EVE-Train-NPL-1.1

Synthetic scanned-document pages with gold key-information-extraction labels for Vietnamese non-performing-loan collateral paperwork. Every value is fabricated — no real borrower, property, contract or person appears. Pages are drawn programmatically, so labels are exact by construction rather than annotated.

Training a QLoRA adapter on this corpus moved field exact-match accuracy on genuine VAMC scans from 64.0% to 81.4% (ysmeta/EVE-OCR-1.0-NPL).

What changed from 1.0

Version 1.0 improved the average but made one real page worse (69.0% -> 65.5%). That page prints addresses abbreviated (Q. Tân Phú, TP.HCM); 1.0 only ever spelled the administrative prefix out (huyện …, tỉnh …), so the adapter learned to normalise toward the single form it had seen.

Version 1.1 adds variation drawn from features observed in real documents:

Added	Seen in
Abbreviated address forms P. X. TT. Q. H. T. TP. (~39% of pages)	the regressed page
Municipality addresses (TP.HCM, TP. Hà Nội …) replacing province/district	real filings
Currency unit varies: VNĐ / đồng / VND / đ	differs per document
Numbered appendix titles (PHỤ LỤC 1:)	one real page
Three term-of-use phrasings: a date, tháng MM/YYYY, Lâu dài	across pages
Mortgage contracts with no number, only a date — gold value null (~25%)	one real page
Certificate numbers containing a space (CV 366117)	one real page
Two more credit-contract number families	real filings

Result on the previously regressed page: 65.5% -> 79.3%, above its untuned score. Every other page improved as well. 1,500 -> 1,800 pages.

Contents

• 1,800 JPEG pages — Vietnamese 1,080 / Korean 360 / English 360
• metadata.jsonl — 30 schema fields per page, plus lang and split_role
• vamc_gen.py — the generator, so the corpus is reproducible and extensible

split_role marks 60 pages (20 per language) held out of training and used only to check whether adapting to Vietnamese degraded the other languages. It did not: Korean rose 63.3% -> 92.2%, English 75.0% -> 91.1%.

Failure modes the generator deliberately induces

Observed in real scans	How it is reproduced
One contract's figures mixed with the totals row	1-3 contract rows plus an always-present totals row
Absent fields filled with junk instead of null	fields randomly omitted, gold value null
Handwritten entries misread (1406 read as 146)	handwriting font in dotted blanks, per-glyph rotation and jitter
Text lost under the company seal	semi-transparent red seal composited over body text
Vietnamese number formats altered	. thousands / , decimal preserved in gold
Low-quality capture	skew, blur, downsample to 150-200 DPI, JPEG artifacts
Vertically merged table cells	merged mortgage/status columns across asset rows

Load

from datasets import load_dataset
ds = load_dataset("ysmeta/EVE-Train-NPL-1.1")

Limitations

• One document family: VAMC valuation-request appendices.
• The generator draws clean geometry then degrades it. Real scanner artefacts, paper texture and print bleed are approximated, not simulated. A 21 pp gap remains between held-out synthetic (93.3%) and real pages (81.4%).
• Korean and English variants reuse the Vietnamese layout with translated labels. They are a language-robustness control, not authentic local paperwork.
• Evaluate on real documents. Training and evaluating on synthetic data alone measures nothing.