Invoice OCR test data: free synthetic invoices with ground truth
GritWorks offers free synthetic invoice PDFs for testing invoice OCR and extraction. Each goods or services invoice comes as a clean PDF and a degraded copy (blur or coffee stain), with JSON ground truth for header fields, line items, tax and totals, plus bounding boxes. Every supplier and amount is fictional.
Two invoices from one fictional supplier to one customer: a goods invoice with three SKU lines, freight and tax, and a services invoice billed by hours and rate against a work order. Both quote the same purchase order number, so the pair also tests reference extraction. The tables below list every annotated field with an example value.
Clean vs degraded: blur and coffee stain
Each invoice ships twice. The goods invoice has a strong-blur copy and the services invoice a coffee-stained copy, and both degraded files are scored against the same expected values as the clean ones. Run the pair side by side to separate errors your pipeline makes on any invoice from errors the capture condition introduces.
How to score invoice extraction
Score header fields (supplier, invoice number, dates, PO), line items and totals separately. A line item passes only when SKU, description, quantity, unit price and amount are all right in the same row. Then reconcile: line amounts should sum to the subtotal, and subtotal plus freight plus tax to the total due.
Normalising dates, currency and whitespace
The invoices print dates as MM/DD/YYYY and amounts with a currency symbol; the JSON stores plain values. Agree the normalisation rules before you compare (date format, currency symbols, thousands separators, case and whitespace) and keep them fixed between runs, or a formatting change will look like an accuracy change.
Linked documents from the same trading relationship
The OCR pack also holds the purchase order, a credit memo and an account statement for the same supplier and customer. The PO total equals the two invoices combined, the credit memo cites the goods invoice, and the statement lists all three, so you can test matching across documents as well as extraction.
FROM THE PACK’S GROUND-TRUTH JSON
Fields with ground truth
Every annotated field in the 2 documents, with one example value. Repeated rows such as line items are shown once. Identifier values are withheld from this page and stay in the pack.
These documents ship in the OCR & extraction pack: 10 PDFs, five clean and five degraded, with structured JSON and bounding boxes. Enter your name, email and company and we’ll email the link. No call or platform account is required for the sample.
The sample is a starting point. We can support the volume your project needs. Book a call to discuss your document types, languages, formats and target volume.
We also provide Excel spreadsheets and images in JPEG, PNG and TIFF formats. Qualified teams can receive the full PDF dataset free after the booking form and call.
The free offer covers PDFs. Larger volumes and other formats are scoped on the call. Product support is available around GritWorks products.
QUESTIONS
Invoice test data, answered.
Are these sample invoice PDFs free to use?+
Yes. They are part of the free OCR & extraction sample pack, licensed CC BY 4.0: you may use, adapt and share them, including commercially, with attribution to GritWorks. Every supplier, customer and amount is fictional.
Do the invoices include line-item ground truth?+
Yes. Each line has its SKU (or service description), quantity or hours, unit price or rate and amount in the JSON, with a bounding box for each value, alongside the header fields, subtotal, freight, tax and total due.
Can I use these invoices to train an invoice OCR model?+
Use them to prototype a training or fine-tuning pipeline and to build regression tests. Two invoice layouts are not a training corpus: a model trained on them would learn these layouts, not invoices in general. Larger and more varied volumes are scoped on a call.
Do you have VAT, GST or non-English invoices?+
Not in the free sample, which holds two US-style invoices. VAT, GST, multi-currency and non-English layouts can be generated with GritRender; book a call to describe the layouts and volume you need.
How do I get the sample?+
Choose “Email me this pack” and enter your name, email and company. We’ll email the download link. No call or platform account is required for the sample.
Can I use my own tools?+
Yes. Use the PDFs and reference files in your own evaluation, sandbox or testing workflow. The included reference data is not an automated scoring system. The sample pack is licensed CC BY 4.0: you may use, adapt and share it, including commercially, with attribution to GritWorks. Third-party names or logos that appear in some documents belong to their owners and are not covered by the license.