Synthetic document pages for training OCR / document-parsing models, mainly Vietnamese with an English counterpart. Each page comes with exact ground truth (block polygons and content) and a ready-to-train target in the LightOnOCR-3 grounding format. The pages are rendered, not real scans; the augment parts are degraded variants (scan / phone / old-paper look).

Every type has three parts: vi and en (clean pages) and augment (degraded pages, Vietnamese and English mixed). slides has a fourth part, bilingual: slides that mix Vietnamese and English text.

Type	Pages `vi` / `en` / `augment`	Total	What
`article`	11,153 / 4,152 / 6,054	21,359	Magazine / newspaper pages
`paper`	7,979 / 3,890 / 4,241	16,110	Scientific-paper pages
`legal`	8,214 / 3,694 / 4,297	16,205	Legal-document pages
`wikimedia`	4,616 / 1,646 / 9,636	15,898	Encyclopedia-style pages
`webcode`	3,241 / 1,637 / 10,500	15,378	Rendered web pages
`standardbook`	4,357 / 3,058 / 13,673	21,088	Book pages
`contract`	14,401 / 9,922 / 11,368	35,691	Company contracts
`finreport`	9,903 / 8,756 / 10,000	28,659	Financial statements and annual-report pages
`invoice`	19,590 / 8,220 / 6,983	34,793	Invoices, receipts, stock slips, utility bills, plus handwritten vouchers and filled forms
`slides`	22,458 / 13,672 / 12,673, plus 15,011 `bilingual`	63,814	Presentation slides
`textbook`	18,077 / 6,818 / 11,740	36,635	School textbook pages
`exam_paper`	13,710 / 6,830 / 5,622	26,162	School exam papers, some filled in by hand; an option the student marked is written `A*.` in the target
`handwriting`	21,597 / 9,652 / 15,590	46,839	Handwritten documents
`form`	10,425 / 7,272 / 12,737	30,434	Forms filled in by hand; check boxes are written `☐` / `☑`
`handnotes`	4,497 / 3,561 / 9,815	17,873	Handwritten notes

426,938 pages in total.

Preview

The Dataset Viewer shows preview/<type>.parquet: about 100 example pages per type with their target (JPEG, long side up to 1400 px). The pages are examples, not a random sample; the full set is in the parts below.

Layout

<type>/<part>/
  images-000.tar, images-001.tar, ...   page images (members "images/..."), PNG / JPG
  gt.tar                                 one .md file per page (the target, for reading)
  shards.tar                             per-shard records (same schema as GroundTruth.jsonl)
  GroundTruth.jsonl                      full record per page (polygons, blocks, target)
  lighton.jsonl                          {"image_path", "target_lighton"}

Extract inside a part directory with for t in *.tar; do tar xf $t; done; image_path is then valid relative to that directory.

Ground truth

target_lighton is the page as LightOnOCR-3 grounding text, one block per paragraph in reading order:

![header](85,50,166,50,166,62,85,62) BÁO BÌNH MINH

![title](198,105,792,105,792,250,198,250) # Việt Nam là đối tác thương mại quan trọng nhất của Phần Lan ở Đông Nam Á

![text](100,252,892,252,892,289,100,289) Các cấp lãnh đạo Phần Lan ...

![label](x1,y1,x2,y2,x3,y3,x4,y4) gives the four corners of the block: top-left, top-right, bottom-right, bottom-left of the block as printed, x normalised to the page width and y to the page height, 0..1000. On a tilted or rotated page the box turns with the text. A + after the label (text+) marks a block continued from a column or page break. Content is Markdown, tables are HTML, formulas LaTeX; a chart is given as its data table, an image as a short description, a diagram as Mermaid.

Labels: text, title, list, table, chart, image, formula, caption, header, footer, page_number, footnote, diagram.

Figure rules:

chart: the plotted data as an HTML table, with the unit in the value headers.
table: an HTML table.
diagram: a ```mermaid block that mermaid 11.4.1 parses.
image: a short English description. Visible text is quoted verbatim. No description is empty or a placeholder.
formula: $$ … $$ LaTeX; inline math and sub/superscripts are $…$ . A literal dollar sign is written &dollar;.
Decoration (icons, emoji, stickers) is not in the target.

A few images in the wikimedia image tars have no target and are not listed in lighton.jsonl.

Licence

Gated, for internal use. The licence terms will be added later.

Limitations

Pages are synthetic; do not use this set alone to measure accuracy on real documents. Strong augment settings can make small text partly unreadable while the target still holds it. Some chart values and picture descriptions are estimates and can be wrong.

Downloads last month: 12,590

Total file size:

241 GB