Synthetic Invoice OCR Ground Truth Dataset Generator

Generates synthetic invoice test documents with exact OCR ground truth: for every page you get a PNG image, a raster-only PDF with the same pixels, and an annotation JSON with every word's bounding box, line structure, reading order, table cells and typed invoice fields. Labels come from the renderer itself, not from a human annotator or a second OCR engine, so they are exact.

Use it to score an OCR, IDP or document-understanding pipeline against known-correct labels, to build a reproducible invoice dataset of any size for CI and regression tests, or to benchmark extraction accuracy across clean and scan-degraded renders of the same geometry. Every page is visibly marked SYNTHETIC TEST DOCUMENT - NOT PAYABLE, and all names, addresses, identifiers and amounts are invented.

Who uses this

• OCR and IDP engineering teams measuring word error rate, box IoU, reading order, and field extraction accuracy against exact labels, without collecting, redacting or licensing real invoices.
• Document AI and LLM evaluation teams that need a labeled invoice dataset for key-information extraction (KIE) benchmarks, described by its input seed instead of stored and shipped around.
• Pipeline and QA teams feeding a document pipeline a fixed, reproducible corpus in CI instead of a folder of real customer files.
• Robustness testing comparing clean renders against deterministic scan degradation at the same geometry.

What you get

Each run produces one to sixteen single-page, invoice-shaped test documents. For every page:

• Page image (PNG) at 144, 200, or 300 DPI
• Raster-only PDF containing exactly the same final pixels as the PNG
• Annotation JSON with the full ground truth:
  • tokens: every word with text, pixel and point bounding boxes, font, and reading order
  • lines: line text, boxes, role, and the tokens each line contains
  • entities: typed field labels (invoice number, dates, vendor and customer blocks, subtotal, tax, total) linked back to their tokens and lines
  • tables: table regions with rows, columns, column names, and per-cell boxes
  • invoice: the structured record behind the page, including line items and reconciled decimal amounts
  • page: dimensions in pixels and points, DPI, bbox format, and coordinate origin
  • full_text, seed, schema_version, and SHA-256 hashes of the delivered image and PDF

Dataset: one item per successfully generated and validated page:

{
  "schemaVersion": "invoice-ground-truth/v0",
  "generatorVersion": "0.1.0",
  "documentId": "SYN-0000002D-0003",
  "seed": 45,
  "templateId": "invoice_grid_v0",
  "profile": "scan_light",
  "dpi": 200,
  "widthPx": 1700,
  "heightPx": 2200,
  "tokenCount": 112,
  "entityCount": 41,
  "recordKeys": {
    "image": "SYN_0000002D_0003_PAGE_PNG",
    "pdf": "SYN_0000002D_0003_PAGE_PDF",
    "annotations": "SYN_0000002D_0003_ANNOTATIONS_JSON"
  },
  "sha256": { "image": "...", "pdf": "...", "annotations": "..." },
  "synthetic": true,
  "payable": false
}

Key-value store: three records per page, named by the recordKeys above: the PNG, the raster PDF, and the annotation JSON. A run-level OUTPUT record reports requestedPages and completedPages.

Fetch the dataset for the index, then pull each artifact by its key. Every dataset item is only pushed after its three artifacts are stored, so a recordKeys entry always resolves.

Input

Field	Type	Default	Notes
seed	integer	required	First deterministic document seed, 0 to 2,147,483,647. Page n of a run uses seed + n.
count	integer	1	Number of pages to generate, 1 to 16.
template_id	string	auto	auto, or one of four structural layouts: invoice_classic_v0, invoice_compact_v0, invoice_grid_v0, invoice_sidebar_v0.
profile	string	clean	clean, scan_light, or scan_hard. Scan profiles apply deterministic degradation and keep the ground-truth geometry aligned to the rendered pixels.
dpi	integer	144	Exactly 144, 200, or 300. Any other value is rejected.
schema_version	string	invoice-ground-truth/v0	Pinned output contract.

Unknown input fields are rejected rather than silently ignored, so a typo fails fast instead of producing an unexpected corpus.

Example inputs

A clean 16-page evaluation batch, mixed layouts (OCR accuracy baseline):

{
  "seed": 1000,
  "count": 16,
  "template_id": "auto",
  "profile": "clean",
  "dpi": 300
}

The same pages degraded as a hard scan (robustness comparison at identical geometry: same seeds, different profile):

{
  "seed": 1000,
  "count": 16,
  "template_id": "auto",
  "profile": "scan_hard",
  "dpi": 300
}

A fixed single-layout CI fixture (regression test that must never change):

{
  "seed": 42,
  "count": 4,
  "template_id": "invoice_grid_v0",
  "profile": "scan_light",
  "dpi": 200
}

Determinism

A given seed, template_id, profile, dpi, and Actor version always produce byte-identical output. The SHA-256 hashes in each dataset item let you verify this. Re-running the same input reproduces the same corpus, and adjacent seeds give you independent documents, so an evaluation set can be described by its input alone instead of being stored and shipped around. Generate larger corpora by launching several runs with non-overlapping seed ranges.

Pricing

$0.05 per generated page. You pay only for pages that are generated, validated, and delivered. A page that fails validation is not charged.

Data source and licence

There is no external data source. Every page is rendered by this Actor from its own templates and a seeded generator; the Actor accepts no caller-supplied document text, logos, addresses, images, URLs, or templates, and reads nothing from the web. Output is yours to use for testing and evaluation under the restrictions below.

Limitations

• 1 to 16 pages per run. Generate larger corpora by launching several runs with non-overlapping seed ranges.
• Single-page English USD invoice-shaped fixtures only.
• Four structural layouts; clean, light-scan, and hard-scan profiles; 144, 200, or 300 DPI.
• Labels are renderer-owned, not human-reviewed. They describe exactly what was drawn.
• This Actor does not reproduce the statistical distribution of real invoices and makes no privacy, compliance, or production-readiness claim.

Not for

These pages are test fixtures, not documents. They contain no real people, companies, or transactions, they are visibly watermarked as synthetic and not payable, and they must not be used for billing, payment requests, identity construction, deception, or fraud.

FAQ

Where can I get an invoice OCR dataset with ground truth?

Here, on demand. Each page ships with word-level boxes, lines, reading order, table cells and typed fields produced by the renderer, so the labels are exact rather than annotated. Set seed and count and the dataset is defined by those two numbers.

What does the OCR ground truth JSON contain?

tokens (every word with text, pixel and point boxes, font and reading order), lines, entities (invoice number, dates, vendor and customer blocks, subtotal, tax, total), tables with per-cell boxes, the structured invoice record with line items, page geometry, full_text, the seed, the schema version and SHA-256 hashes of the image and PDF.

Can I use this to benchmark invoice data extraction or key-information extraction?

Yes. entities links each typed field to its tokens and lines, and invoice carries the reconciled decimal amounts and line items, so you can score field-level extraction as well as raw OCR.

How do I test OCR on scanned documents versus clean ones?

Run the same seed, count, template_id and dpi twice with profile set to clean and then scan_light or scan_hard. The degradation is deterministic and the ground-truth geometry stays aligned to the rendered pixels.

Is the output reproducible?

Yes. The same input on the same Actor version produces byte-identical PNG, PDF and JSON; the SHA-256 hashes on every dataset item let you verify it.

Are these real invoices or real companies?

No. Every name, address, identifier and amount is invented, and every page is visibly marked as a synthetic test document that is not payable.

Related datasets

Other datasets by the same author, all pay-per-record with source, refresh cadence and licence stated on each listing:

Nursing home and long-term care data

• Nursing Home Deficiency Citations (CMS) Scraper
• Nursing Home Fines & Civil Money Penalties (CMS) Scraper
• Nursing Home Directory & Ownership (CMS Form 671) Scraper
• Nursing Home Daily Nurse Staffing (CMS PBJ) Scraper
• Nursing Home Change of Ownership (CMS SNF CHOW) Scraper
• New York Nursing Home Bed Availability (NYSDOH) Scraper

Medicare provider enrollment data

• Medicare Revoked Providers List (CMS Revocations) Scraper
• Medicare Opt-Out Providers List (CMS Affidavits) Scraper
• Medicare Revalidation Due Date List (CMS) Scraper

Public procurement and RFP feeds

• E-Rate Form 470 RFP Scraper (USAC School & Library Bids)
• NYC PASSPort Procurement Scraper (RFPs, Bids & Contracts)