Soxavin/ardb-sft-v5

ARDB Unified Page Understanding (bbox + text, one forward pass)

Vision fine-tuning data for document-understanding VLMs, built from ARDB (Agricultural and Rural Development Bank of Cambodia) daily market-price bulletins. This dataset unifies layout detection (bbox only) and transcription (text only, on pre-cropped regions): one full page image in, one JSON list out, where each region carries its box AND its transcribed text together, in a single forward pass.

Model-agnostic: the box/label/text content isn't tied to any one model family. The only model-specific choice is the box_2d coordinate convention below (a common one, not unique to a single model) — fine-tuning teaches a base model this format directly regardless of that model's own native convention, so no reformatting is needed to reuse this dataset with a different model. Used so far to fine-tune Gemma 4 E2B; also loaded as-is (no schema changes) in a Qwen3.5-0.8B fine-tuning trial.

The source corpus spans a multi-year range (2022–2026), covering both structural bulletin templates in circulation over that period. Supersedes ardb-sft-v4 (same box annotations and template-substitution text; that card documents the split-assignment issue this version fixes).

Stats

Split	Rows (pages)	Docs
train	101	38
validation	9	3
test	18	6

47 documents total (5 excluded: 2 frozen wholesale_retail-era evaluation bulletins, 1 retail_only-era template anchor, 1 non-ARDB test file, 1 additional reserved anchor). Of the 55 documents in the full corpus, 3 further documents produced 0 usable pages each (2 share a minority pagination pattern in the retail-only era; 1 sits exactly at the boundary between the two structural eras) and are counted among the corpus's quarantined pages, not as skipped docs.

Note on split size: splits are clustered by issue date (no adjacent-day leakage) and stratified by template era (both structural templates represented in every split). Validation and test are still small in absolute terms (9 and 18 rows) but span both templates; treat as a training-health signal more than a large-scale generalization claim.

Schema

Field	Type	Description
image	image	Full rendered page
instruction	string	Fixed task instruction
text	string	JSON list of {"box_2d": [y1,x1,y2,x2], "label": category, "text": content}
doc_id	string	Document identifier
source	string	Source PDF filename
page	int	Page index (0-based)
date	string	Bulletin date (page 0 only, else empty)

box_2d is normalized to a 0-1000 grid ([y1, x1, y2, x2], matching Gemma's native detection convention — chosen as a widely-supported convention, not because the dataset is Gemma-specific), and the list is sorted top-to-bottom by box_2d[0]. Categories: Table, Text, Section-Header, Page-Furniture, Picture. text is "" for regions with no content to transcribe (Picture).

Construction

Two independent, already-verified pipelines, zipped by (page, label):

• Boxes come straight from ardb-layout-coco-v5's human-corrected COCO annotations (no new annotation work).
• Text comes from template substitution. The corpus spans two structural templates — a retail-only layout (pre-May 2024) and a combined wholesale/retail layout (May 2024 onward) — each document is classified by era and matched against its own verified template, then that template's layout is filled in with the document's own varying content (table prices/dates, the title's day and month), rather than running OCR independently per document. Page-Furniture (letterhead) and Text (source-note footer) are fully static across every document and were verified directly against rendered crops.

A page is skipped entirely (never emitted with one field silently wrong) if its table can't be aligned to its era's template, or a Section-Header/Page-Furniture region can't be resolved to text.

Limitations

• Two structural templates, era-classified by content (presence of the wholesale-price marker in each document's own text). Splits are stratified by era, though not in exactly proportional counts — cluster integrity (no adjacent-day leakage) takes priority over exact proportionality.
• Still a small corpus (55 source documents) — not large-scale despite the split fix.
• A small number of documents (3 of 55, 9 pages total) are excluded rather than included with unverified content. Each has a table layout that varies just enough from its era's reference template that automatic verification can't confirm the substituted text lines up correctly, so the pipeline drops the page rather than risk silently misaligned data. Recoverable with additional manual template verification; not yet done.

License

cc-by-nc-4.0. Images are reproductions of publicly published ARDB bulletins (research use only; not affiliated with or endorsed by ARDB).