Guides

How to evaluate invoice OCR and extraction datasets

Before purchasing a dataset tier, run this checklist on the free sample or a vendor pilot pack.

Label quality

  • Spot-check 20 random words: does the bbox tightly cover glyphs?
  • Verify totals and tax fields against structured JSON values.
  • Confirm line-item rows align with table structure in the image.

Diversity

  • Count layout families and locales in the manifest statistics.
  • Check multi-page frequency if your AP stack sees long invoices.
  • Review QA contact sheets for duplicate-looking templates.

Degradation realism

  • Compare clean vs. degraded pairs if provided.
  • Train a tiny baseline; if val accuracy is unrealistically high, you may need harder degradation.

Release hygiene

  • SHA-256 checksums verify download integrity.
  • Config hash in manifest documents reproducibility.
  • Train/val/test splits should be template-aware to avoid leakage.

Get the 50-sample pack and run this checklist yourself.

Next: Get 50 free samples · View pricing · All guides