wahaj-labs/arabic-invoice-extractor

Arabic Invoice Extractor

PublicLimited riskRev-ShareOCRPyTorchBilingual2.1.0

Test in Sandbox

5.6K 172K/30d 290

Field F1 (header)

0.93

Field F1 (line items)

0.85

Character error rate

2.1%

Straight-through rate

0.78

About this model

Document-understanding pipeline that extracts structured fields, supplier, CR number, line items, VAT-ready totals, dates, and payment terms, from Arabic, English, and mixed-language invoices. Combines a layout-aware OCR stage with a field-tagging transformer robust to scans, photos, and PDFs.

Intended use

Accounts-payable automation and procurement digitisation for government entities and enterprises processing bilingual supplier invoices.

Training data lineage

Layout backbone pre-trained on document pages from the arabic-gov-docs-corpus, then fine-tuned on 96,000 annotated invoices contributed by launch partners.

Limitations & bias notes

Handwritten amendments and stamps overlapping key fields remain the dominant failure mode. Line-item extraction accuracy drops on multi-page invoices with carried-over subtotals and on low-resolution WhatsApp photos.

Evaluation metrics

Field F1 (header)	0.93
Field F1 (line items)	0.85
Character error rate	2.1%
Straight-through rate	0.78

#ocr#invoices#document-ai#arabic#layout#accounts-payable

Try it

Live sandbox

invoice_gulf_supplies_04417.pdf

fixture

Feedback

Owning entity

waWahaj Labs

Updated: 2026-07-03
Latest version: 2.1.0
License: Revenue Share
Access: Open

Trained on

arabic-gov-docs-corpus →