hyperbots-inc/synthetic-invoice-extraction

Synthetic Invoice Extraction

A fully synthetic set of 200 invoices for testing key-information and line-item extraction. Each invoice comes in three forms: a rendered HTML document, an OCR-style plain-text version that keeps the layout, and gold labels in the same shape as the output of HyperAPI's POST /v1/extract.

It was built by Hyperbots so developers can benchmark HyperAPI, or any other extraction pipeline, on data they're free to share, without exposing real financial documents. It is one of ten datasets in the Hyperbots synthetic document suite, which also covers 31 enterprise document types, parsing, classification, splitting, PII redaction, handwriting, stamps, bad scans and dense pages (see the table below).

All data is synthetic. Every company, person, address, invoice number and amount was generated at random by generate.js. Any resemblance to a real business is coincidental.

What makes it non-trivial

Variation	Values
Layouts	classic, modern, compact, two-column (header split into two columns), ledger (up to 12 line items)
Countries / currencies	US (USD), UK (GBP), India (INR), Canada (CAD), Australia (AUD), Germany and France (EUR)
Date formats	MM/DD/YYYY, DD/MM/YYYY, Mon D, YYYY, D Month YYYY, ISO. Gold labels are always ISO YYYY-MM-DD
Field labels	e.g. "Invoice No." / "Invoice #" / "Bill No."; "Total" / "Amount Due" / "Balance Due" / "Grand Total"
Tax	Sales Tax, VAT, GST or GST/HST at country-appropriate rates, including 0%
Parties	Both vendor and bill-to blocks, so models have to pick the right company name

Every row is arithmetically consistent: the line totals add up to the subtotal, and the subtotal plus tax equals the total.

Schema

Column	Type	Description
id	string	syn-inv-0001 … syn-inv-0200
layout	string	Layout template
country	string	Vendor country code
date_format	string	us, eu, iso, short, long
num_line_items	int	Number of line items
ocr_text	string	Plain text in reading order, the way an OCR engine would return it
html	string	Self-contained HTML rendering of the invoice (render to PDF or PNG for vision models)
labels.entities	object	invoice_number, date, due_date, vendor_name, vendor_address, bill_to_name, bill_to_address, po_number, payment_terms, currency, subtotal, tax_rate, tax_amount, total_amount
labels.line_items	list	description, quantity, unit_price, total

Amounts are strings formatted like HyperAPI's output (e.g. "1,250.00"), and currency is an ISO 4217 code.

from datasets import load_dataset
ds = load_dataset("hyperbots-inc/synthetic-invoice-extraction", split="test")
print(ds[0]["ocr_text"])
print(ds[0]["labels"]["entities"]["total_amount"])

Benchmark HyperAPI in a few minutes

evaluate_hyperapi.py uploads each invoice to HyperAPI's Extract endpoint and scores the six core fields (invoice_number, date, due_date, vendor_name, total_amount, currency) plus line-item recall.

# install the HyperAPI SDK: https://apis.hyperbots.com/docs/quickstart
export HYPERAPI_KEY=hk_test_...      # free key, 30 pages/month: https://apis.hyperbots.com/dashboard/api-keys
python evaluate_hyperapi.py --limit 25          # sends the OCR text as .txt uploads
python evaluate_hyperapi.py --limit 25 --pdf    # renders the HTML to PDF first (pip install weasyprint)
python evaluate_hyperapi.py --dry-run           # checks the scorer with gold labels, no API calls

Regenerating or extending

generate.js is deterministic. The same seed always produces byte-identical output.

node generate.js > data/test.jsonl   # seed 20251113, 200 rows

Change the seed or row count at the bottom of the file to create a bigger training set or a fresh held-out set.

Background

The approach follows Hyperbots' DocuLite research. DocuLite generates layout-preserving synthetic invoice text (InvoicePy) and HTML invoice templates (TemplatePy) so LLMs and VLMs can be adapted to invoice extraction without human-annotated data. The work was accepted at the 4th Deployable AI Workshop at AAAI 2026. Read the summary.

This dataset is a small, public, illustrative set in that spirit. It is not the DocuLite training data and not the private Challenging Invoice Extraction dataset.

Limitations

• The data is clean and English-only. There are no scan artifacts, handwriting, stamps or skew. Render the HTML and add noise if you want to test robustness.
• Product names and address formats come from small lists and are less varied than real-world data.
• Line items are always in a single table with four columns.

Part of the Hyperbots synthetic document suite

Dataset	Focus
synthetic-enterprise-doc-extraction	Field extraction across 31 enterprise document types
synthetic-doc-parsing	Layout blocks, reading order, Markdown and tables
synthetic-pii-redaction	PII spans and boxes for masking and de-identification
synthetic-doc-classification	31-class document classification
synthetic-doc-splitting	Multi-document bundles with true boundaries
synthetic-handwritten-forms	Forms filled in handwriting-style script
synthetic-stamped-documents	Stamps and seals over printed text
synthetic-bad-scans	Skew, blur, noise, shadows and compression at 3 severities
synthetic-dense-documents	Small type, multi-column layouts and long tables
synthetic-invoice-extraction (this dataset)	Invoice fields and line items (text + HTML)

HyperAPI provides parse, extract, classify, split and redact for enterprise documents behind one endpoint, returning schema-validated JSON with per-field confidence. Start free.

License and citation

CC BY 4.0. If you use the dataset, please cite:

@misc{hyperbots2026syntheticinvoices,
  title  = {Synthetic Invoice Extraction: an evaluation set for HyperAPI},
  author = {Hyperbots Inc.},
  year   = {2026},
  url    = {https://huggingface.co/datasets/hyperbots-inc/synthetic-invoice-extraction}
}

Links: HyperAPI · Docs · HyperAPI Space · Python SDK · Hyperbots