A benchmark for Vietnamese OCR and document parsing: 12,700 page images in four subsets. Each subset has its own ground truth format and its own metrics. Report one number per subset; do not average across subsets.

Subset	Pages	What	Ground truth	Licence
`meddies`	973	Real Vietnamese administrative, financial and education pages: 390 text-only, 288 with tables, 295 with merged-cell tables	`meddies/data.jsonl` → `gt_html`: `<div data-bbox data-label>` blocks, tables as HTML with rowspan/colspan	source licence "other", see LICENSES.md
`handwriting`	128	Real handwritten paragraphs by 20 HANDS-VNOnDB writers	`handwriting/data.jsonl` → `text`	research only
`mdpvi`	160	The Vietnamese pages of MDPBench: 40 digital and 120 camera photos (textbooks, papers, slides, exams, ...)	`mdpvi/gt.json`, original MDPBench / OmniDocBench format	Apache-2.0
`vidocparse`	11,439	All of ViDocParse (every split of the source repo): academic_paper 3204, financial_report 1962, newspaper 1233, textbook 1174, handwritten_note 1088, ppt_slide 861, exam_paper 768, magazine 605, book 544	`vidocparse/src/<category>/<id>.json`, OmniDocBench format, unchanged	Apache-2.0

Access is gated: the meddies and handwriting parts come from sources whose terms do not clearly allow open redistribution. Use the data for research evaluation only.

Files

manifest.jsonl                one row per page: sample_id, subset, image, gt (pointer), task, attrs
exclude_from_training.json    every page id of this benchmark: keep them out of any training or tuning data
SHA256SUMS                    checksums of all data files
LICENSES.md                   licence and citation of each source
meddies/      data.jsonl  images/
handwriting/  data.jsonl  images/
mdpvi/        gt.json     images/
vidocparse/   src/<category>/<id>.png + <id>.json   metadata/*.txt   ViDocParse_test.json

All paths in manifest.jsonl are relative to the repository root. Download the whole repo with huggingface-cli download GreenNode/VieDocOCR-Benchmark --repo-type dataset --local-dir VieDocOCR-Benchmark (about 64 GB, of which vidocparse is 63 GB). Use --include "meddies/*" "handwriting/*" "mdpvi/*" "*.jsonl" "*.json" for the small subsets only (about 1.1 GB).

Ground-truth fields

meddies/data.jsonl: sample_id, image, stratum (text-only / table / table-span), gt_html (the label to score against), original_html (the label in the source dataset before correction, kept for comparison only), review_verdict (ok = source label was already right, edited = corrected), reviewer, review_note, unsure, tags, image_sha256. Block labels: Text, Section-Header, Table, Page-Header, Page-Footer, List-Group, Image, Caption, Footnote, Equation-Block, Figure, Complex-Block (bbox values are the source's boxes and were not re-checked).
handwriting/data.jsonl: sample_id, image, text (paragraph transcription), writer (VNOnDB writer id).
mdpvi/gt.json: the MDPBench annotation list for these 160 pages (layout_dets, page_info, extra).
vidocparse/src//.json: the source file as published: a one-element OmniDocBench list, except 140 files that hold the page object itself ({layout_dets, page_info, extra}) without the list. vidocparse/metadata/*.txt are the source's page-id lists (test, train_manual, train_generated, train_mixed, manual, generated), re-saved with LF line endings; manifest.jsonl copies them into attrs.source_split (test / train) and attrs.annotation (manual / generated). We have not verified how the generated annotations were produced; report manual and generated pages separately if that distinction matters to you. vidocparse/ViDocParse_test.json is the 1108 test pages in one OmniDocBench JSON with two fixes that the OmniDocBench scorer needs: 547 header/footer/abandon items without text got text: "" and are ignored, and 15 truncated relations pointing at those items were dropped.

How we score

Use the same prompt and settings for every subset, and give a bootstrap confidence interval over pages (we use 1000 resamples).

Subset	Metrics
`meddies`	Page CER (normalised edit distance) with and without diacritics, diacritic-word error rate (share of GT words whose letters match but whose tone or vowel marks are wrong), TEDS and TEDS-structure-only on tables, loop rate and foreign-script (CJK) insertion rate
`handwriting`	CER, CER without diacritics, rate of pinyin or Chinese-character output, loop rate
`mdpvi`	the official MDPBench / OmniDocBench end-to-end metrics (text edit distance, table TEDS, reading order). They are comparable with published MDPBench numbers
`vidocparse`	OmniDocBench end-to-end metrics per category, plus page CER and the diacritic-word error rate on body text

Header / footer convention. Meddies GT includes page headers, footers and page numbers. ViDocParse page text excludes them. Models differ on whether they transcribe this page furniture, so for page CER we also report a body-only variant that removes GT furniture blocks and drops the prediction lines that match them. Say which variant you report.

Caveats (read before quoting numbers)

Meddies labels were corrected by AI and have not been checked by a human. The source labels (Meddies/meddies-ocr) are noisy: roughly 1.8 errors per 100 words, and the table structure was wrong on 10 of 30 audited table pages. Every page was reviewed against the image twice by Claude models (Opus, then Sonnet subagents). A person settled about 11 pages, and the two reviewers gave the same verdict on 9 of 10 spot-checked pages. The residual error rate is unknown.
- Convention: labels copy the print, keeping typos, missing diacritics, Đ written as D, and number separators. Tables follow the image.
- Out of scope: stamps, signatures, text inside charts, and Arabic script.
- Excluded: pages that are not Vietnamese and pages whose font garbles diacritics (27 of 1000).
- Stamp text may remain on some pages reviewed early.
Meddies text is public. These are real government and institutional documents, and about 150 of the 973 pages share most of their text (≥50% of word 13-grams) with publicly available Vietnamese legal-document corpora. A model trained on such corpora may have seen the text. A model trained on Meddies labels also learned the label conventions and gets an advantage on this subset, so compare it on mdpvi, handwriting and vidocparse as well.
The handwriting images are pen strokes redrawn from online ink, not scans or photos: 523 px wide on a white background. Upscale them if your model needs larger input. The writers are held out from our own training data, but other VNOnDB writers copied the same paragraph texts, so a model trained on VNOnDB text has seen most of these sentences. VNOnDB labels contain a few errors.
MDP-VI is small (40 digital, 120 photo pages), so treat small differences as noise. The MDPBench scorer at the time of writing splits digital pages into pseudo-pages when it averages per page; regroup by page before you compute page averages.
ViDocParse is copied unchanged from an anonymous EMNLP 2026 submission (emnlp2026anon/ViDocParse, Apache-2.0). In this benchmark every split, including the source's train split, is evaluation data. Some scans are very large (up to about 140 MP), so resize them before inference.
Some pages may contain names of real people or organisations, because they are real public documents. If you find personal data that should be removed, open a discussion on this repo.

Citation

If you use a subset, cite its source (see LICENSES.md): Meddies OCR, HANDS-VNOnDB, MDPBench, ViDocParse.

Downloads last month: 27

Total file size:

68.4 GB