DatarrX/myanmar-word-glyphs

🇲🇲 Myanmar Word Glyphs (MWG)

The Myanmar Word Glyphs (MWG) is a curated vocabulary-based synthetic image dataset containing 49,800 high-quality word/phrase glyph images (256x64 pixels, grayscale).

Developed and engineered by Khant Sint Heinn, this dataset is officially published and distributed under DatarrX (Myanmar Open Source Organization, NPO). While our sibling project—MSSG—explores the absolute mathematical grid of theoretical syllables, MWG is designed to map out authentic, practical, and meaningful vocabulary units utilized in everyday written Burmese.

---

📌 Dataset Architecture & Source Material

The cornerstone of the MWG dataset lies in its linguistic authenticity. The base vocabulary comprises 12,451 unique Myanmar words and phrases extracted and cleaned directly from the official Myanmar Wiktionary (my.wiktionary.org) dump.

By utilizing genuine words ranging from short common nouns to highly complex compound phrases (such as "တုံးတိုက်တိုက်ကမ်းတိုက်တိုက်"), this dataset acts as a direct bridge for real-world Optical Character Recognition (OCR), Line Text Recognition, and Text-to-Image verification tasks. It mirrors the exact distribution of character spacing, diacritic placement, and word lengths found in modern digital dictionaries.

---

⚙️ Data Generation & Pipeline Logic

The generation architecture processes the source text line-by-line to transform raw text into robust deep-learning training blocks:

1. Dynamic Scaling (256x64 px): The canvas size is specifically engineered to handle horizontal extension. The 256-pixel width comfortably accommodates long compound phrases, while the 64-pixel height guarantees that upper vowel marks and complex multi-layered subscript ligatures are perfectly bounded without baseline clipping.
2. Global Font Caching: To achieve peak compute efficiency, the pristine open-source Padauk TrueType Font is compiled once globally into memory, eliminating redundant I/O cycles during generation.
3. Robust Data Augmentation Split: To mimic real-world print media variations, scanning artifacts, or lens distortions, every single input word generates an exact 4-stage visual split:
   • original: The digitally crisp reference image.
   • aug_1 (Rotational Shift): Controlled micro-rotations (±5°) optimized specifically for long text lines to maintain structural layout.
   • aug_2 (Affine Translation): Spatial pixel-shifting across both axes to train bounding-box resilience.
   • aug_3 (Gaussian Blur): Micro-blur filters to simulate ink bleeding, lens defocus, or low-resolution scanning conditions.

---

📊 Dataset Specifications

• Total Row Count: 49,800 instances
• Total Footprint: ~42.9 MB (Compressed Local Parquet Stream)
• Unique Input Words: 12,451 vocabulary entries (yielding 4 images per word via the 4 augmentation states)
• Color Space: 8-bit Grayscale (L mode), high-contrast white foreground strokes on a pure black background.

---

👥 Organization & Credits

This repository is built as a community-driven initiative to deliver free, high-performance computational blocks for low-resource language technologies.

• Dataset Creator & Lead Architect: Khant Sint Heinn * Designed the vocabulary parsing pipeline, custom text-centering bounding matrix, and optimized memory-safe automated local Parquet packaging.
• Publisher / Supporting Organization: DatarrX * A dedicated Myanmar Open Source Non-Profit Organization striving for localized technological independence and accessible public AI assets.

---

📜 Citation & Academic Reference

If you incorporate the MWG dataset into your deep learning models, academic publications, industrial OCR pipelines, or linguistic evaluation frameworks, please cite the project using the following official BibTeX reference:

@misc{datarrx_mwg_2026,
  author       = {Khant Sint Heinn},
  title        = {Myanmar Word Glyphs (MWG): A Wiktionary-Driven Vocabulary Dataset for Word-Level OCR and Line Text Recognition},
  year         = {2026},
  publisher    = {Hugging Face},
  organization = {DatarrX Initiative},
  howpublished = {https://huggingface.co/datasets/DatarrX/myanmar-word-glyphs},
  license      = {CC-BY-4.0}
  note         = {Published under DatarrX Initiative. Open-source community asset released under {Creative Commons Attribution 4.0 International}}
}