Saugat212/ne-en-nllb-model

English-Nepali Translation Model

Overview

Fine-tuned translation model for converting English text to Nepali (and vice versa) using Facebook's NLLB-200 distilled model.

Model Details

Property	Value
Model ID	Saugat212/ne-en-nllb-model
Base Model	facebook/nllb-200-distilled-600M
Architecture	m2m_100
Parameters	0.6B
License	apache-2.0

Purpose

• Translate English text to Nepali (EN→NE)
• Translate Nepali text to English (NE→EN)
• Domain-specific translation using custom fine-tuned weights

Contents

File	Description
Fine_Tuning.ipynb	NLLB fine-tuning notebook
Fine_Tuning_nllb.ipynb	NLLB-specific fine-tuning
transformer_finetuning.ipynb	Alternative transformer fine-tuning
data_clean.ipynb	Data cleaning notebook
Data Fetching from translator.ipynb	Fetching parallel data
inference.ipynb	Translation inference notebook
opus-translation.py	OPUS-based translation
Finetune.md	Quick setup guide
NLLB_Finetuning_Documentation.md	Detailed NLLB docs

Usage

Load Model

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_name = "Saugat212/ne-en-nllb-model"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

Translate EN→NE

def translate_en_to_ne(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
    out = model.generate(**inputs, forced_bos_token_id=tokenizer.lang_code_to_id["ne_Latn"], max_new_tokens=128)
    return tokenizer.decode(out[0], skip_special_tokens=True)

print(translate_en_to_ne("Hello, how are you?"))

Translate NE→EN

def translate_ne_to_en(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
    out = model.generate(**inputs, forced_bos_token_id=tokenizer.lang_code_to_id["en_Latn"], max_new_tokens=128)
    return tokenizer.decode(out[0], skip_special_tokens=True)

print(translate_ne_to_en("नमस्ते, तपाईं कस्तो हुनुहुन्छ?"))

Requirements

• transformers
• torch
• pandas
• datasets

Fine-tuning

To fine-tune on custom data:

1. Prepare CSV with English_Sentence and Nepali_Translation columns
2. Use Fine_Tuning.ipynb or Finetune.md as reference
3. Adjust hyperparameters based on GPU memory