d3b4g/dhivehi-corpus

ދިވެހި Corpus — Dhivehi Text Corpus

Clean text corpus for the Dhivehi (Maldivian) language. Built for NLP research and language model training.

Dataset Summary

Split	Docs	Tokens
Train	430,695	~81.6M
Validation	23,924	~4.5M
Test	23,924	~4.6M
Total	478,543	~90.6M

• Language: Dhivehi (dv) — written in Thaana script (Unicode U+0780–U+07BF)
• License: CC-BY-4.0
• Avg quality score: 0.958 / 1.0
• Duplicates: 0 (MinHash LSH deduplication at 80% threshold)

Sources

Source	Type	Docs
sun.mv	News	152,815
vaguthu.mv	News	134,898
mihaaru.com	News	103,012
avas.mv	News	50,828
adhadhu.com	News	33,567
dv.wikipedia.org	Reference	3,423

Domain Distribution (train)

Domain	Docs	%
news	373,991	86%
sports	21,965	5%
business	11,980	2%
entertainment	9,306	2%
lifestyle	6,966	1%
reference	3,081	0.7%
government	2,232	0.5%
religious	720	0.2%
literary	454	0.1%

Fields

Field	Type	Description
body	string	Main article text in Thaana script
title	string	Article headline
source	string	Source site (mihaaru, sun, vaguthu, avas, adhadhu, wikipedia)
domain	string	Content domain (news, sports, business, etc.)
date	string	Publication date (YYYY-MM-DD where available)
author	string	Author name where available
quality_score	float	Composite quality score 0–1
token_count	int	Word count
thaana_ratio	float	Fraction of characters in Thaana Unicode range
url	string	Source URL
doc_id	string	MD5 hash of URL (unique identifier)

Quality Score

Each document is scored 0–1 based on:

• Thaana ratio (35%) — fraction of Thaana script characters
• Length score (30%) — log-normal score peaking at ~200 tokens
• Sentence score (20%) — presence of multiple complete sentences
• Encoding score (15%) — absence of replacement/corrupted characters

Minimum quality score in corpus: 0.663. Average: 0.958.

Usage

from datasets import load_dataset

# Load full corpus
dataset = load_dataset("d3b4g/dhivehi-corpus")

# Load only training split
train = load_dataset("d3b4g/dhivehi-corpus", split="train")

# Filter by domain
news_only = train.filter(lambda x: x["domain"] == "news")

# Filter by quality
high_quality = train.filter(lambda x: x["quality_score"] >= 0.95)

# Get just the text for language model training
texts = train.select_columns(["body"])

Use Cases

• LLM pretraining / fine-tuning — Fine-tune multilingual models (mBERT, XLM-R, mT5) on Dhivehi
• Tokenizer training — Train a Dhivehi-specific BPE/sentencepiece tokenizer
• Text classification — Domain labels ready for news category classification
• NER bootstrapping — Rich source of Maldivian person names, island names, organizations
• Sentiment analysis — Political and social news with strong sentiment signal
• Machine translation — Dhivehi side of a parallel corpus
• Spell checking / autocorrect — Language model for Thaana keyboard input
• Information retrieval — Build Dhivehi search or RAG systems

Limitations

• ~86% news domain — limited literary, legal, and colloquial register coverage
• Date coverage varies by source; many articles missing publication dates
• No manual annotation — domain tags and quality scores are automatically assigned

Citation

@dataset{dhivehi_corpus_2025,
  author    = {d3b4g},
  title     = {Dhivehi Corpus: A Large-Scale Text Corpus for the Maldivian Language},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/d3b4g/dhivehi-corpus}
}

License

CC-BY-4.0. You are free to use, share, and adapt this dataset with attribution. Original article content copyright remains with respective publishers.