---
language:
- ne
- en
tags:
- translation
- fairseq
- fsmt
license: apache-2.0
---

# Nepali-Tamang-English Machine Translation Model

This model was converted from a Fairseq checkpoint to Hugging Face Transformers format. It performs machine translation between Nepali, Tamang, and English.

## Model Details
- **Architecture:** FSMT (Fairseq Machine Translation)
- **Original Framework:** Fairseq
- **Languages:** Nepali, Tamang, English

## Usage

This model requires specific language tokens to indicate the target language:
- `__ne__`: Translate to Nepali
- `__ta__`: Translate to Tamang
- `__en__`: Translate to English

### Installation
```bash
pip install transformers
```

### Python Example
```python
from transformers import FSMTForConditionalGeneration, FSMTTokenizer

model_name = "rishi70612/nepali-tamang-english-mt"
tokenizer = FSMTTokenizer.from_pretrained(model_name)
model = FSMTForConditionalGeneration.from_pretrained(model_name)

def translate(text, target_lang_token):
    # Prepend the language token 
    input_text = f"{target_lang_token} {text}"
    input_ids = tokenizer.encode(input_text, return_tensors="pt")
    
    # Generate with beam search for better quality
    outputs = model.generate(input_ids, num_beams=5, max_length=100)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Examples
print("Nepali:", translate("I love machine learning.", "__ne__"))
print("Tamang:", translate("I love machine learning.", "__ta__"))
print("English:", translate("मलाई नेपाल मन पर्छ।", "__en__"))
```

## Usage in Kaggle
You can use the following code snippet directly in a Kaggle Notebook cell to run inference.

```python
# 1. Install transformers
# !pip install transformers
# !pip install sacremoses

import torch
from transformers import FSMTForConditionalGeneration, FSMTTokenizer

# 2. Load Model
model_name = "rishi70612/nepali-tamang-english-mt"
print(f"Loading model from {model_name}...")

tokenizer = FSMTTokenizer.from_pretrained(model_name)
model = FSMTForConditionalGeneration.from_pretrained(model_name)

if torch.cuda.is_available():
    model = model.cuda()
    print("Moved model to GPU.")

# 3. Define Translation Function
def translate(text, target_lang_token):
    # Prepare input (e.g. "__ne__ I love AI")
    input_text = f"{target_lang_token} {text}"
    
    input_ids = tokenizer.encode(input_text, return_tensors="pt")
    if torch.cuda.is_available():
        input_ids = input_ids.cuda()
        
    outputs = model.generate(input_ids, num_beams=5, max_length=100, early_stopping=True)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# 4. Run
print("Nepali:", translate("I love machine learning.", "__ne__"))
print("Tamang:", translate("I love machine learning.", "__ta__"))
```