mmbert-small-invoice-extraction

Marks the words of an invoice's fields in its text (as pdf.js reads a digital PDF): seller and buyer, their VAT numbers, invoice number, issue and due dates, net, VAT and total, and the IBAN, in the 24 languages of the EU, Norway and Ukraine. The first 12 of mmBERT-small's 22 layers, fine-tuned with BIO tags over words; ONNX weight-only int8 for transformers.js, 47.7 MB plus a 9.3 MB tokenizer.

It runs with a rule layer, as the widget on balticml.eu does: VAT numbers (25 countries) and IBANs checked by their check digits, sums that must add up, dates that must be dates, and seller and buyer shown only where the page names them (otherwise both companies are listed, roles not named). The model labels words, not tokens, each tokenized with the whitespace before it; use the code in the repo (github.com/balticml/invoice-extraction, web/model.js, web/pdftext.js, web/review.js) rather than the generic pipeline.

Results

PDFs from invoicing tools it never trained on, int8 with the rules, threshold 0.6. A field is right when its value equals the invoice's after normalising; "not sure" fields are left empty.

set	fields	right	wrong, shown	not sure
InvoiceShelf, 72 invoices, 24 languages	746	61.7%	0.8%	37.5%
Invoice Ninja, 63 invoices, 6 designs, 23 languages	678	64.7%	0.3%	35.0%
Fakturownia, 20 invoices	220	86.4%	0.0%	13.6%

Most "not sure" fields are the seller and buyer of layouts that do not label them; in their place the widget lists both companies, and 6–8% of the names listed are wrong. On 60 LLM-written invoices checked by hand: 84.5% right, 0.8% wrong. The full tables, and how the test sets were kept apart: the repo's README.

Training data

No real invoices. 9,000 invoices from a generator (six layout families, six languages), 23,259 in layouts composed at random from blocks in 24 languages, 566 written whole by gpt-oss-20b for 30 kinds of business (each also filled again five times with other names and numbers), and 480 exported from local Dolibarr and Odoo installations to made-up companies. Every invoice was read with pdf.js, block by block; each word's label comes from where its value is printed.

Limits

Digital PDFs with a text layer only (no scans). One-page invoices of small businesses tested; no real invoice scored yet. Company names are the weakest fields. Test it on your own suppliers' invoices before trusting it on theirs.