Vietnamese + English training data for a small document VLM. The data covers seven tasks: page parsing (Markdown plus layout boxes), tables, formulas, charts, key information extraction (KIE) and VQA.


Parsing pages	1,495,135: VI 816,724 · EN 678,411
Task rows	8,368,061: train 8,259,714 · dev 50,235 · bench 58,112
Augmented views	568,134: scan-style 510,145 · photo 57,989
Images	~6.9M: page images 2.07M (pages + views) · task images 4,879,988 (VI 543,982 · EN 4,336,936)
Size	~1.1 TB (tar shards + JSONL)

Download by part

You do not need the whole 1.1 TB. Use download.py from this repository:

hf auth login                                         # gated: request access first
python download.py --list                             # parts and sizes
python download.py meta                               # index, mixes, candidates, card (small)
python download.py rows --source cord,mcocr           # task rows of some sources
python download.py row-images --source cord --extract # only the image tars those rows need
python download.py parsing --lang vi                  # parsing shards of VI pools (or --pool cc_pdf_vi,idl)
python download.py augment --pool pdfa_lighton2       # augmented views; augment-photo for G3
python download.py bench --extract                    # benchmarks; raw-sources for raw VI task data

Part	Size	What
meta	~2 GB	index, mixes, `candidates.jsonl`, curation and dedup lists, card
rows	8.9 GB	task rows (JSONL)
row-images	196 GB	images of the task rows (297 tars)
parsing	629 GB	parsing page shards
augment	155 GB	scan-style views
augment-photo	17 GB	photo views (G3)
bench	24 GB	benchmarks
raw-sources	12 GB	raw VI task sources
wip	4.2 GB	planned jobs, not training rows yet: VQA pilot + consensus votes (`wip/README.md`)

Curation workflow

What is inside

Task	Output	Main sources
Page parsing	Markdown, HTML tables, LaTeX; optional layout boxes `![class](x1,y1,x2,y2)`	VI web PDFs and scans, VI public sets, LightOnOCR-mix (pdfa), DocLayNet, IDL, olmOCR-mix, own renders
Table	HTML	PubTabNet, VI table crops, VI table VQA
Formula	LaTeX	SynthFormulaNet, UniMER, VI formula crops
Chart	HTML table	SynthChartNet, CoSyn, VI chart sets
KIE	JSON filled from a typed template	own VI renders (receipts, forms, purchase notes), CORD, MC-OCR (dev / bench)
VQA	short answer	VI and EN VQA sets (tables, charts, infographics, text)

Real vs synthetic

Synthetic: own renders (receipts, forms, charts, tables, math, short and dense pages) and synthetic public sets. Their labels are exact.
Real: web PDFs, scans and public datasets. Labels come from a human ground truth, from the PDF text layer, or from teacher reads (LightOnOCR-3 4B / 0.8B).

Parsing pages

Lengths and training mix

Mix (outputs/p1vi/mix_v2.json):

a target share for each task × language;
inside a group, sources are weighted by √rows × quality;
each source repeats at most 5 times;
parsing pages take 32 % of draws, and layout boxes are asked for on 45 % of them.

Augment

Recipe	What it does	Views
P1	print / photocopy: low ink, show-through, binarize, dust	510,145 in total (`corpus/v3_shards/aug/`)
P2	capture: blur, rescale, JPEG
G1	skew + zoom (+ light P2)
G2	page curl, warp, lens, shading
R90	90° / 270° rotation (text-only targets)
G3	photo of a page: oblique, screen recapture, wrinkle, hard shadow	57,989 (`corpus/v3_shards_g3/img/g3_*`)

Labels follow the parent page. Layout boxes are moved with the geometry, and views that push boxes off the page are rejected.
All probabilities are in corpus/v3_shards/aug.yaml. aug_parents.txt lists the parent of every view.
G3 views have their own index.json and enter the mix as g3_* pools (mix_a3_g3.yaml).

Samples

Image	Task	Prompt → target
	parse + boxes · VI	page → Markdown and `![class](box)` lines
	table · EN	crop → `<table>…</table>`
	formula · EN	crop → `$$\frac{t'(r)}{\rho'(r)}=…$$`
	chart · VI	chart → `<table><tr><th>Nhà cung cấp</th>…`
	KIE · VI	template → `{"ten_cua_hang": "NHÀ HÀNG HƯNG THỊNH", "date": "2024-11-03", …}`
	KIE · VI	template → `{"carrier": "VNPost", "tracking_no": "EX4074493798", …}`
	KIE · EN	template → `{"items": [{"name": "Grilled Salmon Head", …}]}`
	VQA · VI	"Năm nào Nicky Hayden đạt vị trí thứ 3…?" → "Năm 2005"
	VQA · EN	"What is the time signature of this piece?" → "4/4"

Full prompts and targets for every sample: samples.json.

Layout

corpus/p1vi/rows/<source>/*.jsonl    task rows: id, task, lang, split, images, prompt, target, licence, label_origin
corpus/p1vi/img_tars/img-*.tar       row images (path = rows[].images[i]); .tsv lists members
corpus/v3_shards/img/*.tar           parsing pages: <key>.jpg + <key>.json (text, grounding, source, licence, teacher)
corpus/v3_shards/index.json          shard list
outputs/p1/pool_all/candidates.jsonl ready parsing pages with byte offsets into the shards
corpus/v3_shards/aug/, corpus/v3_shards_g3/   augmented views (see Augment)
bench_tars/, raw_task_sources_tars/  benchmarks and raw VI task sources (tar + .tsv)
outputs/p1vi/mix_v2.json             training mix
wip/                                 planned jobs: VQA pilot, consensus votes (not in the mix)

import glob, json
row = json.loads(open(sorted(glob.glob("corpus/p1vi/rows/cord/train-*.jsonl"))[0]).readline())
# image: extract rows' images from corpus/p1vi/img_tars (members are named img/<sha256[:2]>/<sha256>.<ext>)
page = json.loads(open("outputs/p1/pool_all/candidates.jsonl").readline())
with open("corpus/v3_shards/" + page["shard"], "rb") as f:
    f.seek(page["json"][0]); meta = json.loads(f.read(page["json"][1]))      # label
    f.seek(page["img"][0]);  img = f.read(page["img"][1])                    # image bytes

Held-out data and decontamination

The dev and bench splits are never trained on. Benchmarks: OmniDocBench (EN), olmOCR-bench, SEA-Vision VI, MDPBench, VI dev / test, and MC-OCR VI KIE.
Train rows are removed when they match a benchmark: exact sha256, dHash near-duplicate checked against pixels, or a text signature (MC-OCR receipts).
Near-duplicate clusters are capped. Rule filters drop empty, too-long or garbled labels and bad fallback reads.

Licences

Every row and page carries a license field from its source. Some sources are non-commercial or research-only. Use the data for research and check each source's terms.

Planned (not in this release)

Only local models whose licences allow training on their outputs. Every new row will keep the model name, version and prompt id.

VQA on our own pages (VI + EN)

A local Qwen3.8-27B (FP8, non-thinking) reads the page transcript and writes 8 QA pairs per page: extract, lookup, reason and explain.
An extract answer must appear verbatim on the page; pairs that fail this check are dropped.
Pilot (in this release, not yet rows): 1,210 pages, 8,707 QA pairs (VI 999 pages, EN 211) in wip/vqa_pilot/. Full run: ~24k pages.

Consensus voting: several readers label the same page, and only what they agree on is kept. The first extra voter is done and is in wip/consensus/: OvisOCR3 reads of 40.6k pages (plain + boxes, with log-probs) and its boxes on 68k EN scans, matched against the current boxes.

Target	Voters	Kept
Parsing of weak VI pages	OvisOCR3, LightOnOCR-3, PDF text layer	the medoid read (closest to the others)
Layout boxes	OvisOCR3 structured output vs the current boxes	boxes both agree on (IoU match)
KIE on real forms, invoices, receipts	local Qwen with the typed template, 2 samples	fields where the samples agree
Chart → table on real charts	local Qwen, 2 samples	cells where the samples agree

Also planned

VI formulas rendered from textbook LaTeX.
Embedding dedup (SigLIP2 / DINOv2) to break template chains that dHash misses.
Human-checked VI test sets.

Downloads last month: 360

Total file size:

1.11 TB