agustinrojas/insurance-inbox-docs

Synthetic insurance documents (es/en)

530 synthetic insurance back-office documents, 265 in Argentine Spanish and 265 in English, in 13 types, rendered as clean digital files, phone photos, WhatsApp images, PDFs, bad photos and illegible photos, with the full ground-truth text of every file. Made for benchmarking OCR and document routing on the kind of attachments an insurance mailbox actually gets.

Everything is fictitious: companies, clinics, banks, brokers, people and numbers are made up. CUIT/CUIL and CBU numbers have valid check digits, so some may coincide with real ones by chance; they mean nothing.

Contents

Each language folder (es/, en/) has the image and PDF files, manifest.json (file, type, quality, title) and ground_truth.jsonl: one line per file with its full text line by line, the key fields (CUIT/CUIL, CBU, DNI, dates, amounts), the handwritten strings and the stamps.

Type	Count per language
Workers' compensation: accident report, initial care, test order, progress note, medical discharge, procedure authorization, provider invoice	151
Collections: bank transfer, cheque, deposit slip, allocation sheet	76
Quote request	19
Other (not insurance business)	19

Quality	Count per language
digital	84
phone photo	75
bad photo (blur, glare, angle)	78
illegible	13
native PDF	12
WhatsApp	3

How it was made and used

Generated by code (generate_insurance_docs.py in insurance-inbox-bench): templates filled with random fictitious data, handwriting fonts, stamps, then photo simulation (perspective, lighting, blur, JPEG). The same command regenerates the set identically. The benchmark, results and report are in that repository.

License

CC BY 4.0. The handwriting fonts used to render it are under the SIL Open Font License.