Guides

Synthetic invoice dataset tier comparison

Choosing a tier depends on whether you need a quick eval, a first fine-tune, or production-scale volume. All tiers share the same annotation schema.

Decision guide

  • Sample (50, free): mostly clean renders to inspect bbox quality, JSON schema, and manifest format before spending budget.
  • Starter (1,000): first model fine-tune or internal benchmark. Best value for small teams — available now.
  • Professional (10,000): production training runs with full degradation and multi-page coverage.
  • Scan Robust (10,000): 100% degraded samples plus clean/degraded pairs for scan-only OCR pipelines.
  • Enterprise (100,000): maximum volume with COCO export for object-detection workflows.

When Scan Robust beats balanced mode

If your production traffic is photographed or scanned paper, train on degraded-only data. Filter with degradation_tags[] for curriculum schedules. See degradation metadata guide.

Start with the free sample, then upgrade to Starter when your eval checklist passes.


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