yasalma/tatar-ocr-benchmark

Tatar OCR Benchmark

This dataset is a page-level OCR benchmark for Tatar documents. Each item contains:

• the original page image,
• structured OCR/layout annotations in JSONL,
• a visual control image for fast human verification (original | reconstructed).

Full original page images are included.

Acknowledgements

We express our appreciation to Yandex LLC for support in dataset collection.

What Is Included

Export folder structure:

• images/... original page images
• control/... control images (left: original, right: reconstructed)
• dataset.jsonl one JSON object per page

JSONL Schema (dataset.jsonl)

Top-level fields per row:

• page_id (int) stable internal page identifier
• image (str) relative path to original image (under images/)
• control (str) relative path to control image (under control/)
• width (int) original image width in pixels
• height (int) original image height in pixels
• items (list[object]) ordered OCR/layout objects

items[] fields:

• order (int) reading order
• layout_id (int) stable internal bbox identifier
• class_name (str) layout class
• bbox (object) normalized coordinates:
  • x1, y1, x2, y2 (float, range [0,1])
• content_format (str, optional) output format (markdown, html, latex, etc.)
• content (str, optional) OCR content
• caption_targets (list[int], caption-only) referenced layout_id targets

Notes:

• picture items typically do not contain content/content_format.
• page_id and layout_id are kept intentionally for stable linking and reproducibility.

Example Row

{
  "page_id": 42,
  "image": "images/misc/001000.png",
  "control": "control/misc/001000.png",
  "width": 1408,
  "height": 1088,
  "items": [
    {
      "order": 1,
      "layout_id": 1835,
      "bbox": {
        "x1": 0.0557,
        "y1": 0.0958,
        "x2": 0.109,
        "y2": 0.167
      },
      "class_name": "picture"
    },
    {
      "order": 2,
      "layout_id": 1829,
      "bbox": {
        "x1": 0.1144,
        "y1": 0.0966,
        "x2": 0.3049,
        "y2": 0.1648
      },
      "class_name": "text",
      "content_format": "markdown",
      "content": "БӨТЕНДӨНЬЯ ТАТАР\nЯШЬЛӘРЕ ФОРУМЫ"
    },
    {
      "order": 3,
      "layout_id": 1830,
      "bbox": {
        "x1": 0.1026,
        "y1": 0.2243,
        "x2": 0.6247,
        "y2": 0.3098
      },
      "class_name": "section_header",
      "content_format": "markdown",
      "content": "#### СЕРТИФИКАТ"
    },
    {
      "order": 4,
      "layout_id": 1834,
      "bbox": {
        "x1": 0.1199,
        "y1": 0.3424,
        "x2": 0.2885,
        "y2": 0.3751
      },
      "class_name": "text",
      "content_format": "markdown",
      "content": "ТАПШЫРЫЛА"
    },
    {
      "order": 5,
      "layout_id": 1831,
      "bbox": {
        "x1": 0.1942721496928822,
        "y1": 0.6072256985832664,
        "x2": 0.6748469959605824,
        "y2": 0.687212719636805
      },
      "class_name": "text",
      "content_format": "markdown",
      "content": "*\"Татарча диктант\" белем бирү*\n*акциясендә катнашкан өчен*"
    },
    {
      "order": 6,
      "layout_id": 1832,
      "bbox": {
        "x1": 0.6836,
        "y1": 0.7307,
        "x2": 0.9285,
        "y2": 0.9422
      },
      "class_name": "text",
      "content_format": "markdown",
      "content": "**Татарча\nдиктант**\n\n**ЯЗ.**"
    },
    {
      "order": 7,
      "layout_id": 1833,
      "bbox": {
        "x1": 0.32613049853931775,
        "y1": 0.873131134930779,
        "x2": 0.6471898772499778,
        "y2": 0.9243139940149644
      },
      "class_name": "text",
      "content_format": "markdown",
      "content": "*Ихтирам белән, Бөтендөнья татар*\n*яшьләре форумы рәисе Т.М. Яруллин*"
    }
  ]
}

Control Image Semantics

control image is horizontal:

• left half: original image
• right half: reconstructed rendering from the same JSON row

Reconstruction uses the same exported items as source of truth:

• same bbox coordinates,
• same class,
• same order,
• same content.

Each rendered box includes:

• class label,
• reading order,

Key decisions applied:

• OCR preserves source text strictly; typos/errors are not corrected.
• Diacritics are preserved strictly; no simplification to look-alike characters.
• Header hierarchy is preserved; default header level may be adjusted by page context.
• Multi-line headers preserve line breaks (e.g., with <br> in markdown workflow rules).

Intended Use

• OCR benchmarking for Tatar historical/printed documents
• OCR post-processing and reconstruction evaluation

Dataset Evolution Notice

This dataset is an actively evolving benchmark and is subject to change.

Possible updates in future versions include:

• additional pages and sources,
• annotation corrections and quality improvements,
• updated normalization/formatting decisions.

Where possible, backward compatibility will be preserved, but consumers should treat the schema as versioned and monitor release notes.

Limitations

• Current size is small (50 items) and will be expanded.
• Source domain/style coverage is still limited at this stage.
• Some hard OCR cases (quality, uncertainty, complex regions) remain challenging.

Citation

If you use this dataset, please cite:

• yasalma/tatar-ocr-benchmark (Hugging Face)