You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These test items quote Dhivehi text published by news sites and other publishers, who keep the copyright in it. They are shared for evaluating language models. Do not train on them.
Log in or Sign Up to review the conditions and access this dataset content.
Dhivehi evaluation sets
Held-out tests for Dhivehi (Thaana) language models, used to evaluate Nau 1.0. None of these items, or the documents they came from, were in Nau's training data. Please do not train on them.
| Config | Items | What it tests | Scoring |
|---|---|---|---|
auto (default) |
546 | Six Dhivehi skills built from human text (below) | exact match or chrF |
translate300 |
592 | 296 sentence pairs, translated in both directions | chrF |
lawqa |
702 | Questions on the definitions in 22 Maldivian laws, open-book and closed-book | chrF |
prompts |
78 | Open prompts: 42 writing, summary and translation tasks, plus 36 identity, "can you know this" and general-knowledge questions | read by hand |
results |
2,432 | Nau 1.0 and Gemini 3.8 Flash answers to auto, translate300 and prompts |
auto: six skills
dimension |
Items | Task | Scorer |
|---|---|---|---|
cloze |
115 | Fill in the missing word | exact |
continuation |
95 | Continue a news passage | chrF |
fluency_preference |
106 | Pick the natural sentence of two (A/B) | exact |
orthography |
110 | Rewrite a sentence with its spelling mistakes fixed | exact |
translation_dv_en |
60 | Translate news into English | chrF |
transliteration |
60 | Romanised Dhivehi into Thaana | exact |
Thaana is normalised on both sides before exact matching. 85 items are marked leaked_in_runs_1_2: their source sentences were found in parallel datasets used by early training runs. The published Nau scores use the other 461. They report the mean of the six skills, where the 17 remaining transliteration items weigh as much as the 115 cloze items.
Results so far
| Nau 1.0 | Gemini 3.8 Flash | |
|---|---|---|
auto, mean of 6 skills (clean items) |
0.481 | 0.408 |
translate300 (chrF) |
0.696 | 0.677 |
Per-skill scores and the behaviour checks are on the Nau 1.0 model card. The results config has every answer, so any score can be recomputed.
Sources
- News articles from Dhivehi news sites (see nau-ai/dhivehi-corpus), and public datasets by alakxender on Hugging Face.
translate300: sentence pairs held out from public parallel data. Each reference translation was checked with Gemini 3.8 Flash, and only pairs judged accurate were kept.lawqa: laws from vakeelu.com, held out from training whole.
Copyright in the quoted text stays with its publishers.
Contact
Questions, problems or takedown requests: open a discussion on this repository.
- Downloads last month
- 73