politecat314/dhivehi-transliteration

Dhivehi Transliteration Dataset

A comprehensive dataset of Dhivehi-to-Latin script transliteration (300,000 pairs), compiled from news article titles scraped from leading Maldivian news websites.

Dataset Description

This dataset contains parallel Dhivehi and Latin script text pairs designed for training bidirectional transliteration models. The data consists of news article titles collected from prominent Maldivian news sources, providing real-world, contemporary Dhivehi text with corresponding Latin transliterations.

Data Sources

The dataset is compiled from news articles scraped from the following Maldivian news websites:

• Mihaaru News (mihaaru.com)
• Adhadhu (adhadhu.com)
• Dhauru (dhauru.com)
• Dhiyares (dhiyares.com)
• Sun.mv (sun.mv)

Dataset Structure

Data Splits

• Train: Training split for model development (290865 pairs)
• Test: Test split for evaluation (5936 pairs)

Data Fields

• dhivehi: News article titles in Dhivehi (Thaana script)
• latin: Corresponding Latin transliteration of the Dhivehi text

Usage

Basic Loading

from datasets import load_dataset

dataset = load_dataset("politecat314/dhivehi-transliteration")

Loading the Dataset to train a bidirectional model

from datasets import load_dataset, DatasetDict, concatenate_datasets

def create_bidirectional_dataset(repo_id: str, dv_col: str, la_col: str) -> DatasetDict:
    """
    Loads translation data from the Hugging Face Hub, duplicates it for bidirectional
    training (dv->la and la->dv), and returns a prepared Hugging Face DatasetDict.
    
    Args:
        repo_id (str): Hugging Face dataset repository ID.
        dv_col (str): Column name for the Dhivehi text.
        la_col (str): Column name for the Latin text.
        
    Returns:
        DatasetDict: A DatasetDict containing 'train' and 'test' splits,
                    prepared for bidirectional training.
    """
    # Load dataset from the Hugging Face Hub
    raw_datasets = load_dataset(repo_id)
    bidirectional_datasets = {}
    
    for split in raw_datasets.keys():
        dataset = raw_datasets[split]
        
        # 1. Dhivehi -> Latin
        ds1 = dataset.map(lambda example: {
            'source': example[dv_col],
            'target': example[la_col],
            'prefix': PREFIX_DV2LA
        })
        
        # 2. Latin -> Dhivehi
        ds2 = dataset.map(lambda example: {
            'source': example[la_col],
            'target': example[dv_col],
            'prefix': PREFIX_LA2DV
        })
        
        # 3. Concatenate and shuffle
        bidirectional_datasets[split] = concatenate_datasets([ds1, ds2]).shuffle(seed=42)
    
    return DatasetDict(bidirectional_datasets)

# Load the dataset
DHIVEHI_COL = "dhivehi"
LATIN_COL = "latin"
PREFIX_DV2LA = "dv2la: "
PREFIX_LA2DV = "la2dv: "

dataset = create_bidirectional_dataset("politecat314/dhivehi-transliteration", DHIVEHI_COL, LATIN_COL)
print("Bidirectional dataset prepared and split:", dataset)

# Example usage
print("\nExample from the training set:")
print(dataset['train'][0])

Limitations

This dataset contains Dhivehi and Latin transliteration pairs from news article titles. Care should be taken when models trained on this dataset are used in other domains.

Citation

If you use this dataset in your research, please cite:

@dataset{dhivehi_transliteration_2025,
  title={Dhivehi Transliteration Dataset},
  author={Ahmed Aman Ibrahim, Maadh Ali Zameer, Mushfau Saeed},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/politecat314/dhivehi-transliteration}
}

Acknowledgments

Special thanks to Maadh (https://github.com/itsMaadh) and Mushfau (https://github.com/mushfau) for scraping the data and the news websites Mihaaru, Adhadhu, Dhauru, Dhiyares, and Sun.mv for providing high-quality transliterations that form the foundation of this dataset.

Contact

For questions, issues, or contributions, please reach out at ahmedamanibrahim@gmail.com or create an issue in the repository.