Training data for document understanding models

Document understanding models are trained on real documents with their outcomes. Useful records include invoices, forms and contracts, redline and approval histories, and extracted fields with corrections. Corrections show ground truth, and the records vary across years and vendors, which reflects real layouts and messy scans.

Document understanding models records have personal details removed, and the supplier approves each release. A rights review runs on invoices, forms, contracts and approval histories before a sample is prepared. That sample is shared under NDA only after supplier approval.

Document understanding models are sourced through a request describing the records, volume and uses you need. SourceX matches suitable suppliers in the U.S., Canada, the EU, the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia and runs a rights review. A prepared sample is shared under NDA after supplier approval, and the license sets allowed uses, term and deletion.