OCR fine-tuning checklist for synthetic invoices

Synthetic data accelerates early training, but you still need a disciplined eval loop. Use this checklist after downloading the free sample.

1. Baseline on clean data

Train on clean renders only; measure word-level CER/WER on the provided val split.
Spot-check 20 random word boxes against the image — tight alignment indicates good DOM labels.

2. Add degradation gradually

Filter samples with degradation_tags — start with tilted, then add stained and jpeg.
Compare val metrics clean vs. degraded; if gap is huge, prioritize Scan Robust tier.

3. Hold out real data

Keep 50–200 real invoices (with permission) as a sim-to-real eval set.
Do not mix real scans into training until synthetic baseline is stable.

4. Scale volume

Starter (1,000) for a first production candidate; Professional or Scan Robust for robustness sweeps.
Log config hash from manifest so experiments stay reproducible.

Ready to scale? View pricing tiers or read how to evaluate datasets.

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