hyperbots-inc/synthetic-bad-scans

Synthetic Bad Scans

120 enterprise documents rendered and then degraded the way real scanners, fax machines and phone cameras degrade them. There are 40 images at each of three severity levels (mild, moderate, severe). Every image comes with clean gold text, field values and boxes mapped into the skewed image's coordinates, plus the exact degradation parameters, so you can plot accuracy against each kind of damage.

HyperAPI deskews, corrects rotation and enhances images before extraction. Its documentation recommends scanning at 300 DPI or higher, because very low resolution reduces accuracy. This set shows where that threshold sits for your own pipeline.

Degradations

Effect	mild	moderate	severe
Rotation / skew	±1.5°	±4°	±8°
Effective resolution (downsample, then upsample)	112-142 DPI	75-112 DPI	50-82 DPI
Gaussian blur	up to 0.6 px	up to 1.2 px	up to 2 px
Sensor noise (sigma)	up to 6	up to 14	up to 26
Salt & pepper, dust specks	light	medium	heavy
Contrast	0.85-1.0	0.65-0.9	0.45-0.75
JPEG quality	0.55-0.80	0.30-0.55	0.12-0.32
Shadows, folds, punch holes, faded toner streaks	occasional	frequent	most pages

About a quarter of pages also carry a stamp, and some forms are handwritten.

Columns

Column	Description
image	1240 x 1754 degraded JPEG
severity	mild / moderate / severe
degradation	Exact parameters (rotation_deg, effective_dpi, blur_px, noise_sigma, salt_pepper, contrast, brightness, jpeg_quality, shadow, fold_y, punch_holes, toner_streaks, dust_specks, paper)
has_stamp	Whether a stamp was added before degrading
fields	{key, label, value, bbox, handwritten} with boxes in degraded-image coordinates
blocks	Layout blocks with mapped boxes
labels	JSON string of normalised gold values
ocr_text	Clean gold text

from datasets import load_dataset
ds = load_dataset("hyperbots-inc/synthetic-bad-scans", split="test")
severe = ds.filter(lambda x: x["severity"] == "severe")
print(severe[0]["degradation"])

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('bad_scans'). 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 (this dataset)	Skew, blur, noise, shadows and compression at 3 severities
synthetic-dense-documents	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{hyperbots2026badscans,
  title  = {Synthetic Bad Scans},
  author = {Hyperbots Inc.},
  year   = {2026},
  url    = {https://huggingface.co/datasets/hyperbots-inc/synthetic-bad-scans}
}