BeitTigreAI/tigre-sonar-encoder

Tigre Low-Resource Language Resource Collection (Tigre-Data 1.0)

Overview

This repository introduces the first comprehensive public collection of resources for the Tigre language — an under-resourced South Semitic language within the Afro-Asiatic family. The release aggregates multiple modalities (text + speech) and provides baseline models for NLP tasks including language modeling, automatic speech recognition (ASR), and machine translation.

The goal of Tigre-Data 1.0 is to accelerate research in low-resource NLP, morphologically rich language modeling, speech technologies, and cross-lingual transfer.

---

tigre-sonar-encoder

A Tigre–English semantic similarity and quality-checking encoder, fine-tuned from the SONAR universal embedding model.

Key Capabilities

• Generates 1024-dimensional embeddings for Tigre and English text
• Computes cosine similarity for translation validation and filtering
• Supports retrieval, clustering, and cross-lingual semantic tasks

---

Model Description

Input Language: Tigre (tig, script: Ethiopic — tig_Ethi)
Base Model: facebook/nllb-200-distilled-1.3B
Model Type: Encoder-only (text embedding model)
Purpose: Align Tigre embeddings with the universal SONAR cross-lingual space

---

Training Method: Knowledge Distillation

The model was trained with a teacher–student distillation pipeline:

1. Model & Tokenizer Preparation

• Initialized from the NLLB-200 distilled encoder
• Extended tokenizer with Tigre-specific vocabulary
• New token embeddings initialized by averaging sub-token embeddings

2. Teacher Embedding Generation

• SONAR embedding model used as the Teacher
• English translations of Tigre sentences encoded into 1024-dimensional vectors

3. Distillation Fine-Tuning

• Minimized Mean Squared Error (MSE) loss between Student (Tigre encoder) and Teacher embeddings
• Forced the Tigre model to align with the universal cross-lingual space

---

Training Details

• Dataset: train_tig_parallel_text.parquet
• Contents: Tigre sentences paired with gold-standard SONAR embeddings
• Objective: MSE loss between model output and SONAR target vectors
• Tokenizer: Extended NLLB tokenizer with Tigre-specific vocabulary

---

Evaluation Results

Metric	Result	Description
Accuracy (Source → Target)	0.88	Retrieval accuracy when querying with Tigre text
Accuracy (Target → Source)	0.78	Retrieval accuracy when querying with English text
BLEU	30.74	(BLEU relates to a separate MT evaluation, not this encoder)

---

Usage Example (Python)

pip install transformers torch

```python

from transformers import AutoTokenizer, M2M100ForConditionalGeneration
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"

model_id = "BeitTigreAI/tigre-sonar-encoder"
seq2seq = M2M100ForConditionalGeneration.from_pretrained(
    model_id,
    subfolder="model"  
)
encoder = seq2seq.get_encoder().to(device).eval()
tokenizer = AutoTokenizer.from_pretrained(model_id, subfolder="model")

@torch.inference_mode()
def embed(texts, lang):
    tokenizer.src_lang = lang
    batch = tokenizer(texts, return_tensors="pt", padding=True, truncation=True, max_length=512).to(device)
    out = encoder(**batch, return_dict=True)
    mask = batch["attention_mask"].unsqueeze(-1).float()
    pooled = (out.last_hidden_state * mask).sum(dim=1) / mask.sum(dim=1).clamp_min(1.0)
    return torch.nn.functional.normalize(pooled, p=2, dim=1)

def score_pair(tig, eng):
    t = embed([tig], "tig_Ethi")
    e = embed([eng], "eng_Latn")
    sim = float((t*e).sum())
    return round(sim*100, 1)

print(score_pair("እት እድንየ እግል ትርኤ ተሐዜዮ ተቅዪር ግበእ", "Be the change that you wish to see in the world"))
print(score_pair("ክል ዶል ኢገብእ መስል እስከ ይከለስ", "It always seems impossible until it's done"))
   
---

## License

**CC BY-SA 4.0**