Invoice-Check JP: A 3B Model Reads the Invoice, Checks Decide Whether to Trust It
A QLoRA-tuned vision-language model extracts Japanese qualified invoices; a check digit, a registry lookup and tax arithmetic decide what gets automated. On synthetic invoices: 86.8% auto-approved, and none of the approved invoices had a wrong field those checks can catch. Scope note: everything here is measured on synthetic invoices with fictitious companies (600 test invoices, 200 hold-out…
Japanese qualified invoices now include a registration number, tax amount per rate, and tax per rate. A model was built to extract these invoices and verify their accuracy through five checks: schema, check digit, registry lookup, issuer name match, and tax arithmetic under three rounding conventions. On synthetic invoices, the model achieved 86.8% auto-approval with zero wrong fields in approved invoices.
Real tests found that 4.6% of approved invoices still had errors in fields not verified by checks, with OCR baseline being worse at 32%. The model had limitations with real invoices, such as clustered registered numbers and lack of individual data.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
