README.md

metadata

license: cc-by-4.0
task_categories:
  - object-detection
tags:
  - document-layout
  - layout-analysis
  - coco
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: test
        path: manifest.csv

RealDocBench-Layout

A 1,500-page document-layout benchmark for evaluating layout-detection models on real-world documents. COCO-style annotations across 9 block classes.

Contents

• images/ — 1,500 page images (PNG / JPG / occasional WebP-as-PNG; see Caveats).
• annotations/<pageId>.json — per-page COCO files, each with a single image record, an annotations list, a categories list, and a page_info block.
• manifest.csv — pageId → domain + source URLs. The canonical row list.

Block classes (9)

text, heading, section_heading, header, footer, page_number,
figure, table, key_value

class	what it covers
text	body text and other inline content
heading	document title
section_heading	section / subsection titles
header	running page header
footer	running page footer
page_number	page-number indicators
figure	images, diagrams, charts, barcodes, QR codes
table	tabular data
key_value	form fields / key-value pairs

The companion benchmark loader at extend-hq/realdoc-bench walks each annotation file and produces 9-class block predictions; see realdoc_bench/layout/normalizers/coco.py.

Loading

from huggingface_hub import snapshot_download
path = snapshot_download(repo_id="Extend-AI/RealDocBench-Layout", repo_type="dataset")

Caveats

Mixed image formats. 197 of the 1,500 images carry a .png extension but their actual byte stream is WebP (171), GIF (23), or BMP (3). They render correctly in PIL / browsers, but services that validate the magic bytes (AWS Textract, Azure Document Intelligence) will reject them without a transcode step. Re-encoding to true PNG with dimensions preserved is a lossless one-liner.

License

Annotations: CC-BY-4.0. Page images carry their original per-source licenses; the manifest's sourceUrl column points back to each one.