EVE-Train-NPL-1.0

Synthetic scanned-document images with gold key-information-extraction labels, built to train a vision-language model on Vietnamese non-performing-loan (NPL) collateral appraisal paperwork.

Every value in this dataset is fabricated. No real borrower, property, contract or person appears here. Images are drawn programmatically, so the gold labels are exact by construction rather than annotated by hand.

Why it exists

Real VAMC appraisal letters are scanner output with no text layer at all. Measuring a Tesseract baseline and a Qwen2.5-VL-7B zero-shot baseline on genuine documents showed the hard part is not character recognition — the base model already reproduces 92.6% of gold values somewhere in its output — but field discipline: returning exactly the value a schema field asks for, and returning null when a field is absent. Exact-match field accuracy was 51.9%.

This dataset reproduces the specific failure modes that were measured:

Observed failure	How the generator induces it
Row confusion — one contract's figures mixed with the totals row	1-3 credit-contract rows plus an always-present totals row
Absent fields filled with junk instead of null	Fields randomly omitted; gold value is null
Handwritten entries misread (1406 read as 146)	Handwriting font in dotted blanks, per-glyph rotation and baseline jitter
Text lost under the company seal	Semi-transparent red seal composited over body text at random positions
Vietnamese number formats altered	. thousands separator and , decimal separator preserved in gold
Low-quality capture	Skew, blur, downsampling to 150-200 DPI, JPEG artifacts

Contents

• 1,500 JPEG page images: Vietnamese 900 / Korean 300 / English 300
• metadata.jsonl — 30 schema fields per image, plus lang and split_role
• vamc_gen.py — the generator, so the corpus can be reproduced or extended

split_role marks 60 images (20 per language) held out of training and used only to check whether adapting to Vietnamese degraded the other two languages.

Schema

Document metadata, credit-contract totals, collateral registry details, address components, and optional amounts. Field names are identical across languages; only the values change language. Fields absent from a document are null.

Load

from datasets import load_dataset
ds = load_dataset("ysmeta/EVE-Train-NPL-1.0")

Limitations

• Layout is modelled on one document family (VAMC valuation-request appendices).
• The generator draws clean geometry then degrades it; genuine scanner artefacts, paper texture and print bleed are only approximated.
• Korean and English variants reuse the Vietnamese layout with translated labels. They are a language-robustness control, not authentic local paperwork.
• Evaluation of models trained on this data should use real documents. Training and evaluating on synthetic data alone measures nothing.