---
license: odc-by
license_link: https://opendatacommons.org/licenses/by/
language:
- my
- en
tags:
- c4
- web_crawl
size_categories:
- 100K<n<1M
extra_gated_prompt: "Please let me send you my newsletters @@, or access the data from Kaggle."
extra_gated_fields:

  I will let you send newsletter to my email: checkbox

---

<div align="center">
  <img src="https://cdn.alexbeatson.me/blog/static/media/NerdyNerd.svg" alt="Nerdy Nerd" width="200"/>
</div>

<div align="center" style="line-height: 1;">
  <a href="https://github.com/alexanderbeatson" target="_blank" style="margin: 2px;">
    <img alt="GitHub" src="https://img.shields.io/badge/GitHub-Alexander%20Beatson-blue?logo=github" style="display: inline-block; vertical-align: middle;"/>
  </a>
  <a href="https://opendatacommons.org/licenses/by/" target="_blank" style="margin: 2px;">
    <img alt="ODC-BY" src="https://img.shields.io/badge/License-ODC--BY-green" style="display: inline-block; vertical-align: middle;"/>
  </a>
</div>

# Burmese C4 dataset (Cleaned)

## Overview
This dataset is a cleaned version of the Burmese subset of the [C4 multilingual dataset](https://huggingface.co/datasets/allenai/c4/tree/main/multilingual). I clean existing script patterns and noises while retaining high-quality Myanmar (\U1000-\U109F) and English text. If I ever publicize something, it must be under the open-source license [ODC-BY](https://opendatacommons.org/licenses/by/), as always.

## Why?
All of Myanmar scripts are very low-resource in this AI era. High-quality data is always needed for training robust NLP models. Garbage in - garbage out. This cleaned version of Burmese C4 is just my another _tiny_ contribution towards the development of Myanmar LLM models, enforcing character coverage, and handling legacy encoding issues like Zawgyi.

## Cleaning Process
I tried the following steps to refine the data:
- **Remove original URLs:** Too bored to retain them, please reach out to me in case you need the original URLs.
- **Script & HTML Cleaning:** Removed any residual JavaScript, HTML tags, and other non-text elements.
- **Character Coverage Filtering:** Retained only documents where English and Myanmar Unicode characters constitute at least 95% of the text.
- **Whitespace Normalization:** Standardized spaces, tabs, and line breaks.
- **Zawgyi Detection & Conversion:** Converted auto-convertible Zawgyi text to Unicode and discarded non-convertible documents.
- **Tokenizer Enforcement:** Any content outside Burmese and English language constraints is replaced with `<unk>`.

## Usage
This dataset is designed for:
- Pretraining and fine-tuning multilingual NLP models
- Burmese-English language modeling
- Low-resource NLP research

## Dataset Format
- **Language Coverage:** Burmese (Unicode), English (Latin)
- **File Type:** JSON

## Citation
Cite me plz?
```
@dataset {alexanderbeatson,
	author       = { {Alexander Beatson} },
	title        = { Burmese C4 dataset (Cleaned) },
	year         = 2025,
	url          = { https://huggingface.co/datasets/alexbeatson/c4_mm },
	doi          = { 10.57967/hf/4865 },
	publisher    = { Hugging Face },
	note     = {ORCID: 0000-0002-1829-5965} 
}
```

## License
This dataset is released under the [ODC-BY License](https://opendatacommons.org/licenses/by/), same as the original license. You are free to share and adapt the dataset with proper attribution.