nchhaya's picture
Match author spelling to the published paper
85ce76c verified
|
Raw History Blame Contribute Delete
8.24 kB
metadata
license: cc-by-4.0
language:
  - en
pretty_name: Synthetic Dense Documents
size_categories:
  - n<1K
task_categories:
  - image-to-text
tags:
  - synthetic
  - document-ai
  - enterprise-documents
  - hyperapi
  - dense-documents
  - multi-column
  - tables
  - reading-order
  - digitisation

Synthetic Dense Documents

60 high-resolution pages (1654 x 2339, A4 at 200 DPI) packed with small type. These are the pages where generic OCR loses reading order, merges neighbouring entries and drops table rows. There are five kinds, 12 pages each:

kind What it looks like
directory_listing Aged 3 to 4 column business and residential directory pages, with abbreviated occupations, house numbers and phone numbers, in serif or typewriter faces on tinted, lightly skewed paper
terms_and_conditions 2 to 3 column legal terms with up to 20 numbered clauses
long_statement Bank statements with 57 to 72 printed transactions in one table
financial_statements Balance sheets and P&L statements with five sections and comparative years
price_list Parts and price catalogues with 62 to 72 printed rows and 7 columns

The generator deliberately produces more content than fits. Overflow is cut at the page edge and the page is flagged truncated (24 of 60 pages: 6 directory, 3 terms, 7 statement and 8 price-list pages). Gold labels cover only what is printed.

The directory pages are modelled on a real digitisation project: HyperAPI parsed more than 1 million dense, multi-column directory pages in 10 days, keeping column boundaries, reading order and entry breaks intact, and turned each listing into a structured record.

Benchmarking a document AI vendor? Score HyperAPI on this dataset with a free API key (30 pages a month), try it with no code in the live demo, or request a proof of concept on your own documents.

Columns

Column Description
image 1654 x 2339 JPEG
kind, title Layout family and page title
blocks {id, order, label, bbox, text} in reading order (column by column)
tables {block_id, n_rows, n_cols, html}
num_blocks, num_table_rows Density indicators
truncated true if the generator's content didn't fit on the page. Gold labels cover only what's printed
aged_scan, degradation For directory pages: the aging and skew parameters (JSON). Boxes are already mapped
markdown, ocr_text Gold page content
from datasets import load_dataset
ds = load_dataset("hyperbots-inc/synthetic-dense-documents", split="test")
dirs = ds.filter(lambda x: x["kind"] == "directory_listing")
print(dirs[0]["ocr_text"][:400])

Suggested metrics

Use reading-order accuracy (does the text come out column by column?), entry-level exact match for directory listings, row recall and TEDS for tables, and CER on body text.

Related research

SAVIOR: Sample-efficient Adaptation of Vision-Language Models for OCR Representation (Bhat, Naganna, Haq, Khatri, Tamataam, Arun, Chhaya and Bhattacharya; WACV 2026 Workshops) adapts vision-language models to the OCR failure cases that public benchmarks under-represent; complex multi-column layouts and fine print are among the failure patterns it targets.

How the data was made

Every document is generated from a seeded random number generator by docgen.js and rendered on an HTML canvas in Chrome on macOS. Images use the system's print fonts and, for handwriting, the handwriting-style fonts installed with macOS (Bradley Hand, Noteworthy, Marker Felt, Chalkboard SE, Snell Roundhand, Apple Chancery, Savoye LET, Brush Script MT). Labels are taken from the generator itself, so there is no human annotation error. Field and block boxes are recorded at the moment each element is drawn.

All companies, people, addresses, account numbers and amounts are fictional. Phone numbers use ranges reserved for fiction where these exist (US 555-01xx, UK 07700 900xxx, AU 0491 570 1xx). Email addresses use the reserved example.com / example.org domains, and US ID numbers start with 9, which is never issued as an SSN.

To rebuild the dataset, load generator/docgen.js and generator/datasets.js in a browser page and run await DocSets.build('dense'). Rendering depends on installed fonts, so pixels can differ on other systems, but text and labels are identical for the same seed.

Limitations

  • The documents are synthetic. Their layouts, vocabulary and value ranges are much narrower than real enterprise traffic. Use them for benchmarking, regression tests and pipeline development, not as a replacement for evaluation on your own documents.
  • All content is in English, with values formatted for 9 locales (US, UK, IN, CA, AU, DE, FR, SG, AE).
  • Handwriting is rendered from fonts with per-character jitter. It is a stand-in for real handwriting, not a sample of it.

About HyperAPI

HyperAPI is Hyperbots' document intelligence API for enterprise documents. It provides parse, extract, classify, split and redact behind one endpoint and returns schema-validated JSON with a confidence score on every field. It has a REST API and a Python SDK, with an MCP server for agents available on request, and you can start free. See the HyperAPI Space for a walkthrough.

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 (this dataset) Small type, multi-column layouts and long tables
synthetic-invoice-extraction Invoice fields and line items (text + HTML)

License and citation

CC BY 4.0. Generator code is MIT.

@misc{hyperbots2026densedocs,
  title  = {Synthetic Dense Documents},
  author = {Hyperbots Inc.},
  year   = {2026},
  url    = {https://huggingface.co/datasets/hyperbots-inc/synthetic-dense-documents}
}