Multi-Lang-OCR

Synthetic document images with full layout ground truth, in nine languages across six scripts. Built to fine-tune OCR and document-understanding models on documents that are not English — and not only in Latin script.

Every page is composed from real corpus prose, rendered as HTML in headless Chromium, and exported with the bounding box, reading order and text of every block the browser actually laid out. The ground truth is read back off the rendered DOM, so it describes the pixels rather than the intent.

Languages

Code	Language	Script	Tokenization	Text source
ar	Arabic	Arabic (RTL)	word	Shamela book pages
fr	French	Latin	word	Wikipedia 20231101.fr
en	English	Latin	word	Wikipedia 20231101.en
es	Spanish	Latin	word	Wikipedia 20231101.es
de	German	Latin	word	Wikipedia 20231101.de
ru	Russian	Cyrillic	word	Wikipedia 20231101.ru
zh	Chinese (Simplified)	Han	character	Wikipedia 20231101.zh
ja	Japanese	Kana + Han	character	Wikipedia 20231101.ja
hi	Hindi	Devanagari	word	Wikipedia 20231101.hi

Per language: 29,997 original pages in 3 shards, plus 3 augmented variants of each in 9 more. Across all nine that is 269,973 originals and 1,079,892 images.

Loading

from datasets import load_dataset

# one language
ds = load_dataset("OCR-Data/Multi-Lang-OCR", "fr", split="train", streaming=True)

# originals only, no augmented variants
ds = load_dataset("OCR-Data/Multi-Lang-OCR", "originals", split="train", streaming=True)

# everything
ds = load_dataset("OCR-Data/Multi-Lang-OCR", "all", split="train", streaming=True)

sample = next(iter(ds))
image = sample["png"]          # or "jpg" for an augmented variant
annotation = sample["json"]

Layout

data/<lang>_001.tar .. <lang>_003.tar              originals, PNG + JSON
data_aug/<lang>_001_aug1.tar .. _003_aug3.tar      variants,  JPEG + JSON

Sample keys carry the language prefix (fr_0000000), so keys are unique across the whole dataset and shards can be unpacked into one directory without overwriting each other. An original and its three variants share a stem: fr_0000000 → fr_0000000_01, _02, _03.

Annotation

{
  "markdown": "...",            // the page as markdown, in reading order
  "blocks": [                   // every laid-out block
    {
      "top_left_x": 40, "top_left_y": 80,
      "bottom_right_x": 754, "bottom_right_y": 131,
      "content": "...",
      "type": "title",          // title | text | table | image | header | footer
      "reading_index": 1,
      "confidence_scores": null
    }
  ],
  "tables": [                   // one entry per table, as cleaned HTML
    {
      "id": "tbl-0.html",
      "content": "<table><tbody><tr><td rowspan=\"2\">..</td>..</tr></tbody></table>",
      "format": "html",
      "word_confidence_scores": [ ... ]
    }
  ],
  "images":  [ { "id": "img-0", "top_left_x": .., "reading_index": .. } ],
  "header": "...", "footer": "...",
  "dimensions": { "width": 824, "height": 1090, "dpi": 96 },
  "confidence_scores": { "word_confidence_scores": [ ... ] },
  "meta": {
    "template": "a4_paper",
    "language": "fr", "script": "latin", "direction": "ltr",
    "page_font": "NotoSerif-Regular.ttf",
    "title_font": "Montserrat[wght].ttf",
    "hybrid": null,
    "augmentation": null         // or { "name": "rotation", "params": {...} }
  }
}

Tables and merged cells

Roughly one table in five carries merged cells — rowspan, colspan, or both. Spans are placed on an occupancy grid, so every table expands to a complete rectangular grid with no cell claimed twice and no hole left unclaimed, and no span ever reaches past the last row. Merged cells carry text like any other cell.

This matters because merged cells are where table-recognition models fail, and because synthetic table data that only ever shows uniform grids teaches a model that merges do not exist.

What the pages look like

14 page templates: A4 reports, academic papers with LaTeX equations and theorem boxes, dedicated table pages, chart pages, image galleries, literary pages with verse, invoices, formal letters, receipts, book covers, business cards, ID cards and minimalist layouts. A page may also carry a small out-of-place "hybrid" block (7% of pages), which is deliberate: real documents are not homogeneous.

Text direction, alignment and boilerplate follow the language. Arabic pages are genuinely right-to-left, not Latin layout with Arabic glyphs dropped in.

Augmentations

Three variants per original, each a different transform drawn without replacement from: gaussian blur, bad photocopy, bleed-through, colour shift, letterpress, rotation (±5°, bounding boxes rewritten), reflected light, shadow cast, watermark, gaussian noise. meta.augmentation records which one and with what parameters. Watermark text is drawn in the page's own language, in a face verified to cover it.

Variants are JPEG (quality 90); originals are lossless PNG.

Fonts

Each language draws from faces verified against that language's script by reading the font's own cmap, not from a hand-maintained list. Complex scripts additionally require the right OpenType shaping tables — a font with Arabic codepoints but no arab GSUB renders unjoined letterforms that look like text and are not, and is excluded.

Corpus characters a page's font cannot render are dropped before the text is placed, so the annotation never claims a character the image does not show.

No font file is redistributed in this dataset — only rendered images.

Caveats

• Synthetic. The layouts are real layouts and the prose is real prose, but no page is a photograph of a real document. Use it to teach layout and script, and validate on real scans.
• Not deduplicated against anything else, including OCR-Data/ocr_data.
• Not reproducible. Generation is not seeded; re-running produces a different dataset.
• Hindi repeats source text more than the others. hi.wikipedia has about 163,000 articles against 29,997 samples, so the pool is shallower than English's.
• Confidence scores are synthetic — always 1.0. They exist so the schema matches real OCR output, not because anything was measured.

Licence

Images and annotations: CC-BY-4.0. Source text derives from Wikipedia (CC-BY-SA-3.0) and public-domain book scans; text reproduced in the images carries its original licence.