You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This corpus is text published by Maldivian news sites, fiction sites and other publishers, who keep the copyright in it. It is shared for research and non-commercial use, such as training and studying Dhivehi language models. Do not republish the articles or stories, and remove any document on a publisher's request.
Log in or Sign Up to review the conditions and access this dataset content.
Dhivehi web corpus
1.82 billion characters of Dhivehi (Thaana) text from 25 public Maldivian websites, cleaned and split into passages. It is the web-crawled part of the pretraining data for Nau 1.0.
| Config | Passages | Characters | Sources |
|---|---|---|---|
news |
651,434 | 1,575M | 18 news sites (below) |
fiction |
25,052 | 205M | esfiya.com (serialised novels), laragevaahaka, vaahaka.mv (short stories) |
majlis_web |
4,525 | 14M | Public pages and documents on the People's Majlis website |
religious |
3,429 | 14M | dhisalafiyyah.net |
laws |
3,193 | 12M | Full texts of Maldivian laws from vakeelu.com |
gazette |
1,869 | 5M | The Government Gazette, gazette.gov.mv |
all (default) |
689,502 | 1,824M | Everything above |
News sites, largest first: vaguthu, mihaaru, sun, avas, raajje, dhauru, khabaru, psmnews, fiyaonline, thepress, hurihaa, halinews, atollnews, masverin, xeetimes, haftha, asuru, gaafu.
Fields
| Field | |
|---|---|
id |
crawl:<site>:<article>[:<part>]. Long articles are split into parts |
site |
Short site name, e.g. mihaaru |
genre |
The config it belongs to |
url |
The article's address |
title |
The article's title |
date |
As shown by the site, when available. Some sites give no date, and some dates may be when a page was last changed |
text |
The passage |
Processing
- Thaana normalised (stray zero-width characters and a few common encoding faults fixed). Passages that are too short or mostly non-Thaana, and duplicates, are removed.
- The 5,000 source documents of nau-ai/dhivehi-eval, and every passage containing one of its test sentences, are removed (23,483 passages). The 22 laws used for the
lawqatest are not included. - The spelling is the publishers' own. It is not corrected to the Dhivehi Academy standard.
Not included
Nau 1.0 was also pretrained on text that is not published here:
- the public datasets by alakxender on Hugging Face, which are available there: dhivehi-news-corpus, dhivehi-sentences-extended, dhivehi-sentences-normalized and dhivehi-speeches
- English text, used to keep the model's English
Personal information
This is news. It names public figures and people in the news, and some articles print phone numbers and email addresses, almost all of them official hotlines and offices. Nothing was added, but nothing was removed either.
Copyright and takedown
The text belongs to its publishers. It is shared here, gated, for research and non-commercial use. If you are a publisher and want your site's text removed, open a discussion on this repository. Every passage carries its site and URL, so removal is quick.
- Downloads last month
- 45