---
license: cc-by-4.0
language:
- as
- brx
- grt
- kha
- lus
- mni
- njz
- pbv
- trp
- nag
tags:
- sentence-transformers
- embeddings
- retrieval
- northeast-india
- low-resource
- multilingual
- RAG
base_model: sentence-transformers/LaBSE
---

<p align="center">
  <img src="https://huggingface.co/MWirelabs/ne-embed/resolve/main/Ne_embed_Mwirelabs_banner.jpg" alt="NE-Embed Banner">
</p>

<div align="center">

### Purpose-built Multilingual Embeddings for Northeast Indian Languages

> **10 languages • 201k parallel pairs • 768 dimensions • Built on LaBSE**

Semantic search, Retrieval and RAG for low-resource Northeast Indian languages.

</div>

---

## Highlights

- Supports **10 Northeast Indian languages**
- Optimized for **Semantic Search, Retrieval and RAG**
- Up to **7× higher retrieval accuracy** than raw LaBSE on low-resource languages
- Built on **sentence-transformers/LaBSE**
- Trained on **201,738 balanced English ↔ Northeast language parallel pairs**
- Released under the **CC-BY-4.0** license

## What is NE-Embed?

NE-Embed is a multilingual sentence embedding model designed for semantic understanding across Northeast Indian languages. It is optimized for **semantic search**, **dense retrieval**, **Retrieval-Augmented Generation (RAG)**, and **cross-lingual information retrieval**, where general-purpose multilingual embedding models often perform poorly.

The model is fine-tuned from **LaBSE** using **201,738 balanced English↔Northeast language parallel pairs** spanning **10 languages**. It substantially improves retrieval quality for several low-resource languages—including Garo, Khasi, Nyishi, Pnar, and Kokborok—while maintaining strong multilingual alignment.

---

## Why NE-Embed?

General-purpose multilingual embedding models are trained on hundreds of languages, but many Northeast Indian languages receive little or no representation during training. As a result, semantically similar sentences are often mapped far apart, leading to poor retrieval performance.

NE-Embed addresses this gap through targeted contrastive fine-tuning on balanced parallel data, producing embeddings that better capture semantic similarity for low-resource Northeast Indian languages while preserving multilingual compatibility.

---

## Supported Languages

| Code | Language | Script | Tier | Training Pairs |
|------|----------|--------|------|----------------|
| `asm` | Assamese | Bengali | ✅ Supported | 25,000 |
| `brx` | Bodo | Devanagari | ✅ Supported | 25,000 |
| `grt` | Garo | Latin | ✅ Supported | 25,000 |
| `kha` | Khasi | Latin | ✅ Supported | 25,000 |
| `lus` | Mizo | Latin | ✅ Supported | 25,000 |
| `mni` | Meitei | Meitei Mayek | ✅ Supported | 25,000 |
| `njz` | Nyishi | Latin | ✅ Supported | 25,000 |
| `trp` | Kokborok | Latin | ⚠️ Limited | 12,545 |
| `pbv` | Pnar | Latin | ⚠️ Limited | 6,034 |
| `nag` | Nagamese | Latin | ⚠️ Limited | 1,996 |

> **Supported** = strong retrieval performance. **Limited** = model has coverage but quality is lower; use with caution in production.

---

## Performance

Evaluated on 500 samples per language. CLRI = Cross-Language Retrieval Interference (lower is better).

| Language | R@1 (Base) | R@1 (NE-Embed) | CLRI (Base) | CLRI (NE-Embed) |
|----------|-----------|----------------|-------------|-----------------|
| Assamese | 95.6 | **97.4** | 1.8% | 4.6% |
| Bodo | 55.8 | **99.8** | 61.0% | **3.0%** |
| Garo | 13.2 | **90.8** | 88.8% | **3.0%** |
| Khasi | 28.6 | **95.6** | 65.0% | **3.4%** |
| Mizo | 46.6 | **91.8** | 58.4% | **9.4%** |
| Meitei | 13.6 | **34.2** | 90.8% | **19.8%** |
| Nyishi | 10.2 | **75.0** | 71.0% | **17.4%** |
| Pnar | 27.2 | **86.2** | 79.6% | **8.0%** |
| Kokborok | 26.4 | **71.6** | 63.8% | **11.8%** |
| Nagamese | 77.0 | **88.0** | 17.8% | **8.4%** |

> Base = raw LaBSE zero-shot. All CLRI reductions represent genuine cross-lingual confusion fixed by fine-tuning.

---

## Quick Start

```python
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("MWirelabs/ne-embed")

sentences = [
    "Where is the nearest hospital?",                          # English
    "Ngi lah ia shong ha ki shnong baroh",     # Khasi
    "Pilakchin an·senganiko man·na am·tokenga.",               # Garo
]

embeddings = model.encode(sentences, normalize_embeddings=True)
similarities = model.similarity(embeddings, embeddings)
print(similarities)
```

### Recommended for RAG / Hybrid Retrieval

```python
# Hybrid: NE-Embed dense + BM25 char 3-gram sparse
score = 0.7 * ne_embed_score + 0.3 * bm25_score
```

---

## Training

- **Base model:** `sentence-transformers/LaBSE`
- **Loss:** `MultipleNegativesRankingLoss`
- **Data:** 201,738 English↔NE language parallel pairs, capped at 25k per language to prevent Assamese attractor bias
- **Epochs:** 3 · **Batch size:** 64 · **Max seq length:** 128
- **Hardware:** 1× NVIDIA A40 (48 GB) · **Training time:** ~1.3 hours

---

## Intended Uses

- Semantic search
- Dense retrieval
- RAG
- Cross-lingual retrieval
- Clustering

---

## Citation

```bibtex
@misc{mwirelabs2026neembed,
  title        = {NE-Embed: Multilingual Text Embeddings for Northeast Indian Languages},
  author       = {MWire Labs},
  year         = {2026},
  publisher    = {HuggingFace},
  howpublished = {\url{https://huggingface.co/MWirelabs/ne-embed}},
  note         = {CC-BY-4.0}
}
```

---

<div align="center">

Built with ♥ in Shillong, Meghalaya · [MWire Labs](https://mwirelabs.in) · Part of the **NE-Stack**

*NE-LID · NE-BERT · NE-Embed · Kren · Aganbo · Klam*

</div>