DatarrX/Burmese-Handwritten-Sentence-Dataset

Burmese Handwritten Sentence Dataset (BHSD)

BHSD is a sentence-level Burmese handwriting dataset developed for optical character recognition (OCR), handwritten text recognition (HTR), error analysis, robustness testing, and research on low-resource scripts.

The dataset was created by Ah Maung Oo and DatarrX through the voluntary contributions of 54 handwriting writers.

This dataset would not have been possible without its volunteers. Every handwritten image in BSHD exists because a volunteer gave their time, effort, and trust to support open Burmese-language research. Their contributions form the foundation of the entire dataset.

Dataset Summary

Collection	Writers	Sentences per writer	Images
Core collection	54	50	2,700
Supplementary coverage collection	10	15	150
Complete dataset	54 unique writers	—	2,850

All ten supplementary writers also participated in the core collection. The complete dataset therefore contains 54 unique writers.

Core Sentence Categories

Every core writer contributed the same 50 handwritten items:

Category	Labels	Count	Description
Natural (NAT)	NAT_01–NAT_30	30	Natural Burmese sentences for normal sentence-level handwriting and OCR testing.
Rare (RARE)	RARE_01–RARE_10	10	Sentences containing less-common names, places, numbers, codes, technical vocabulary, or mixed-script content.
Nonce (NONCE)	NONCE_01–NONCE_10	10	Constructed Burmese-like words used to test whether a recognition system reads the handwriting instead of relying only on familiar-word guessing.

Supplementary Coverage Collection

The supplementary collection contains 15 additional natural Burmese sentences labelled SUP_01–SUP_15. It was created after analysing the core prompts for missing or underrepresented characters and grapheme combinations.

The supplementary collection is a coverage extension and should be kept separate from the original 50-sentence core benchmark.

Data Format

Each image represents one cropped handwritten sentence or nonce item. Images are stored inside their respective writer folders. The filename contains only the sentence label; it does not contain the writer ID.

Example filenames:

NAT_01.png
RARE_03.png
NONCE_10.png

The downloadable sentence-label spreadsheet provides the exact Burmese transcription assigned to each NAT, RARE, NONCE, and SUP label.

Supplementary coverage images are named numerically from 1.png to 15.png within each writer’s folder.

Recommended Metadata Fields

Machine-readable metadata should contain one row per image with at least the following fields:

Field	Description
file_name	Relative path to the image file.
text_id	Sentence label such as NAT_01 or SUP_10.
writer_id	Pseudonymous writer identifier.
category	NAT, RARE, NONCE, or SUP.
ground_truth	Exact Unicode Burmese transcription of the image.
collection	core or coverage.

Dataset Splits

The initial public release does not define an official training, validation, and test split. Researchers creating their own splits should split the data by writer rather than by individual image. Images from one writer should not appear across multiple splits, as this can cause writer leakage and produce overly optimistic results.

Collection Process

Fifty-four volunteers were asked to write the provided prompts in their normal handwriting. Each volunteer in the core collection wrote the same 50 items. Ten of these volunteers also contributed 15 supplementary natural sentences designed to improve coverage of characters and grapheme combinations that were underrepresented in the core prompts.

The volunteers were not simply participants in the collection process; they made the dataset possible. Their handwriting is the central research contribution represented by BSHD, and their time and effort should be acknowledged whenever the dataset is used or discussed.

The released data consists of cropped sentence images. Direct personal identifiers should not be included in public filenames or metadata.

The prompt set contains researcher-created text and sentences drawn from or adapted from Burmese-language text resources, including the kalixlouiis/burmese-text-corpus and chuuhtetnaing/myanmar-language-dataset-collection resources. Additional prompts were created or revised by the BSHD research team.

Intended Uses

BSHD is intended for:

• Burmese OCR and handwritten text recognition;
• sentence-level transcription experiments;
• character and grapheme coverage analysis;
• robustness and preprocessing evaluation;
• error analysis using character error rate and related metrics;
• educational and non-commercial or commercial research permitted by the license.

Out-of-Scope Uses

The dataset should not be used to:

• identify or authenticate individual writers;
• infer sensitive personal characteristics from handwriting;
• reproduce signatures or identity documents;
• build handwriting-based surveillance systems;
• misrepresent generated handwriting as belonging to a contributor.

Limitations

• The core collection uses a fixed set of 50 prompts for all core writers.
• The dataset may not represent every Burmese handwriting style, demographic group, region, writing instrument, or capture condition.
• The supplementary collection has fewer participating writers than the core collection.
• Some graphemes or combinations may remain underrepresented even after the coverage extension.
• Cropping, transcription, blur, lighting, and other collection errors may remain despite manual review.
• Results obtained from randomly splitting images may be misleading because the same writer can appear in both training and evaluation data.

Privacy and Responsible Release

The public release should contain only images approved for public dataset distribution. Original full-page photographs, private identity mappings, names, email addresses, phone numbers, signatures, student identifiers, location information, and embedded EXIF metadata should not be included.

Users of the dataset are responsible for complying with the license and for respecting the privacy and dignity of handwriting contributors.

License

BSHD is released under the Creative Commons Attribution 4.0 International License. Users may share and adapt the dataset, including for commercial purposes, provided that appropriate credit is given and changes are indicated.

Citation

If you use BHSD, please cite the dataset:

@dataset{oo_2026_bhsd,
  author    = {Oo, Ah Maung},
  title     = {Burmese Handwritten Sentence Dataset (BSHD)},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Drew2456/Burmese-Sentence-Handwriting-Dataset}
}

Maintainer

Ah Maung Oo
DatarrX — datarrx.org

Acknowledgements

The deepest thanks go to all 54 volunteers who contributed their handwriting to BSHD. This dataset could not exist without their generosity, patience, time, and willingness to support Burmese-language technology. Every core image and every supplementary coverage image comes directly from their contribution.

Their collective effort has created an open resource that can support future work in Burmese OCR, handwritten text recognition, and low-resource language technology. The creators sincerely ask users of BSHD to recognise and respect the volunteers whose work made this release possible.

Special thanks to kalix.louiss.

Thank you to every volunteer, from Andrew and DatarrX.