bks-lab/xrechnung-invoice-fields

XRechnung invoice fields from page images

34 German test invoices of the official XRechnung test suite, rendered to 118 page images, with twelve fields per invoice that a model has to read from the images alone: invoice number, issue and due date, seller and buyer name, seller VAT id, IBAN, net, VAT and gross total, amount due and currency. The correct value of every field comes from the invoice's XML, so the gold answers are exact and nobody had to annotate them.

This is test C3 of lab-bench, the bench BKS-Lab uses to decide which open model can do a task on hardware we control. The recorded outputs of every model we ran are included, so you can compare your own model with ours invoice by invoice.

• Built from commit 3584dfc8fddeb353916650f6da1f449cf20e2ccb of bks-lab/lab-bench.
• Method: results/c3/method.md, findings: results/c3/findings.md.
• Browse the outputs: bks-lab/xrechnung-invoice-fields-viewer.

Configs

pages (default). One row per page image: 200 dpi, greyscale, no text layer. Each row carries the invoice id, the page number and the gold fields of its invoice.

invoices. One row per invoice: gold fields, number of pages, and the paths of the CII XML and the rendered PDF in this repository.

predictions. One row per invoice and model arm, from our runs on one RTX 4090 with 24 GB through Ollama: the predicted fields, a correct flag per field (exact match after normalisation, see the method), the model's Hub id where there is one, quantisation, licence and seconds per invoice. kind says how the arm reads the invoice: vision from the page images, ocr+text through docling OCR and then a text model. The arm with reference_arm set reads the PDF text layer instead of an image. It shows what a text model can do when the text is already clean and is not a scan result.

Results so far

A regular expression over the PDF text layer gets 27.0% of the 408 fields right. The best vision model, gemma4-26b, gets 98.0%. The scores per model with 95 % bootstrap intervals are in the findings linked above. One run per arm, temperature 0.

Limits

• The invoices are test cases: names and numbers are the test suite's placeholders, layouts are the one official visualisation. Real scans with stamps, skew, handwriting or other layouts are harder.
• 34 invoices give wide intervals. Two models whose intervals overlap are not shown to differ.
• 17 of the 408 gold values are null because the invoice has no such field (no due date, no IBAN, no seller VAT id). A model gets those right by answering null. The render check in the repository found all 391 other gold values in the text of the rendered PDF.

Licence

The test invoices are from the KoSIT XRechnung test suite (commit 1ce1daf), rendered with the KoSIT XRechnung visualisation (v2026-08-31). Both are Apache-2.0, and so are the gold rows and outputs derived from them. The PDFs embed subsets of Source Serif Pro under the SIL Open Font License 1.1. The licence files are included.