synthetic-receipts-ocr

32,000 synthetic thermal receipts across 5 locales (US/UK/DE/IT/FR) — each a clean render plus a photo-degraded twin, with pixel-exact word boxes, full transcription, and structured KIE fields.

Samples train-000357 (US), train-000073 (UK), eval-001179 (DE), train-000222 (IT), train-000711 (FR) — real dataset rows, not mockups. Each receipt is its sample's image_photo, cut out along its own homography quad; no retouching beyond composition.

Built for receipt OCR, document KIE (Donut/LayoutLM-style), text detection, and denoising. Every number on every receipt adds up: line totals × quantities, per-class contained VAT (ENTH. MwSt 7%, DONT TVA 5,5%, DI CUI IVA 22%), US sales tax, tendered and change. Models trained on receipts whose math is wrong learn to hallucinate totals; these don't lie.

What's in a sample?

column	what
image_clean	PNG, the rendered receipt (thermal-printer aesthetic, 2 font families)
image_photo	JPEG, photo-realistic degradation: perspective on a surface, uneven lighting, shadow bands, thermal fade, noise, blur, JPEG grunge
full_text	exact printed text, line by line
words	word-level boxes on image_clean — exact by construction (captured during rendering, never re-OCR'd)
words_photo	the same boxes mapped into image_photo space (axis-aligned hulls)
homography	the 3×3 matrix mapping clean → photo coordinates, so you can map anything else
fields	structured KIE ground truth: merchant, address, phone, tax id (P.IVA / USt-IdNr / SIRET / VAT No), date, time, receipt no, line items (name, qty, unit price, total, tax rate, source product title), subtotal, per-class taxes, total, payment, tendered, change, loyalty
locale, font, degradations, n_items	metadata for slicing

The ground-truth contract: a field is non-null if and only if its value is printed on the image. No invisible labels to hallucinate.

What does a training pair look like?

Every sample carries the same receipt twice — image_clean and image_photo — so you can train on the hard image and supervise with the easy one, or train denoising directly.

Sample train-000009 — its actual image_clean (left) and image_photo (right), unmodified apart from uniform scaling. Real dataset row, not a mockup.

And this is that same sample's fields — check it against the image above (line items truncated to 2 of 3 for display):

{
  "locale": "US",
  "merchant": "JANSEN'S EXPRESS",
  "address": "274 BROADWAY",
  "phone": "+1 (519) 644-6909",
  "tax_id": null,
  "date": "09/09/2024",
  "time": "13:56",
  "receipt_no": "1-95787-340",
  "cashier": "TILL 3",
  "currency": "USD",
  "lines": [
    {
      "name": "DII COT TERRY WINDOWPANE",
      "qty": 1,
      "unit_price": 0.98,
      "total": 0.98,
      "tax_rate": 0.095,
      "source_title": "DII Cotton Terry Windowpane Dish Cloths, 12 x 12\" Set of 6, Machine Washable and Ultra Absorbent Kitchen Bar Towels-Soli"
    },
    {
      "name": "AMERICAN ANGLER CLLCTN",
      "qty": 1,
      "unit_price": 49.99,
      "total": 49.99,
      "tax_rate": 0.095,
      "source_title": "American Angler Collection Toy Fish Set | Toy Fish Figurines | | Largemouth Bass Catfish Bluegill Crappie Striper | Educ"
    }
    ...
  ],
  "subtotal": 100.96,
  "taxes": [
    {
      "label": "TAX",
      "rate": 0.095,
      "amount": 9.59
    }
  ],
  "tax_included": false,
  "total": 110.55,
  "payment": "CONTACTLESS",
  "tendered": null,
  "change": null,
  "loyalty": null
}

Note tax_id: null — this US receipt prints no tax id, so the field is null. The contract again.

How exact are the boxes?

words boxes are captured during rendering, never re-OCR'd. words_photo maps them into photo space through the sample's homography — and since the matrix ships with every sample, you can map anything else the same way.

Sample train-000651 — real dataset row, not a mockup. Every box drawn 1:1 from the sample's words / words_photo JSON, no adjustment; the inset is a plain 3× crop of the clean render around the total.

Why synthetic — and what does that honestly mean?

Real receipt datasets are small (CORD ~1k, SROIE ~1k) because receipts are private. Synthetic generation trades realism risk for scale, perfect labels, and zero privacy exposure. This dataset is 100% synthetic and says so — nothing here is presented as real. What keeps it less artificial than most:

• Real product vocabulary — 60,000 titles from the Amazon ESCI dataset (Apache-2.0), run through a thermal-printer abbreviator (NNSTCK FRY PAN 12), with the untruncated source title kept in the ground truth.
• Enforced locale realism — real per-country VAT rate structures, locale-true currency/decimal/date formats (1.234,56 EUR, 5,5%), localized section labels (SCONTRINO/BELEG/TICKET, TOTALE/SUMME, RESTO/RUECKGELD), locale-correct phone formats and address order, printed tax registration numbers.
• Three rounds of adversarial QA before publication — independent inspectors recomputed every number from pixels and ground truth, checked box alignment at 4× zoom, and audited locale consistency. A pilot round found 7 systematic defect classes (misbound quantity lines, glyph-clipping boxes, a VAT-label rounding bug, unprintable ground truth); a second round on production samples found 2 more (a printed-name/label desync on narrow layouts, a trademarked footer slogan); category-blind VAT assignment and anglophone merchant names on EU receipts were also caught (reduced VAT now goes to food-adjacent items only; Italian receipts come from MERCATO ROSSI, not ROSSI'S GROCERY). Everything was fixed and all 32,000 samples regenerated. Finally, every sample passed a machine validation gate — arithmetic, locale conventions, the printed-only contract — shipped in this repo as generator/validate.py.

Limitations, plainly: monospace thermal-style receipts only — no proportional-font layouts, no handwriting, no 3D crumple, no logos/barcodes yet. Five locales, Latin script. The degradations are parametric, not photographs. Transfer to real receipt photos is plausible but unmeasured until someone evaluates on real data — if you do, please share numbers in a discussion.

What does the eval split actually measure?

Generalization, not memorization. eval (2,000 receipts) shares no font (JetBrains Mono, unseen in train), no product title (held-out vocabulary slice), and no merchant surname with train (30,000). Same generator, disjoint surface distribution — eval performance means your model learned receipt structure, not these particular pixels.

How do I use it?

from datasets import load_dataset
import json

ds = load_dataset("albertobarnabo/synthetic-receipts-ocr", split="train")
s = ds[0]
img = s["image_photo"]                 # PIL.Image, decoded for you
fields = json.loads(s["fields"])
words = json.loads(s["words_photo"])
print(fields["merchant"], fields["total"], len(words), "words")

Train a Donut-style KIE model on (image_photo, fields); OCR/text recognition on (image_photo, full_text); text detection on (image_photo, words_photo); denoising on (image_photo → image_clean). The homography supports anything geometric.

Where does it come from?

The entire generator ships in this repo (generator/): content model, renderer, degradation pipeline, validation gate, and the OFL fonts. Every receipt is a pure function of (seed, index) — regenerate, extend to new locales, or scale to millions. Product vocabulary derives from tasksource/esci (Apache-2.0). Fonts: Space Mono, IBM Plex Mono, JetBrains Mono (OFL).

Related work from the same author: the e-commerce search stack — the ESCI-trained retrieval models whose corpus supplies this dataset's product vocabulary.