Synthetic OCR + Document AI Training Data

Generate labeled synthetic documents for OCR and document AI.

DocSet Generator is a Windows & Linux desktop generator for OCR training data, document AI evaluation, and extraction pipeline testing. 33 document types, controllable degradation, and ML-ready exports — COCO, LayoutLM, FUNSD, and DocVQA — with zero real PII.

Purchase — $199 → Try the Live Demo One-time purchase. Free updates. Runs locally — nothing leaves your machine.
court_document.pdf
STATE OF CALIFORNIA
LOS ANGELES COUNTY SUPERIOR COURT
━━━━━━━━━━━━━━━━━━━━━━━
ASHLEY VINCENT, Plaintiff,
Case No.: 2020-CV-3360
invoice.pdf
INVOICE #7291
Date: March 24, 2026
Subtotal: $12,450.00
Tax (8%): $996.00
TOTAL: $13,446.00
receipt_Corrupt_Etext.pdf — OCR: 15%
RÒB1Ñ$OIN BI<
Dsle: QE/7?|2026 7îlna.
Srih7c+a] $3zQ20
Yoy {/%> $7Z91
TQLÂI $8q2·07
OCR DEGRADATION ACTIVE — 15%
33
Document Types
300
Docs / Second (clean)
0
Real PII Generated
100%
Local — Nothing Uploaded

See the output before you commit.

Download 50 real documents generated by DocSet Generator — across multiple types, clean and corrupted. No email required. No strings attached.

  • PDF documents across 10+ types
  • Clean and OCR-degraded versions
  • Corrupt text layer examples
  • Native formats included (.docx, .eml, .csv)
  • 50 documents total, ready to inspect
↓ Download Free Sample Pack

// 50 documents · .zip archive · No signup required · .7z version

// Or generate your own right now — try the live browser demo →

Corrupted receipt showing OCR degradation at 15%
Receipt — OCR Degraded 15%
Clean receipt output
Receipt — Clean 100%

Clean interface. No configuration required.

DocSet Generator main interface
Main Interface 33 document types organized by category. Select individual types or entire families. Estimated output count updates in real time.
Completed generation run with output files
Generation Complete COMPLETE status, generation time, output folder open with timestamped files. 5 documents with Bates stamping, watermarks, and corrupt text layer — 1.3 seconds.

Fast enough to not be your bottleneck.

ModeDocumentsTimeWorkers
Clean (100% quality)330 docs1.1s8 workers
OCR Degraded (50%)330 docs11s8 workers
Corrupt Text Layer330 docs11s4 workers
Corrupt Text Layer (scale)3,300 docs104s4 workers

// Worker count scales automatically based on your CPU and workload type. No configuration required.

01 — EXPORTS

ML-Ready Annotations

Export directly to COCO, LayoutLM, FUNSD, and DocVQA — word-detection boxes, token/label/normalized-box JSONL, question/answer/header roles with semantic links, and derived question-answer pairs. Train and fine-tune without writing a labeling pipeline.

02 — GROUND TRUTH

Canonical Text + Geometry

A pre-render document model captures the exact text of every type — titles, paragraphs, list items, tables with merged-cell spans — before any corruption. Ships as a ground-truth text sidecar per document plus word → line → block boxes and key/value pairs.

03 — DATASETS

Splits & Recipes

Deterministic, type-balanced train / validation / test splits with a configurable seed. Every run writes a replayable dataset_recipe.json, SHA-256 checksums, and a quality_report.json so datasets are reproducible and auditable.

04 — DEGRADATION

OCR Corruption Slider

Control exactly how hard your model has to work. From clean ground truth to heavily degraded scans — continuous spectrum, not presets. Realistic character substitutions based on actual OCR failure patterns.

05 — STEALTH

Corrupt Text Layer

Clean visual page, corrupted hidden text layer. Forces OCR fallback on tools that read embedded text directly. Tests the gap between what a document looks like and what an extractor actually reads.

06 — LEGAL

Bates & Image-Only PDFs

Sequential Bates identifiers on every page for eDiscovery and legal AI pipelines, plus fully flattened image-only PDFs with no selectable text — a true visual-only extraction challenge, generated at scale.

07 — REPRODUCIBLE

Seeded & Byte-Identical

Seed a run and regenerate it exactly, down to byte-identical PDFs on the same platform. Derived per-document seeds, a fixed reference date, and normalized metadata make clean/degraded twin pairs and dataset replays deterministic.

08 — PRIVACY

Zero Real PII

All names, addresses, companies, SSNs, account numbers, and financial figures are synthetically generated — reserved .example domains, 555-01xx phones, invalid SSNs. Mathematically accurate but entirely fake.

09 — LOCAL

Runs On-Prem

Windows and Linux desktop application. No internet required after install. No data sent to any server. Your training data stays on your machine — critical for regulated industries.

33 document types across every domain your pipeline will encounter.

General

  • Letter
  • Memo
  • Report
  • Fax
  • Meeting Notes
  • Scheduler
  • Transmittal

Business

  • Email
  • Mass CC Email
  • Invoice
  • Receipt
  • Check
  • Financial
  • Corporate
  • Presentation
  • Real Estate

Legal & Govt

  • Agreement
  • Court Document
  • Government
  • Patent
  • Certificate
  • Form

Data & Misc

  • Media
  • Documentation
  • Personal Info
  • Publication
  • Table / List
  • Transcript

Native Formats

  • Word (.docx)
  • Excel (.xlsx)
  • Email (.eml)
  • CSV (.csv)
  • Plain Text (.txt)

From OCR test data to a full training dataset.

New in v1.5

COCO · LayoutLM · FUNSD · DocVQA exports

Every run can emit ML-ready annotations: COCO word boxes in image pixels, LayoutLM token/label/normalized-box JSONL, FUNSD question/answer/header roles with links, and DocVQA Q&A pairs with word and page references. Drop it straight into training.

New in v1.5

Canonical ground truth & rich annotations

A pre-render model captures the true text of all 33 types — headings, tables, merged-cell spans — before any corruption. Annotation schema 2.1 adds word → line → block geometry, table/row/cell records, entities, and key/value pairs, with a ground-truth text sidecar per document.

New in v1.5

Reproducible datasets & recipes

Deterministic, type-balanced train/validation/test splits, a replayable dataset_recipe.json, byte-identical seeded PDFs, SHA-256 checksums, and a quality_report.json with per-type completeness metrics. Manifest schema 2.0 with built-in JSON Schema validation.

New in v1.5

Now on Linux, too

Ships as a Linux AppImage and Flatpak alongside the Windows build, with high-fidelity LibreOffice rendering for native formats when available. Automatic parallel generation in the GUI, and coverage expanded from 33 tests to 104.

Read the full v1.5 changelog →

One price. No subscriptions. No usage limits.

Built for ML engineers and QA teams who need training data now, not after a procurement process. Buy once, generate as many documents as you need, keep every update.

Questions? Contact us at
[email protected]

$199
One-time purchase · Windows & Linux
  • Windows .exe + Linux AppImage & Flatpak
  • 50-document sample pack included
  • 33 document types across 5 categories
  • COCO, LayoutLM, FUNSD & DocVQA exports
  • Ground-truth text + word/line/block annotations
  • Train / validation / test dataset splits
  • OCR degradation slider + corrupt text layer
  • Bates stamping, watermarks, image-only PDFs
  • Seeded, reproducible generation with recipes
  • Free updates — re-download anytime
  • Runs fully offline — no data leaves machine
Purchase Now →

Or download the free sample first

Common questions.

Anything else? Email [email protected]

Is any of the generated data real?
No. All names, addresses, companies, SSNs, EINs, account numbers, and financial figures are synthetically generated. The math is accurate but every entity is entirely fabricated. Safe to use in any environment.
Does it require an internet connection?
Only for the initial download. After that it runs entirely offline. No telemetry, no license servers, no API calls. Your generated data never leaves your machine.
What's the difference between OCR Degradation and Corrupt Text Layer?
OCR Degradation visually corrupts the document — characters are substituted, text becomes hard to read. Corrupt Text Layer keeps the document looking clean but corrupts the hidden embedded text, forcing tools to fall back to visual OCR. They can be used independently or together.
What OS does it run on?
Windows and Linux. Windows ships as a standalone .exe; Linux ships as an AppImage and a Flatpak. No Python installation or dependencies required on either platform.
What annotation and dataset export formats does it produce?
Every run can write ML-ready exports: COCO word-detection boxes in rendered-image pixels, LayoutLM token/label/normalized-box JSONL, FUNSD question/answer/header roles with semantic links, and DocVQA question-answer pairs with word and page references. It also emits word/line/block annotations with page geometry, a ground-truth text sidecar per document, clean page images, and deterministic train/validation/test splits — all validated against bundled JSON Schemas.
What happens when you release updates?
Updates are free for all existing customers. Re-download the latest version from your original purchase link anytime.
Can I use the generated documents for commercial ML training?
Yes. Documents generated by DocSet Generator are yours to use however you need, including as training data for commercial ML products.