Separate the document from the task
A request for invoices can support very different projects: classifying document types, locating fields, extracting line items, or answering questions. State the intended task and the exact output that will count as a correct result. This helps determine whether labels need to describe a whole file, a page, a region, or an individual field.
Describe meaningful diversity
Ask for variation that resembles deployment: languages, countries, issuing organizations, scan quality, and document layouts. State whether duplicate templates are useful or should be limited. Distinguish the number of files from the number of unique layouts and business cases.
Specify the original and derived files
List required formats, including source documents, page images, extracted text, tables, or annotations. Explain how they should be linked. For multi-page documents, decide whether page order and document boundaries must be retained.
Set handling rules before samples
Describe excluded information and the desired replacement or masking process. Establish a way to evaluate redaction quality without distributing sensitive originals unnecessarily. Commercial use, confidentiality, and onward-sharing permissions should be discussed for the particular source.
Review checklist
- Document types or table schemas
- Languages, regions, and date coverage
- Volume by category
- Ground truth and sensitive-data requirements
- Intended use and licensing scope