TajikOCR is a synthetic dataset of page images containing texts in Tajik (Cyrillic script). The pages imitate old book leaves with classical poetry (Ferdowsi, Saadi, Rumi, Nizami, Bedil, Asadi, and others). Each page features various backgrounds (from plain white to aged paper with vignetting), borders, ornaments, and realistic defects (noise, blur, ink stains, folds, fiber texture). The dataset is designed for training robust OCR models and multimodal systems.

Curated by: TajikNLPWorld
Language(s) (NLP): Tajik (Cyrillic script)
License: Apache 2.0

Dataset Sources

Repository: TajikOCR on Hugging Face
Original corpus: TajPersParallelCorpusFull

Uses

Direct Use

The dataset can be used for:

Optical Character Recognition (OCR) for Tajik Cyrillic text
Visual Question Answering (VQA) on document images
Image Captioning
Multimodal Retrieval (text-to-image and image-to-text)

Out-of-Scope Use

The dataset is synthetic and may not fully represent real-world historical documents. It is not intended for:

Production systems without additional validation on real data
Tasks requiring exact historical authenticity or original page layouts
Applications involving personal or sensitive information

Dataset Structure

The dataset is organised in the following directory structure:

data/
├── Asadi/
│   ├── page_023456.jpg
│   └── ...
├── Bidel/
│   ├── page_025200.jpg
│   └── ...
├── Ferdousi/
│   ├── part_0/
│   │   ├── page_008670.jpg
│   │   └── ...
│   └── ...
├── Nezami/
│   ├── page_020766.jpg
│   └── ...
├── Saadi/
│   ├── page_014183.jpg
│   └── ...
└── ... (other categories)

Image files: JPG images, each 1200×800 pixels.
Annotation file: annotations.parquet (in the root of the repository) with one JSON object per line.

Each annotation contains:

id: unique page identifier (integer)
image_path: relative path to the image file (e.g., data/Asadi/page_023456.jpg)
text: full text on the page, lines separated by newline characters
bboxes: list of line objects, each containing:
- text: the line text
- bbox: [x_min, y_min, x_max, y_max] coordinates
- words: list of word objects with word and bbox
author: author of the text (if available)
category: source category (e.g., masnavi, shahnameh, Saadi)

Примеры страниц (Example Pages)

Dataset Creation

Curation Rationale

There is a lack of publicly available OCR datasets for the Tajik language, especially with realistic document degradation. This dataset was created to fill that gap by generating synthetic pages that mimic aged manuscripts and books, using existing high-quality Tajik text from a parallel corpus. The synthetic nature allows for a large volume of varied and annotated data for training deep learning models.

Source Data

The source text data comes from TajPersParallelCorpusFull, a Tajik–Persian parallel corpus with 456,307 unique sentence pairs. From this corpus, records belonging to the following categories were selected: Saadi, masnavi, shahnameh, Ferdousi, Nezami, Bidel, Asadi.

Data Collection and Processing

The generation pipeline (implemented in Python using Pillow and NumPy) includes:

Loading and filtering the corpus to keep only Tajik texts from the specified categories.
Splitting each text into short verse fragments (1–4 lines).
Creating a random page background – white, cream, gray, gradient, aged (with vignette), or textured paper.
Adding decorative elements – borders (simple, double, decorative, ornamental), corner ornaments, header/footer decorations, page numbers.
Rendering the text using the Noto Serif font (supporting the Tajik alphabet) with random alignment (justified, left, center).
Applying augmentations – rotation, affine shift, blur, noise, brightness/contrast adjustment, paper texture, ink stains, fold lines, and fibers.
Saving the page image and recording the exact text and word/line bounding boxes in JSONL format.

All processing steps are deterministic given a random seed; the released dataset was generated in a single run. The complete generation code is available upon request from the authors.

Who are the source data producers?

The original text authors are classical Tajik/Persian poets (e.g., Ferdowsi, Saadi, Rumi, Nizami, Bedil).
The parallel corpus was compiled and curated by the TajikNLPWorld community.
The synthetic page images were generated programmatically by TajikNLPWorld.

Annotations

Annotation process

Annotations (text and bounding boxes) were generated automatically during image creation. The exact line and word coordinates are recorded as the text is drawn on the page. No manual annotation or correction was performed.

Who are the annotators?

No human annotators were involved. All annotations are produced by the generation script.

Personal and Sensitive Information

The dataset contains classical poetry and does not include any personal, sensitive, or private information.

Bias, Risks, and Limitations

Domain bias: The dataset is limited to classical poetry and may not generalize well to other text types (e.g., newspapers, handwritten notes).
Synthetic artifacts: Although the images include realistic degradations, they may not capture the full variability of real-world documents (e.g., lighting conditions, scanner noise, arbitrary rotations).
Language coverage: Only Tajik (Cyrillic) is represented; other languages or scripts are not included.
Font dependency: The text is rendered using a single font family (Noto Serif). Models trained solely on this dataset might overfit to this font.
Size: While the dataset contains over 26,000 images, it may still be insufficient for very large models without additional data.

Recommendations

Users should:

Augment this dataset with real scanned documents if available.
Evaluate models on real-world Tajik OCR benchmarks before deployment.
Consider fine-tuning on additional fonts and degradations to improve robustness.
Be aware of the domain limitation (poetry) when applying the model to other text genres.

Citation

BibTeX for TajikOCR:

@dataset{tajikocr_2026,
    title        = {TajikOCR: Synthetic Pages of Tajik Poetry for OCR and Multimodal Tasks},
    author       = {TajikNLPWorld},
    year         = {2026},
    publisher    = {Hugging Face},
    url          = {https://huggingface.co/datasets/TajikNLPWorld/tajik-ocr}
}

BibTeX for the original corpus:

@dataset{tajpers_parallel_corpus_full_2026,
    title        = {TajPersParallelCorpusFull – A Unified Tajik–Persian Parallel Corpus},
    author       = {TajikNLPWorld},
    year         = {2026},
    publisher    = {Hugging Face},
    url          = {https://huggingface.co/datasets/TajikNLPWorld/TajPersParallelCorpusFull}
}

APA:

TajikNLPWorld. (2026). TajikOCR: Synthetic Pages of Tajik Poetry for OCR and Multimodal Tasks [Dataset]. Hugging Face. https://huggingface.co/datasets/TajikNLPWorld/tajik-ocr

Glossary

Beit (бейт): A couplet in Persian/Tajik poetry; in this dataset, texts were split into fragments of 1–4 lines.
Masnavi: A poetic form consisting of rhyming couplets, often used for narrative poetry.

More Information

The generation code is available upon request from the dataset authors. For any questions, open an issue on the Hugging Face repository.

Dataset Card Authors

TajikNLPWorld

Dataset Card Contact

Open an issue on the Hugging Face repository: TajikOCR or contact the TajikNLPWorld team.

Downloads last month: 52

Total file size:

11 GB