EuroDocBench

1,440 synthetic European invoices in 12 locales and 5 layouts — PDF, scan, text layer and exact ground truth — for benchmarking key-information extraction and locale-aware normalisation.

All documents are synthetic. No real invoice is included, and the layouts are generated by five templates. A high score here shows that a system handles European number, date, VAT and identifier formats on clean generated documents; it is not evidence of accuracy on real invoices, and the keyword/regex baseline's 100 % exists only because it was written for these templates.

Built by Since AI. Code, specification and scorer: github.com/sinceaihq/eurodocbench. Browse documents and baseline predictions in the results Space.

# pip install "datasets[vision]"   (Pillow decodes the scan images)
from datasets import load_dataset

ds = load_dataset("sinceaihq/eurodocbench", split="test")
row = ds[0]
row["image"], row["text"], row["fields"]

Task

Read an invoice (the text track uses the PDF text layer, the scan track the image) and return 12 fields in canonical form:

field	canonical form
invoice_number	as printed
issue_date, due_date	YYYY-MM-DD (due_date may be null)
currency	ISO 4217 (EUR, SEK, DKK, PLN, CZK)
seller_name, buyer_name	as printed
seller_vat_id, buyer_vat_id	country prefix + number, no spaces (buyer_vat_id may be null)
iban	no spaces (may be null)
net_amount, vat_amount, total_amount	plain decimal, e.g. 1234.50; vat_amount is the total over all VAT rates (invoices with two rates print only the per-rate lines)

Score with eurodocbench-score --gold gold/test.jsonl --pred predictions.jsonl (see the specification).

Baseline results (test split)

system	track	field accuracy	field F1	document accuracy	hallucinated values	format errors
Keyword/regex rules	text	100.0	100.0	100.0	0	0
Qwen2.5-1.5B-Instruct (zero-shot)	text	82.7	82.5	27.8	146	1,581
EasyOCR → keyword/regex rules	scan	66.5	75.5	2.9	24	0
EasyOCR → Qwen2.5-1.5B-Instruct	scan	67.3	66.7	4.3	194	1,770

The keyword/regex rules were written for these documents and show that every answer is recoverable; the other rows are general systems. Per-field, per-locale and per-layout results and an error analysis are in the technical report.

Files

path	content
data/{dev,test}.parquet	one row per document: id, locale, language, country, layout, font, date_style, currency_style, image (scan, JPEG), pdf (bytes), text (PDF text layer), fields (ground truth), vat_breakdown, lines (line items)
gold/{dev,test}.jsonl	ground truth in the scorer's format
manifest.json	SHA-256 of every generated file, generator version and seed
runs/	baseline predictions (*-test.jsonl), their scores (*.score.json) and run settings (*.meta.json) from the report, and the EasyOCR output for both splits (ocr-{dev,test}.jsonl)

Splits

split	documents	per locale × layout	seed
dev	240	4	2026
test	1,200	20	2026

Locales: fi-FI, sv-SE, de-DE, fr-FR, nl-NL, it-IT, es-ES, pl-PL, pt-PT, cs-CZ, da-DK, en-IE. Layouts: classic, modern, compact, grid, narrow.

How the data was made

Everything is generated by code (eurodocbench-generate --seed 2026): seller/buyer names from invented syllables plus a sector word and legal form, real city names, line items from a translated catalogue, 2026 VAT rates, amounts with commercial rounding, rendering with reportlab and the bundled Liberation fonts, and simulated scanning (skew, blur, uneven exposure, noise, JPEG). VAT numbers and IBANs have the correct national format but deliberately wrong check digits, so they cannot belong to any real company or bank account. No personal names appear.

Intended use and limitations

For comparing document-AI systems on European formatting conventions. The documents are single-page, template-based and cleaner than real invoices; good scores here do not demonstrate real-world performance. Because the generator is public, results of models trained on generated documents should be reported as such; use a private seed for contamination-free evaluation.

Licence

CC BY 4.0 (data); the generator is Apache-2.0. Cite as:

@misc{eurodocbench_2026,
  title  = {EuroDocBench: a synthetic benchmark for European invoice extraction},
  author = {{Since AI}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/sinceaihq/eurodocbench}
}