Hugging Face's logo

Datasets:
Heliosoph
/
DocBank

Tasks:
Image-to-Text
Object Detection
Languages:
English
Size:
1K<n<10K
ArXiv:
Tags:
docbank
ocr
document-ai
document-layout
layout-analysis
text-recognition
License:
Dataset card Files
xet
 Community
DocBank / README.md
flyingbertman's picture
flyingbertman
Upload folder using huggingface_hub
459bfd7 verified
preview code
|
Raw
History Blame Contribute Delete
6.06 kB
metadata 
license: apache-2.0
pretty_name: DocBank (sampled) — document pages with token-level OCR ground truth
task_categories:
  - image-to-text
  - object-detection
language:
  - en
size_categories:
  - 1K<n<10K
tags:
  - docbank
  - ocr
  - document-ai
  - document-layout
  - layout-analysis
  - text-recognition
  - arxiv
source_datasets:
  - doc-analysis/DocBank

DocBank (sampled) — document pages + exact OCR ground truth

A sampled, reshaped mirror of DocBank (Li et al., COLING 2020), packaged for one-row-per-page ingestion. The upstream release is 54 GB with page images split across a ten-part archive (DocBank_500K_ori_img.zip.001….010) and annotations as a separate 3 GB zip of per-page .txt files. This mirror carries a small cut — 5,000 train + 1,000 test pages — with images repackaged into a single zip per split and the token annotations reshaped into gzipped JSONL, so a pipeline can fetch one split and join image ↔ annotation without reassembling a multi-part archive or parsing thousands of loose text files.

Re-hosted under Heliosoph for ingestion-pipeline stability; the page pixels and token ground truth are DocBank's, unchanged in content. Pin a revision for reproducible fetches.

Credit: Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, Ming Zhou (DocBank). Source documents are LaTeX submissions on arXiv.org.

Why a mirror?

Three frictions make raw DocBank awkward to ingest:

  1. Ten-part split zip. The images ship as *.zip.001 … *.zip.010; you must download all ~54 GB and reassemble before a single page is readable.
  2. Loose per-page text files. Annotations are one .txt per page (tab-separated tokens) inside a 3 GB zip — millions of tiny files.
  3. Scale. 500K pages is far more than a test corpus needs.

This mirror resolves all three: a bounded sample, images as one zip per split, and annotations pre-joined into JSONL (one page per line). The token text and boxes are byte-faithful to DocBank's .txt ground truth.

What this repo contains 

docbank-train-images.zip   # 5,000 page images (PNG/JPG), zip entry name == page_id + ext
docbank-train.jsonl.gz     # 5,000 lines, one JSON object per page
docbank-test-images.zip    # 1,000 page images
docbank-test.jsonl.gz      # 1,000 lines

Annotation JSONL — one page per line 

{
  "page_id": "2004.xxxxx.tar_2004.01234.gz_paper_3_ori",
  "image_file": "2004.xxxxx.tar_2004.01234.gz_paper_3_ori.jpg",
  "width": 612,
  "height": 792,
  "text": "Full-page reading-order text, tokens space-joined ...",
  "tokens": [
    {"text": "Attention", "x0": 120, "y0": 88, "x1": 210, "y1": 104, "label": "title", "font": "NimbusRomNo9L-Medi"}
  ]
}
  • image_file is the exact entry name inside the images zip — join on it directly, no filename parsing.
  • text is every token joined by spaces in reading order — ready for an OCR-vs-truth diff.
  • tokens[] preserves DocBank's per-token ground truth: text, the bounding box (x0,y0)–(x1,y1) normalized to 0–1000, one of twelve semantic labels, and the font.

The twelve semantic labels

title, author, abstract, paragraph, section, list, caption, equation, figure, table, reference, footer.

⚠️ Ground truth is machine-perfect, not human

DocBank's labels come from the LaTeX source, not from reading the pixels. The transcription is exact by construction — but it reflects source tokenization (ligatures, hyphenation, math markup) rather than what a pixel-level OCR model would emit. Evaluate with a fuzzy / normalized-edit-distance metric, not exact string match. Bounding boxes are in a 0–1000 normalized space; multiply by width/1000 and height/1000 to project onto the image.

How to use 

import gzip, json, zipfile, io
from PIL import Image

# annotations
with gzip.open("docbank-test.jsonl.gz", "rt", encoding="utf-8") as f:
    pages = [json.loads(line) for line in f]
print(len(pages), pages[0]["text"][:200])

# images — join by image_file
with zipfile.ZipFile("docbank-test-images.zip") as z:
    img = Image.open(io.BytesIO(z.read(pages[0]["image_file"])))
    print(img.size, "vs", (pages[0]["width"], pages[0]["height"]))

Dataset specs

Spec
Pages 5,000 train / 1,000 test (sampled from DocBank's official splits)
Row granularity one page (image + full text + token array)
Bounding boxes per token, normalized to 0–1000
Semantic labels 12 (title / author / abstract / paragraph / section / list / caption / equation / figure / table / reference / footer)
Domain academic papers (arXiv), English, two-column + equations
Format images as zip per split; annotations as gzipped JSONL
Ground truth machine-exact (from LaTeX source) — use fuzzy scoring

When to pick DocBank

  • Document OCR eval: run a recognizer over the page image, diff against the exact text. Real printed pages, not synthetic, not scene text.
  • Layout / detection: token boxes with semantic labels support detection and reading-order experiments.
  • Document-AI prototyping: per-token text + box + role is the shape LayoutLM-style models consume.

For photographed text in the wild use TextOCR / HierText; for receipts, CORD; for isolated handwritten characters, EMNIST.

License

Apache-2.0, as released by the DocBank authors. Permits commercial use, modification, and redistribution with attribution. The underlying documents are arXiv submissions used under their distribution terms; cite the DocBank paper and arXiv.