You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Nau Embed is licensed under CC BY-NC-SA 4.0 and is also subject to Google's Gemma Terms of Use, because it is built on EmbeddingGemma. Non-commercial use is free: credit Nau and share anything you build from it under the same licence. Commercial use, including running it inside a paid product or service, needs a separate commercial licence from the Nau authors; ask by opening a discussion titled "Commercial licence" on this repository.

Nau Embed 1.0 (ނައޫ)

A Dhivehi text-embedding model for search and retrieval. It turns Dhivehi (Thaana), romanised Dhivehi and English text into 768-number vectors, so a question in any of the three finds the right Dhivehi passage.

It is a fine-tune of Google's EmbeddingGemma-300m: 300M parameters, small enough to embed an archive on an ordinary CPU.

Results

NTREX English↔Dhivehi bitext mining (MMTEB task NTREXBitextMining, eng_Latn-div_Thaa, 1,997 sentence pairs, macro F1, same metric as MMTEB). NTREX was never used for training.

Model	Size	en→dv	dv→en
Nau Embed 1.0	300M	0.979	0.980
gemini-embedding-001	API	0.952	0.971
EmbeddingGemma-300m (base)	300M	0.386	0.394
Qwen3-Embedding 0.6B–8B (MMTEB)	0.6–8B	0.09–0.16
multilingual-e5-large-instruct (MMTEB)	560M	0.078	0.113
bge-m3 (MMTEB)	568M	0.025	0.068

KaLM-Embedding-Gemma3-12B reports 0.984 / 0.987 on MMTEB, slightly higher, at 40 times the size.

Dhivehi retrieval (held-out dev set: 6,751 queries against one pool of 12,378 Dhivehi passages; Recall@10, the share of queries whose correct passage is in the top 10).

Query type	Nau Embed 1.0	gemini-embedding-001	Base
English question → Dhivehi passage	0.997	0.997	0.584
English keyword search → Dhivehi passage	0.990	0.991	0.511
Romanised Dhivehi headline → article	0.951	0.762	0.394
Thaana headline → article	0.920	0.893	0.715
Dhivehi question → answer	0.750	0.601	0.365
Word → dictionary definition	0.632	0.549	0.357

The dev set comes from the same kinds of sources as the training data (held out by document).

English-only retrieval dropped from the base model (NanoBEIR mean nDCG@10: 0.545 vs 0.614). Use it for Dhivehi and cross-lingual search; for English documents with English queries, the base model is better.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("nau-ai/Nau-Embed-1.0")   # gated: log in and accept the licence first

docs = ["ރައްޔިތުންގެ މަޖިލީހުގެ ނައިބު ރައީސްކަމަށް ހޮވުނީ ...", "..."]
doc_vecs = model.encode(docs, prompt_name="document")                    # "title: none | text: "
q_vecs = model.encode(["Who was elected deputy speaker?",
                       "majlis ge naibu raees hovun"], prompt_name="query")  # "task: search result | query: "
scores = model.similarity(q_vecs, doc_vecs)

Always encode questions with prompt_name="query" and passages with prompt_name="document". Leaving the prompts out lowers accuracy.
Query and document vectors must come from this same model. They are not comparable with vectors from any other model, including gemini-embedding-001.
To match English and Dhivehi sentences one-to-one (translation pairs), use prompt_name="BitextMining" on both sides.
Vectors can be truncated to 512, 256 or 128 numbers (Matryoshka training). Re-normalise after truncating, and truncate queries and documents the same way.
Use float32 or bfloat16, not float16 (EmbeddingGemma does not support float16).
Speed: about 2,000 characters of Thaana per second on a 12-thread desktop CPU.

Training

Contrastive fine-tuning with in-batch and mined hard negatives, inside a Matryoshka loss (768/512/256/128), for one epoch over 772k pairs. The result was then merged with the base model (0.8 × fine-tuned + 0.2 × base) to recover English retrieval.

Pairs	Count	Query side	Passage side
English question / keyword search → passage	195k	Gemini-written English	human Dhivehi
English ↔ Dhivehi sentences	169k	English: Gemini translations and checked parallel data	human Dhivehi
Romanised headline → article	120k	human (news sites publish romanised headlines)	human Dhivehi
Thaana headline → article	80k	human	human Dhivehi
Romanised ↔ Thaana sentences	60k	human	human Dhivehi
English sentence → Dhivehi sentence (search prompts)	50k	as above	human Dhivehi
Word → definition	30k	Radheef dictionary	Radheef dictionary
Dhivehi question → answer	8.6k	alakxender/dhivehi-qa-dataset
English question → answer (to keep English working)	60k	sentence-transformers/gooaq

Every Dhivehi text in training is human-written; machine-written text appears only on the English side. Pairs whose Dhivehi was under 90% dictionary words were dropped. All evaluation documents and sentences, and all of NTREX, were removed before training.

Licence and attribution

CC BY-NC-SA 4.0, with a commercial licence on request, and subject to the Gemma Terms of Use (see LICENSE and NOTICE).

Non-commercial use (research, teaching, personal and public-interest projects) is free. Credit Nau Embed, and share anything you build from it (fine-tunes, merges, quantisations, distilled models) under the same licence.
Commercial use, including running it inside a paid product or service, needs a commercial licence from the Nau authors. Its terms include publishing the improvements you make. To ask for one, open a discussion titled "Commercial licence" on this repository.
Because it is built on EmbeddingGemma, use is also governed by Google's Gemma Terms of Use and Prohibited Use Policy. These apply to everyone, including commercial licensees.

Access requires accepting these terms in the form at the top of this page.

Related: Nau 1.0, the Dhivehi chat model.

Questions or problems: open a discussion on this repository.

Downloads last month: 2

Safetensors

Model size

0.3B params

Tensor type

F32

Model tree for nau-ai/Nau-Embed-1.0

Base model

google/embeddinggemma-300m

Finetuned

(292)

this model