DastKhat

Persian Handwritten Text Dataset

DastKhat (دستخط) is an open Persian handwritten text dataset designed for Optical Character Recognition (OCR) and Handwritten Text Recognition (HTR). The dataset contains sentence-level handwritten samples collected from multiple writers and aims to support research on Persian handwriting recognition, writer-independent OCR systems, and document understanding.

The dataset contains handwritten Persian sentences collected from multiple participants. Each participant wrote a collection of predefined Persian sentences, producing handwritten samples with different writing styles and characteristics.

---

Dataset Statistics

Metric	Value
Writers	8
Unique Sentences Available	700
Handwritten Samples	560
Image Type	RGB
Language	Persian
Annotation Level	Sentence
License	mit
Current Version	v0.1

---

Sample Handwriting

---

Why DastKhat?

Despite the growing availability of OCR datasets for English and several other languages, publicly available Persian handwritten datasets remain limited in size, accessibility, and writer diversity.

DastKhat was created to provide an open resource for developing and evaluating Persian OCR and HTR systems while encouraging community participation from students, researchers, and handwriting contributors.

---

Dataset Contents

The dataset consists of handwritten Persian sentence images and a metadata file describing the corresponding annotations.

The metadata includes the following fields:

Field	Description
sample_id	Unique identifier for the handwritten sample
participant_id	Identifier of the participant who wrote the sample
sentence_id	Identifier of the source sentence
text	Ground-truth Persian transcription
split	Dataset split (train, validation, or test)

Each metadata record links a handwritten image to its corresponding transcription and additional information. Example:

sample_id,participant_id,sentence_id,text,split
000001,P001,000001,امروز هوا بسیار خوب است.,train

---

Dataset Structure

The dataset is organized around handwritten sentence images.

Each image is associated with:

• A unique sample identifier
• The participant who wrote the sample
• The sentence identifier
• The original Persian text
• A dataset split

The dataset is designed to preserve writer diversity, allowing models and experiments to work with handwriting from different individuals.

---

Data Collection Pipeline

1. Participants receive a generated handwriting sheet.
2. Each sheet contains predefined Persian sentences.
3. Participants write the sentences by hand.
4. Sheets are scanned.
5. Handwritten regions are automatically detected.
6. Individual sentence images are cropped.
7. Metadata is generated and verified.
8. Samples are added to the dataset.

---

Intended Uses

DastKhat can be used for:

• Persian Handwritten Text Recognition (HTR)
• Persian Optical Character Recognition (OCR)
• Computer Vision
• Deep Learning
• Document Image Analysis
• Writer Identification
• Handwriting Style Analysis
• Data Augmentation Research
• Sequence Recognition

The dataset is particularly suitable for experiments involving image-to-text recognition of handwritten Persian sentences.

---

Recommended Evaluation Metrics

For OCR and HTR experiments we recommend:

• Character Error Rate (CER)
• Word Error Rate (WER)
• Sequence Accuracy

---

Loading the Dataset

You can load the dataset using the Hugging Face datasets library:

python -m pip install datasets

from datasets import load_dataset

dataset = load_dataset("MrFarahmand/DastKhat")

Inspect the dataset:

print(dataset)

Access a sample:

sample = dataset["train"][0]

print(sample)

the image can be accessed from the image field:

image = sample["image"]
text = sample["text"]

print(text)
image.show()

---

Dataset Splits

The dataset includes split information associated with the samples.

The available splits may include:

• train
• validation
• test

The exact number of samples in each split may change as the dataset is expanded.

For the most up-to-date dataset statistics, please refer to the dataset viewer and dataset metadata on this page.

---

Limitations

DastKhat is an actively developing dataset.

Potential limitations include:

• The number of participants may be limited compared to large-scale handwriting datasets.
• The dataset currently focuses on handwritten Persian sentences.
• Handwriting styles may not represent the full diversity of Persian handwriting.
• Dataset size and composition may change as new samples are collected.

Users should consider these limitations when evaluating model performance.

---

Contributing

We welcome new handwriting contributors.

You can participate through:

https://www.alacrity.ir

Contributors may be acknowledged in future releases and project publications.

GitHub Repository

The GitHub repository contains the project's documentation, metadata, and data processing tools:

https://github.com/AmirRezaFarahmand/DastKhat

---

Citation

If you use DastKhat in your research or project, please cite this dataset.

@dataset{farahmandfar2026dastkhat,
  author = {Farahmandfar, AmirReza},
  title = {DastKhat: A Persian Handwritten Text Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/MrFarahmand/DastKhat}
}

---

Version History

v0.1

• Initial public release
• 8 writers
• 560 samples
• 700 sentences

Planned v0.2

• Additional writers
• Expanded sentence set
• Improved metadata

---

Acknowledgements

Special thanks to all participants who contributed their handwritten Persian samples.

Their contributions made the creation of the DastKhat dataset possible.

---

Maintainer

AmirReza Farahmand

• GitHub: https://github.com/AmirRezaFarahmand
• Hugging Face: https://huggingface.co/MrFarahmand