ilprl-docse/NepTam20k-transformer-49.07M

NepTam20K-Transformer

📝 Model Overview

NepTam20K-Transformer is a vanilla Transformer model fine-tuned for Nepali ↔ Tamang neural machine translation.

Originally implemented in Fairseq, the model has been converted to HuggingFace MarianMT format for easier deployment and inference.

The model was developed by the Information and Language Processing Research Lab (ILPRL) at Kathmandu University, as part of low-resource MT research.

It uses Hindi (hi_IN) as a proxy for Tamang in tokenization, due to the absence of Tamang in standard tokenizers.

---

🌍 Supported Languages

Intended language	Tokenized as
Nepali	ne_NP
Tamang	hi_IN (proxy)

Supported directions:

• Nepali → Tamang
• Tamang → Nepali

---

⚙️ Training Setup

🔧 Transformer Architecture

• Encoder: 5 layers
• Decoder: 5 layers
• Embedding size: 512
• Feed-forward hidden: 2048
• Attention heads: 8
• Dropout: 0.3
• Weight decay: 1e-4
• Learning rate: 5e-4
• Scheduler: inverse square root
• Optimizer: Adam
• Epochs: 50
• Batch size: 16 effective

📦 Dataset

• NepTam20K (15K train / 5K test)
• Devanagari script
• Cleaned and normalized parallel corpus
• Source: ilprl-docse/NepTam-A-Nepali-Tamang-Parallel-Corpus

---

📊 Evaluation

Model	Param Size	Direction	BLEU	chrF	chrF++	METEOR	COMET
NepTam_Transformer	49.07M	ne→tam	37.71	71.71	67.74	58.20	75.00
		tam→ne	38.01	69.89	66.30	57.60	77.37

---

🚀 Usage (Inference)

This model uses the MarianMT HuggingFace API. The following code shows how to load and translate with it:

import torch
from transformers import MarianMTModel, MarianTokenizer

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

# Load the model and tokenizer
model_path = "ilprl-docse/NepTam20k-transformer-49.07M"
tokenizer = MarianTokenizer.from_pretrained(model_path)
model = MarianMTModel.from_pretrained(model_path).to(device)

# Language map
LANG_MAP = {
    "Nepali": "__ne__",
    "Tamang": "__taj__"
}

def translate_with_transformer(text, src="Nepali", tgt="Tamang"):
    src_lang = LANG_MAP[src]
    tgt_lang = LANG_MAP[tgt]
    
    # Prefix target language token (MarianMT-style)
    input_text = f"{tgt_lang} {text}"
    encoded = tokenizer(input_text, return_tensors="pt", padding=True).to(device)
    
    with torch.no_grad():
        generated_ids = model.generate(
            **encoded,
            max_length=512,
            num_beams=5,
            early_stopping=True
        )
    
    return tokenizer.decode(generated_ids[0], skip_special_tokens=True).strip()

# Example
text = "म आज धेरै खुसी छु।"
translation = translate_with_transformer(text, src="Nepali", tgt="Tamang")
print("Translation (Tamang):", translation)