Skip to content
Bayanat Documents

Arabic documents with the labels your model needs.

Build a dataset for the documents your product reads, from bilingual invoices to handwritten forms. Get text, layout, extracted fields, and evidence for the answers you need.

What we can build together

Invoices and receipts

Arabic-English line items, totals, taxes, discounts, and stamps, with the scan quality and layout variation your system encounters.

Forms and records

Printed fields, handwriting, checkboxes, corrections, and missing entries in applications, delivery documents, or service reports.

Contracts and procurement

Clauses, tables, attachments, and cross-references, with questions answered from specific supporting passages.

Labels beyond transcription

Choose reading order, layout, normalized fields, evidence locations, and checks for missing information or inconsistent totals.

How we put the data together

Start with one document type

We define source permissions, layout diversity, capture conditions, and page count around the problem you need to solve.

Keep sources identifiable

Genuine, de-identified, and synthetic documents remain distinguishable. Template diversity is reported separately from page count.

Test unfamiliar documents

Evaluation splits can separate issuers, templates, sources, and writers to check performance on documents the model has not seen.

Checking an invoice total

A bilingual invoice has a discount and a handwritten correction. Labels locate both amounts so the system can extract them, recompute the total, and flag a mismatch.

Common questions

Can I order OCR labels without extraction labels?

Yes. You can request transcription, coordinates, and reading order on their own, or add layout, field extraction, document questions, and validation checks. The labels should match the model you are building.

How do you handle repeated invoice or form layouts?

For a custom collection, we agree the template coverage and evaluation splits with you. Variations of one template remain identifiable, so page count does not get mistaken for the number of independent layouts.