nau-ai/Nau-OCR-1.0
Image-to-Text • Updated
nau-ai
non-profit
AI & ML interests
None defined yet.
Recent Activity
Organization Card
Nau · ނައޫ
Open language models for Dhivehi (Þ‹Þ¨ÞˆÞ¬Þ€Þ¨), the language of the Maldives, written in Thaana.
Nau 1.0 writes, translates, summarises and corrects Dhivehi, and runs on an ordinary laptop without an internet connection. It is a 5.4B-parameter model built on Gemma 4 E2B, with a tokenizer extended for Thaana.
| On held-out tests | Nau 1.0 | Gemini 3.8 Flash |
|---|---|---|
| Translation, Dhivehi ↔ English | 0.696 | 0.677 |
| Dhivehi skills (6 tasks) | 0.481 | 0.408 |
| Romanised Dhivehi → Thaana | 0.765 | 0.118 |
| General knowledge (8 questions) | 4/8 | 8/8 |
Models
- Nau-1.0: weights for transformers, plus GGUF files for llama.cpp, LM Studio and Ollama
Datasets
- dhivehi-corpus: 1.8 billion characters of Dhivehi from 25 public websites (news, fiction, religious writing, laws, the Gazette)
- nau-sft: 386k Dhivehi instruction examples
- dhivehi-eval: held-out test sets, with model answers
Everything is open for research. The text datasets are gated because the publishers keep the copyright in their articles and stories. To report a problem or ask for a takedown, open a discussion on the repository concerned.