drew-ipp/invoice-extraction-benchmark

Invoice Extraction Benchmark v1

A synthetic test set for invoice data extraction (invoice OCR, intelligent document processing, accounts-payable capture): 181 documents with answer keys and a scorer. Run any invoice reader over the documents, write its output as one JSON file, and score it field by field.

Source, scorer and generator: https://github.com/DrewKraken/invoice-extraction-benchmark (this dataset is a mirror of corpus/v1/ there; the GitHub repository is canonical).

Who made it. The set was made by InvoiceParser Pro (IPP), a company that sells an invoice extraction product. It is not an independent benchmark; weigh IPP's own results below with that in mind.

Contents

• corpus/v1/documents/: 181 documents. 113 PDF (digital, with a text layer), 53 JPEG (phone photos and handwritten), 15 PNG (scans). 220 pages in all.
• corpus/v1/answers/<doc_id>.json: the answer key for each document.
• corpus/v1/manifest.json: every document with category, difficulty, damage, sha256 and answer key.
• corpus/v1/metadata.jsonl: the same, one row per document (file_name is relative to corpus/v1/). This is what the dataset viewer shows.

Category	Docs	What it tests
clean	31	Digital PDFs: varied headers and tables, US/UK/EU/ZA date and number formats
currency	19	USD, ZAR, NAD, EUR, GBP, ZMW; bare $ on a Namibian invoice; dual-currency payable boxes; lines in several currencies
tax	14	VAT columns, VAT-inclusive totals, "Sub Total" after VAT, zero-rated, percentage lines, withholding
credit	10	Credit notes printed with minus, R -, brackets, CR, or no sign at all
multipage	10	2 to 8 pages, carried/brought-forward rows
freight	16	Carrier and clearing & forwarding invoices, trip/load references
scan	15	5 scan looks x 3 grades (grayscale, noise, fax, punch holes, stamps)
photo	45	15 phone-photo problems x 3 grades (blur, perspective, rotation, low light, shadow, JPEG, crumples, stains, pen marks, crops, screen photo...)
handwritten	8	Invoice books filled in by hand, three legibility grades
hard	13	Several invoices in one file, statements, order number beside invoice number, no invoice number, totals that do not add up

Every vendor, customer, number and amount is generated; no customer document was read, copied or imitated. Company names are built from common words (place + trade + legal form) and may coincide with a real business by accident. 10 documents of the original 191 (a document type that is not public yet) are withheld; IPP's published figures leave them out too.

Scoring

git clone https://github.com/DrewKraken/invoice-extraction-benchmark
cd invoice-extraction-benchmark
pip install -r requirements.txt
python -m scorer.evaluate predictions.json --markdown scores.md

Amounts exact to the cent, dates as ISO dates, vendor names after normalising legal forms, invoice numbers after normalising case and labels, line items by greedy match (F1 per document), and a flag that should fire on inconsistent invoices and multi-invoice files. A document is fully correct when every key header field it is scored on is right. Full rules and the predictions format are in the GitHub README.

Results

System	Date	Docs	Every key header field right	Key header fields right	Line items, mean F1	Inconsistent invoices flagged
InvoiceParser Pro (production)	2026-10-03	181	96.1% (174 of 181)	98.7% (1,555 of 1,576)	0.983	5 of 5

IPP's own measurement through its live production service; details and every miss at https://invoiceparserpro.com/accuracy. Results from other tools are welcome as pull requests on GitHub.

Limitations

Synthetic documents from one layout engine, so far less layout variety than real supplier invoices; mostly English with Southern African, UK, US and EU conventions; simulated (not photographed) damage; small per-category counts.

License and citation

Data CC BY 4.0; code (on GitHub) MIT.

@misc{ipp_invoice_extraction_benchmark_2026,
  title        = {Invoice Extraction Benchmark v1: 181 synthetic invoices, photos, scans and handwritten documents with answer keys},
  author       = {{InvoiceParser Pro}},
  year         = {2026},
  month        = oct,
  howpublished = {\url{https://github.com/DrewKraken/invoice-extraction-benchmark}},
  note         = {Synthetic corpus, seed 42; results at https://invoiceparserpro.com/accuracy}
}