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 dataset is built from Dhivehi text published by news sites, fiction sites and other publishers, who keep the copyright in it. It is shared for research and non-commercial use. Do not republish the text as your own, and remove any document on a publisher's request.
Log in or Sign Up to review the conditions and access this dataset content.
Nau instruction data
The instruction-tuning data behind Nau 1.0, a Dhivehi (Thaana) language model. Every row is a chat in the usual messages format (role, content).
| Config | Rows | What it is |
|---|---|---|
nau-1.0 (default) |
382,695 train, 3,907 validation | The tasks Nau 1.0 was fine-tuned on |
identity |
253 | Its name, and answers that decline to invent live or private facts |
gemini-translations |
37,404 | English translations of human-written Dhivehi sentences by Gemini 3.8 Flash. Not used to train Nau 1.0 |
What is in nau-1.0
Tasks built from human Dhivehi text, in Dhivehi and English:
- translation between Dhivehi and English (from public parallel datasets and crawled bilingual news)
- summaries, headlines and lead paragraphs of news articles
- continuing a text, and putting shuffled sentences back in order
- spelling correction toward the Dhivehi Academy standard, including synthetic errors
- romanised Dhivehi ("aharen mi hafthaa…") to Thaana
- choosing the more natural of two sentences (A/B)
- questions about the definitions in Maldivian laws
- vocabulary tasks from the Radheef dictionary
Almost every task is built mechanically from human text: a word blanked out, sentences shuffled, errors injected, a headline paired with its article. About 550 tasks (short summaries, English explanations, questions and multiple-choice items) were written with Claude and checked automatically. The question–answer pairs from alakxender/dhivehi-qa-dataset are included as published.
What was removed
- Every test item. The 5,000 source documents of the test sets and every test sentence were kept out, so nau-ai/dhivehi-eval stays a clean test.
- Unreleased sources. 44,742 examples built from text that is not published here were removed, so this config is about 10% smaller than the data Nau was actually trained on.
Sources and licences
The tasks are derived from:
- text crawled from Dhivehi news, fiction, religious and legal websites (see nau-ai/dhivehi-corpus), whose publishers keep the copyright
- public datasets by alakxender on Hugging Face, each under its own licence (mostly Apache 2.0 or MIT;
dhivehi-finetranslationanddv-synthetic-errors-lgdeclare none) - the Radheef dictionary of the Dhivehi Academy
Because of the mix, there is no single licence. Use it for research and non-commercial work. The gemini-translations config was generated with Gemini 3.8 Flash and is also subject to Google's terms for Gemini outputs.
Contact
Questions, problems or takedown requests: open a discussion on this repository.
- Downloads last month
- 45