extractingbench

Accuracy

Is the value right?

Every field printed in the document is compared with a checked answer after fixed normalization rules: money to the cent, dates as full dates, identifiers kept exact.

Broken down by field type (text, number, date, list, computed) and by document type (receipts, invoices, government forms, emails, tables, documents with many missing fields).

Accuracy share of fields in the document read right

Clairic
Cloudflare
NuExtract3
NuMind
GLiNER2
Fastino

Share of fields printed in the document that the model returned correctly, after the normalization rules on the methodology page. Every document type counts equally.

Test items and precision

Items
600 documents per test set, about 6,000 fields, across six document types
Precision
About ±2 points between two models overall; about ±5 within one document type.

Full method and scoring rules