andree2103/VieOCR-Real-Dataset

VieOCR Real Dataset

A mirror of the Vietnamese part and the English full-page part of MonkeyDocv2 (ModelScope zenosai/MonkeyDocv2, Apache License 2.0), released with the MonkeyOCRv2 paper (arXiv 2607.11562). All credit for the data goes to the MonkeyDocv2 authors. Files are byte-identical to the original release; the only change is that files larger than 5 GiB are split into parts so they fit the Hub's per-file limit.

Contents

Path	What	Size
real/FinePDF/page/image/vie_Latn.tar.zst	103,065 Vietnamese full-page images (FinePDFs)	60 GiB
real/FinePDF/page/json/{full_rec,layout,layout_rec}/vie_Latn.json	page-level labels: reading-order content, layout boxes, layout + content	0.9 GiB
real/FinePDF/block/{text,html,formula}/image/vie_Latn.tar.zst	cropped text / table / formula blocks	64 GiB
real/FinePDF/block/{text,html,html/json_otsl,formula}/json/vie_Latn.json	block labels (tables as HTML and OTSL)	1 GiB
real/FinePDF/page/image/eng_Latn.tar.zst	162,889 English full-page images	89 GiB
real/FinePDF/page/json/*/eng_Latn.json	English page-level labels	1.2 GiB
synth40m/table/vi.tar.zst	synthetic Vietnamese table images + labels	89 GiB
synth40m/text/vi.tar.zst	synthetic Vietnamese text images + labels	344 GiB

Each label file is a JSON list of chat samples: {"messages": [user prompt with <image>, assistant answer], "images": ["MonkeyDocv2/real/FinePDF/page/image/vie_Latn/<n>.png"]}. In layout_rec, the answer is a list of {bbox, label, content}; bbox values are normalized to 0–1000 per axis; tables are HTML, formulas LaTeX. About 27% of Vietnamese pages (28,234) contain at least one table.

Split files

A file over 5 GiB is stored as <file>.part000, <file>.part001, ... with a <file>.SHA256SUMS:

cat vi.tar.zst.part* > vi.tar.zst
sha256sum -c vi.tar.zst.SHA256SUMS --ignore-missing

Label quality

The real-page labels are pseudo-labels from an expert-model pipeline (agreement of dots.mocr, PaddleOCR-VL and Qwen3-VL; low-agreement pages dropped), not human annotation. Shared mistakes survive the vote, including Vietnamese diacritic errors (e.g. "Hội năng Mây" for "Hội nàng Mây"). Use them as training data, not as an evaluation ground truth. The synth40m labels are exact by construction (rendered text).

License

Apache License 2.0, as the original MonkeyDocv2 release. The page images come from FinePDFs (ODC-By); check the source documents' terms for your use.