GaborMadarasz/synthetic_OCR_dataset

Synthetic Hungarian Document OCR Dataset

Dataset Description

This dataset contains 3,000 synthetically generated Hungarian document images paired with their ground-truth text transcriptions. It is designed for training and evaluating optical character recognition (OCR), document layout analysis, and multimodal language models on Hungarian administrative and business documents.

Key Characteristics

• Language: Hungarian (hu), with full support for diacritics (á, é, í, ó, ö, ő, ú, ü, ű)
• Document Types:
  • Invoices (SZÁMLA) with itemized tables, VAT calculations, and bank details
  • Formal letters (HIVATALOS LEVÉL) with sender/recipient headers and structured body text
  • Forms (ŰRLAP) with labeled fields (name, address, tax ID, TAJ, etc.)
  • Freeform documents with mixed content (headers, snippets, lists, amounts)
• Visual Variability:
  • Random system fonts (TTF/OTF), font sizes (10–22 pt)
  • DPI range: 150–300
  • Optional Gaussian noise (30% probability, severity 0.02–0.08)
  • Optional slight rotation (±3°, 20% probability)
  • Two rendering backends: Pillow (default) and ReportLab+ImageMagick (for structured invoices, if dependencies available)
• Text Content:
  • Realistic Hungarian names, addresses, cities, street names
  • Valid-format tax codes (HUxxxxxxxx), IBAN/BIC codes, invoice numbers
  • HUF currency formatting with thousand separators
  • Domain-specific vocabulary (products, services, legal phrases)

Supported Tasks

• Optical Character Recognition (OCR)
• Handwritten-style text recognition (synthetic degradation)
• Document layout understanding
• Multimodal pretraining for Hungarian document AI
• Synthetic data augmentation for low-resource OCR

Data Fields

Field	Type	Description
images	list[string]	List containing one relative path to the PNG image (e.g., images/img_000001.png)
texts	string	Ground-truth Hungarian text, newline-separated, exactly as rendered in the image

Example Record

{
  "images": ["images/img_000042.png"],
  "texts": "SZÁMLA\nNagy Kft.\nBudapest 1052, Fő utca 15.\nAdószám: HU12345678\n\nVevő: Kovács Anna\nSzeged 6721, Kossuth Lajos utca 42.\n\nSzámla szám: FV/842101\nKészítés dátuma: 2025. 03. 15.\n----------------------------------------\nTétel                     Menny.   Egységár    Összesen\n----------------------------------------\nSzámítógép                2        150 000 Ft    300 000 Ft\n..."
}

Data Splits

• Single split with 3,000 samples
• Composition:

  • ~60% template-based documents (invoices, letters, forms; ~1,800 samples)

  • 40% freeform documents with randomized structure (1,200 samples)

Creation Process

1. Content Generation: Hungarian text snippets, names, addresses, and domain-specific fields are sampled from curated lists.
2. Layout Assembly: Documents are assembled using predefined templates (for structured types) or randomized composition (freeform).
3. Image Rendering:
   • Text is rendered onto white-background PNGs using PIL/Pillow with random fonts, sizes, and colors.
   • Optional post-processing: Gaussian noise, slight rotation.
   • For invoices: optional high-fidelity PDF rendering via ReportLab, converted to PNG via ImageMagick (if available).

Considerations for Using the Data

• Synthetic Nature: Images are computer-generated; they lack real-world scanning artifacts (e.g., skew, shadows, paper texture). Consider domain adaptation for production OCR.
• Hungarian Specificity: Diacritics and Hungarian orthography are preserved; models trained on this data may require fine-tuning for other languages.
• Privacy: All personal data (names, addresses, IDs) is synthetically generated and does not correspond to real individuals.

Limitations and Bias

• Font Coverage: Depends on system fonts available during generation; may not represent all typographic styles used in Hungarian documents.
• Layout Diversity: Template-based samples follow fixed structures; freeform samples add variability but may not cover all real-world document layouts.
• No Handwriting: All text is machine-rendered; not suitable for handwritten text recognition without additional augmentation.

Citation

@dataset{synthetic-hu-docs-2026,
  title = {Synthetic Hungarian Document OCR Dataset},
  author = {Gabor Madarasz},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/GaborMadarasz/synthetic_OCR_dataset}
}

Contact

For questions or contributions, please open an issue on the dataset repository or contact the author directly.

---