ComingOn documents
French invoice extraction
How many of your invoices will go through without a human fixing them?
The protocol for this benchmark is written. No model has been queried yet, so this hub shows no figure for this task.
What this test will measure
Extract the fields of a French supplier invoice received as a scan or a photo: amounts, VAT rates and total, company number, dates, line items. The cases that make French invoicing hard — VAT exemption, reverse charge, credit notes, deposits, several rates on one invoice — will have to be represented in the test set. A field absent from the document is part of the test: inventing a value counts as a hallucination.
What will be graded
- Net, VAT and gross amounts
- SIRET and VAT numbers
- Issue and due dates
- Line items
- Trap cases
What it is waiting for
No public dataset exists: these documents never leave the company. They will be collected from partners, with their consent, and only the scores will be published.
- Place on the roadmap
- Wave 4
- Target sample
- 25 invoices
- What the model receives
- Documents (images or PDF)
One business function at a time, one task at a time. Every wave ends with a publication.
No figure is shown here because there is none. The day this task is measured, its ranking will appear on this very page.
How the hub measures — how an answer is verified, the rubric and the verdicts, on the task already measured.