GaborMadarasz/synthetic_OCR_dataset_v2

Hungarian Synthetic OCR Dataset

A synthetically generated image–text dataset for fine-tuning OCR models on Hungarian text, with particular focus on correct recognition of Hungarian-specific accented characters (á, é, í, ó, ö, ő, ú, ü, ű).

The ground-truth text is sourced from Hungarian Wikipedia. Images are rendered programmatically using PIL and augmented with Albumentations, across four visual profiles that simulate real-world document conditions.

---

Dataset Summary

Property	Value
Language	Hungarian (hu)
Total images	~100,000
Ground truth source	Hungarian Wikipedia
Image generation	PIL + Albumentations
License	CC BY-SA 4.0
Task	Optical Character Recognition (OCR)
Format	HuggingFace datasets, Image feature + JSON conversation

---

Visual Profiles

The dataset contains four distinct visual profiles, each targeting a different real-world document type:

modern (30,000 images)

Clean, single-line printed text on a near-white background. Minimal augmentation. Simulates laser-printed or digitally typeset documents.

• Background: white to very light grey (RGB 245–255)
• Fonts: DejaVu Sans family, Liberation Sans, FreeSans (regular, bold, oblique, condensed, mono)
• Font size: derived from fixed 48px canvas height, fill ratio 55–78%
• Augmentation: subtle brightness/contrast shift, very light Gaussian noise

scanned (30,000 images)

Single-line serif text on a warm off-white background. Moderate augmentation simulating photocopier or flatbed scanner output.

• Background: warm off-white with subtle grain (RGB ~220–240 / 212–232 / 185–210)
• Fonts: DejaVu Serif family, Liberation Serif, FreeSerif (regular, bold, italic, condensed)
• Font size: derived from fixed 52px canvas height, fill ratio 55–78%
• Augmentation: mild Gaussian blur, brightness/contrast variation, light noise, occasional JPEG compression

archiv (20,000 images)

Single-line serif text on aged yellowed paper. Simulates historical or archival documents.

• Background: yellowed paper with stronger grain and occasional foxing spots (RGB ~195–225 / 180–210 / 145–175)
• Fonts: FreeSerif, DejaVu Serif, Liberation Serif, FreeMono (italic and condensed variants)
• Font size: derived from fixed 52px canvas height, fill ratio 55–78%
• Augmentation: Gaussian blur, stronger brightness/contrast shift, moderate noise, JPEG compression, gamma variation

paragraph (20,000 images)

Multi-line text blocks (3–8 lines per image) on a scanned-style background. Simulates book pages, newspaper columns, or typed documents.

• Background: warm off-white (same as scanned profile)
• Fonts: mixed serif and sans-serif (DejaVu, Liberation, Free families)
• Font size: 13–19px fixed range; word-wrapped to fit canvas width (740px)
• Layout variation: random line spacing (2–10px), occasional first-line indent (30% probability)
• Augmentation: same pipeline as scanned

---

Dataset Structure

Each sample contains the following fields:

Field	Type	Description
image	Image	PIL image (JPEG, variable size)
messages	string	JSON-serialized chat conversation (see below)
profile	string	Visual profile: modern, scanned, archiv, paragraph
multiline	bool	True for paragraph profile, False for single-line profiles
char_count	int32	Character count of the ground truth text

messages format

The messages field is a JSON string containing a three-turn conversation in the format expected by vision-language model instruction tuning (compatible with Unsloth FastVisionModel and standard chat templates):

[
  {
    "role": "system",
    "content": "Te egy pontos optikai karakterfelismerő (OCR) asszisztens vagy. ..."
  },
  {
    "role": "user",
    "content": [
      {"type": "image"},
      {"type": "text", "text": "Írd le pontosan a képen látható szöveget!"}
    ]
  },
  {
    "role": "assistant",
    "content": "A képen látható pontos szöveg."
  }
]

Six prompt variants are used in rotation to improve robustness:

• Írd le pontosan a képen látható szöveget!
• Mi a képen szereplő szöveg?
• Olvasd el a képen látható szöveget, és írd le szó szerint!
• Kérlek, írd át a képen lévő szöveget!
• Add meg a képen látható feliratot pontosan!
• Milyen szöveg szerepel ezen a képen?

---

Text Corpus Statistics

Ground truth sentences were extracted from Hungarian Wikipedia and filtered to ensure quality:

Metric	Value
Total sentences	200,000
Mean character count	68.9
Median character count	67
Character range	8–140
Min. Hungarian char ratio	1.5%
Mean Hungarian char ratio	10.0%

Bucket distribution (used to ensure variety across sentence lengths):

Bucket	Range	Count
short	8–40 chars	60,000
medium	41–90 chars	80,000
long	91–140 chars	60,000

Hungarian character frequencies (in the 200k sentence corpus):

Char	Count
á	433,936
é	376,628
ó	116,001
ö	126,808
ő	103,651
í	68,898
ü	59,812
ú	34,677
ű	26,662

---

Generation Pipeline

The full pipeline consists of 3 stages:

01_corpus_preparation.ipynb
  Hungarian Wikipedia (HuggingFace dataset)
  → sentence tokenization (NLTK punkt)
  → quality filtering (length, HU char ratio, URL/markup removal, deduplication)
  → corpus_clean.parquet (200k sentences, 3 length buckets)

02_image_generator.ipynb
  corpus_clean.parquet
  → PIL text rendering (fixed canvas height, auto-fit font size)
  → background generation (profile-specific color + grain)
  → Albumentations augmentation (profile-specific pipeline)
  → synthetic_ocr_dataset/ (100k JPEG images + metadata.parquet)

03_dataset_builder.ipynb
  synthetic_ocr_dataset/
  → sharded Parquet writing (2,000 images/shard, memory-efficient)
  → HF Image feature encoding
  → instruction-tuning conversation format
  → train/val split (95/5)
  → hf_ocr_dataset/dataset/

---

Fonts Used

All fonts are open-source and support the full Hungarian character set (including ő and ű which are absent from many common fonts):

Font family	Style variants	Profile
DejaVu Sans	Regular, Bold, Oblique, BoldOblique, ExtraLight, Condensed, Mono	modern
Liberation Sans	Regular, Bold, Italic, BoldItalic, Mono	modern
FreeSans	Regular, Bold, Oblique, Mono	modern
DejaVu Serif	Regular, Bold, Italic, BoldItalic, Condensed	scanned, archiv, paragraph
Liberation Serif	Regular, Bold, Italic, BoldItalic	scanned, archiv, paragraph
FreeSerif	Regular, Bold, Italic, BoldItalic	scanned, archiv, paragraph
FreeMono	Regular, Oblique	archiv

---

Intended Use

This dataset is designed for fine-tuning vision-language models on Hungarian OCR, specifically:

• Improving recognition accuracy of Hungarian accented characters (ő, ű, í, á, é, ó, ö, ú, ü)
• Adapting pre-trained OCR models (e.g. DeepSeek-OCR via Unsloth) to Hungarian typography
• Training or evaluating document understanding models on Hungarian text

Out-of-scope uses

• This dataset does not contain real scanned documents; it is fully synthetic
• It is not suitable for handwriting recognition (all text is computer-rendered)
• It should not be used as a source of factual information

---

Limitations

• Synthetic only: real scanned documents contain distortions (page curl, ink bleed, uneven lighting) not fully captured by the augmentation pipeline
• Font diversity: only open-source fonts with Hungarian support were used; proprietary typefaces (Times New Roman, Arial, etc.) are not represented
• No handwriting: all text is rendered with digital fonts
• Wikipedia domain: ground truth text is encyclopedic; performance may vary on other domains (legal, literary, colloquial)
• Single column: paragraph images simulate single-column layout only; multi-column newspaper layouts are not included

---

License

This dataset is released under Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0), consistent with the license of the source material (Hungarian Wikipedia).

You are free to share and adapt this dataset, provided you give appropriate credit and distribute any derivative works under the same license.

---

Citation

@dataset{GaborMadarasz/synthetic_OCR_dataset_v2,
  author    = {Madarász, Gábor},
  title     = {Hungarian Synthetic OCR Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/GaborMadarasz/synthetic_OCR_dataset_v2},
  license   = {CC BY-SA 4.0},
  note      = {Synthetically generated from Hungarian Wikipedia text
               using PIL and Albumentations. Designed for fine-tuning
               vision-language OCR models on Hungarian accented characters.}
}

---

Acknowledgements

• Ground truth text: Hungarian Wikipedia (CC BY-SA 4.0)
• Image rendering: Pillow (PIL)
• Augmentation: Albumentations
• Fine-tuning framework: Unsloth
• Fonts: DejaVu Fonts, Liberation Fonts, GNU FreeFont