DatarrX/myanmar-numeral-glyphs

🇲🇲 Myanmar Numeral Glyphs (MNG)

The Myanmar Numeral Glyphs (MNG) dataset is a curated, high-quality hybrid image dataset designed for optical character recognition (OCR) and image classification tasks targeting native Burmese digits (၀ to ၉). Released under DatarrX, this dataset bridges the gap in low-resource language resources by combining clean, human-annotated handwritten data with robust computer-generated font variations.

---

📌 Dataset Overview

• Total Images: 1,488 images
• Target Classes: 10 classes (၀, ၁, ၂, ၃, ၄, ၅, ၆, ၇, ၈, ၉)
• Format: MNIST-standardized format (28x28 pixels, grayscale, black background with white foreground strokes).
• Subsets: - Handwritten: Authentic human handwriting samples gathered under a structured pipeline.
  • Computer-generated: Synthetic data rendered using multiple native Burmese typographic fonts (including MyanmarNayone, YoeYarOne, MyanmarSquare & Pixel) with variations in weight, slant (italicization), and augmentations.

---

📊 Dataset Distribution & Insights

The dataset is strictly audited to ensure high data quality and balancing across all ten digits. Below is the detailed breakdown of the image distribution:

Label (ENG)	Label (MM)	Computer-generated	Handwritten	Total (Digit Breakdown)
0	၀	16	114	130
1	၁	16	120	136
2	၂	16	120	136
3	၃	16	120	136
4	၄	16	120	136
5	၅	16	114	130
6	၆	16	120	136
7	၇	16	120	136
8	၈	16	120	136
9	၉	16	120	136
Total	Summary	160 (11.87%)	1188 (88.13%)	1348 (100%)

🎯 Data Balance Check:

• Maximum Samples: Digit 1, 2, 3, 4, 6, 7, 8, 9 (150 images each)
• Minimum Samples: Digit 0, 5 (144 images each)
• Note: The minor variance is the result of strict data cleaning and human-in-the-loop noise filtration to ensure no distorted or corrupted strokes are present.

---

⚙️ Data Preparation & Augmentation

To make deep learning models highly robust against real-world shifts, both handwriting and font subsets underwent precise processing pipelines:

1. True MNIST Look: Canvas padding, object centering, and bounding-box margin isolation (20% margins).
2. Augmentations Applied: - Random rotations (±5° to ±10°)
   • Translations/Shifts (±2 pixels)
   • Gaussian blurring (simulating ink bleeding or camera blur)
   • Morphological Dilation/Erosion (simulating variation in stroke thickness)

---

👥 Roles & Credits

This dataset is published as part of an open-source initiative to support localized AI solutions for Myanmar.

• Publisher / Organization: DatarrX (Myanmar Open Source, Non-Profit Organization)
• Lead Developer: Khant Sint Heinn (Engineered the pipeline, image cropping scripts, interactive cleaning dashboards, and Parquet pipeline)
• Data Annotators (Handwritten Contributors): Khant Sint Heinn & Thazin Nyein (Contributed an exact 50% split each to the handwritten handwriting acquisition phase)

---

🛠️ How to Use

Since the dataset is pre-converted into the optimized .parquet standard format, you can load it directly into your deep learning project via the Hugging Face datasets library with a single line of code:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("DatarrX/myanmar-numeral-glyphs")

# Example: Accessing a sample
sample = dataset['train'][0]
print(sample['label'])       # Output: 0 (Integer for ML training)
print(sample['label_mm'])    # Output: '၀' (Native Burmese character)
print(sample['source_type']) # Output: 'Handwritten'

---

📜 Citation Info

If you use this dataset in your research, production systems, or academic publications, please cite it using the following BibTeX format:

@misc{datarrx_mng_2026,
  author       = {Khant Sint Heinn and Thazin Nyein},
  title        = {Myanmar Numeral Glyphs (MNG): A Hybrid Dataset for Native Burmese Digits Recognition},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {https://huggingface.co/datasets/DatarrX/myanmar-numeral-glyphs},
  note         = {Published under DatarrX Initiative}
}