchmark

The benchmark balances diversity, difficulty, and representativeness. Pages are embedded with ResNet-50 features, clustered via FAISS, and stratified by difficulty into equal easy/medium/hard splits, yielding up to 100 pages per language across 82 languages (5,575 samples). In the paper, it is complemented by 144 curated formula samples and multilingual chart data across 15 languages, for 5,862 pages and 9 evaluation tasks in total.

This repository contains the 5,575 page-level samples; the formula and chart subsets are not part of it.

Ground truth comes from the DocAtlas annotation pipelines (differential rendering of native DOCX documents and LaTeX-based synthesis for right-to-left scripts), which produce structural annotations without learned models for core annotation. Pages are further annotated with page-level attributes such as document type, column layout, watermark presence, and background color, predicted via Qwen3-VL and confirmed by human inspection; these attributes enable controlled evaluations under specific visual or layout conditions.

Dataset structure

A single test split with one row per page.

Column	Type	Description
id	string	Unique page identifier. Predictions are matched to pages by this id.
image	image	Rendered page image.
language	string	ISO 639 language code of the page (e.g., en, ar, zh, hi), as in docatlas_instruct.
language_name	string	Language name used in the evaluation ground truth (e.g., english, arabic).
data_source	string	Document type: academic_literature, book, colorful_textbook, exam_paper, magazine, newspaper, note, research_report, or PPT2PDF.
markdown	string	Page content in Markdown, with tables as HTML. Same format as in docatlas_instruct.
doctags	string	Page annotation in the DocTag format, with bounding boxes in image pixels. Same format as in docatlas_instruct; null for 225 pages.
annotation	string	JSON string with the page ground truth used by the evaluation code (see below).

DocAtlas-Bench.json at the root of the repository holds the same ground truth as the annotation column, as one JSON list, and is the file the evaluation code reads.

Ground-truth format

The ground truth follows the OmniDocBench annotation format:

{
  "page_info": {
    "page_no": 10, "height": 2526, "width": 1786,
    "image_path": "images/<id>.jpg",
    "page_attribute": {"language": "english", "data_source": "book", "layout": "single_column", "special_issue": []}
  },
  "layout_dets": [
    {"category_type": "text_block", "poly": [x1, y1, x2, y1, x2, y2, x1, y2], "order": 0, "text": "...", "attribute": {"text_language": "text_english", "text_background": "white", "text_rotate": "normal"}},
    {"category_type": "table", "poly": [...], "order": 1, "html": "<table>...</table>", "attribute": {"with_span": false, "line": "wireless_line", "...": "..."}}
  ],
  "extra": {"relation": []}
}

Element types are text_block, section_header, and table. poly gives the four corners of the element in image pixels and order its position in the reading order.

Usage

import json
from datasets import load_dataset

bench = load_dataset("ahmedheakl/DocAtlas-Bench", split="test")

page = bench[0]
print(page["id"], page["language"], page["data_source"])
print(page["markdown"])                      # Markdown ground truth
elements = json.loads(page["annotation"])["layout_dets"]
page["image"].show()

Evaluation

Following the paper, each model produces a Markdown output per page, which is evaluated with a three-stage pipeline: extraction (tables, formulas, and paragraphs), matching of predictions to ground truth using Normalized Edit Distance, and metric computation, with breakdowns by language, document type, and layout.

Metric	Description
Text Edit ↓	Normalized edit distance between predicted and ground-truth text.
Table TEDS ↑	Tree-Edit-Distance-based Similarity between predicted and ground-truth tables, converted to HTML.
Read Order Edit ↓	Edit distance between the predicted and ground-truth reading order.
Overall ↑	Average of text and table scores, after converting text edit distance to accuracy.

The evaluation code is available at github.com/ahmedheakl/DocAtlas:

git clone https://github.com/ahmedheakl/DocAtlas && cd DocAtlas/eval
pip install -r requirements.txt

python prepare_benchmark.py --out data                    # ground truth + page images
# write one predictions/<model>/<id>.md per page, then:
python run_eval.py --gt data/DocAtlas-Bench.json --pred predictions/<model> --out results

Results

Quantitative comparison across OCR systems reported in the paper. Overall is the average of text and table scores, after converting text edit distance to accuracy.

Type	Method	Params	Text Edit ↓	Table TEDS ↑	Formula Edit ↓	Read Order Edit ↓	Overall ↑
General VLMs	Gemini-2.0-Pro	-	0.090	68.50	0.356	0.050	79.75
	GPT4o	-	0.117	62.26	0.425	0.065	75.30
	Qwen3-VL	3B	0.081	51.86	0.420	0.089	71.87
	Qwen2.5-VL	2B	0.174	50.59	0.453	0.117	66.59
	InternVL3.5	2B	0.095	70.80	0.543	0.060	77.20
Expert VLMs	DotsOCR	3B	0.068	65.40	0.321	0.037	79.29
	PaddleOCR-VL	1B	0.078	73.90	0.241	0.052	80.10
	DeepseekOCR	3B	0.082	71.54	0.242	0.053	81.66
	MonkeyOCR-pro	1.2B	0.095	72.80	0.295	0.065	78.25
	Dolphin	400M	0.160	58.30	0.465	0.066	71.17
	Nanonets-OCR-s	4B	0.088	71.90	0.518	0.059	81.53
	Nanonets-OCR2	3B	0.088	66.24	0.471	0.060	78.70
	Chandra	9B	0.071	69.79	0.262	0.042	81.33
	MinerU2.5	1.2B	0.267	72.79	0.273	0.096	73.07
	DocAtlas-Deepseek (Ours)	3B	0.055	72.24	0.237	0.049	83.37

Citation

@article{heakl2026docatlas,
  title={DocAtlas: Multilingual Document Understanding Across 80+ Languages},
  author={Heakl, Ahmed and Mohamed, Youssef and Sohail, Abdullah and Elbadry, Rania and Nassar, Ahmed and Staar, Peter WJ and Khan, Fahad Shahbaz and Razzak, Imran and Khan, Salman},
  journal={arXiv preprint arXiv:2605.12623},
  year={2026}
}

License

This dataset is released under the CC-BY 4.0 license. All source documents were collected from publicly accessible web content and academic repositories under permissive licenses, with automated PII detection applied to remove sensitive content.