Yaser77/mbart-en-te

English → Telugu Translation (mBART)

This repository contains an English to Telugu neural machine translation model based on mBART-50, fine-tuned and deployed using the Hugging Face ecosystem.

The model translates English text (en) into Telugu (te) and is suitable for research, learning, and demo purposes.

---

🔎 Live Demo

You can test the model live using the Hugging Face Space below:

👉 https://huggingface.co/spaces/Yaser77/mbart-en-te-demo

---

🧠 Model Details

• Model name: mbart-en-te
• Base model: facebook/mbart-large-50-many-to-many-mmt
• Task: Machine Translation (English → Telugu)
• Framework: 🤗 Transformers
• Model size: ~600M parameters
• Precision: FP32
• Tokenizer: SentencePiece (mBART tokenizer)

---

📌 How It Works

• Source language: English (en_XX)
• Target language: Telugu (te_IN)
• The model uses forced BOS token decoding, which is required for mBART-based translation models.

---

🚀 Intended Uses

✅ Direct Use

• Translating English text into Telugu
• Educational demos and learning projects
• NLP experimentation with Indic languages

🔄 Downstream Use

• Can be fine-tuned further on domain-specific parallel corpora
• Can be integrated into web apps or APIs for translation services

❌ Out-of-Scope Use

• Not intended for real-time production systems without optimization
• Not suitable for legal, medical, or safety-critical translations

---

⚠️ Limitations & Risks

• Translation quality depends heavily on sentence complexity
• May struggle with idioms, slang, or highly technical language
• Biases present in the original training data may be reflected
• Slower inference on CPU due to model size

---

🧪 How to Use the Model (Code)

from transformers import MBartForConditionalGeneration, MBart50TokenizerFast

model_id = "Yaser77/mbart-en-te"

tokenizer = MBart50TokenizerFast.from_pretrained(model_id)
model = MBartForConditionalGeneration.from_pretrained(model_id)

tokenizer.src_lang = "en_XX"
forced_bos_token_id = tokenizer.lang_code_to_id["te_IN"]

text = "Hello, how are you?"

inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(
    **inputs,
    forced_bos_token_id=forced_bos_token_id,
    max_length=128
)

print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])

---

📊 Training Details

Training Data

• Based on multilingual translation data used by mBART
• No custom dataset is claimed beyond fine-tuning and experimentation

Training Method

• Sequence-to-sequence Transformer
• Teacher forcing during training
• Cross-entropy loss

---

🧾 Evaluation

• Qualitative evaluation using common English sentences
• No official BLEU or automatic metrics reported
• Intended mainly as a demo and learning project

---

🌍 Environmental Impact

• Hardware: CPU (Hugging Face Spaces – Basic tier)
• Compute Region: Managed by Hugging Face
• Precision: FP32
• Carbon Emissions: Not explicitly measured

---

🧩 Technical Specifications

• Architecture: Transformer encoder–decoder (mBART)

• Libraries:

  • transformers
  • torch
  • sentencepiece
  • gradio

---

📖 References

• mBART Paper: https://arxiv.org/abs/2001.08210
• Hugging Face Transformers: https://huggingface.co/docs/transformers

---

Acknowledgements

This project was developed as part of a hands-on learning workshop conducted by Vijender P at Alumnx AI Labs during the "GPU Hours" session on fine-tuning and deploying large AI models.

The deployment, debugging, Hugging Face integration, and demo Space were independently implemented by the author as a learning exercise.

---

👤 Author

T Mohamed Yaser Computer Science Engineering Student Interested in NLP, ML deployment, and real-world AI applications

---

📬 Contact

For questions, feedback, or collaboration:

• Hugging Face: https://huggingface.co/Yaser77