---
license: mit
tags:
  - machine-translation
  - mbart
  - multilingual
  - huggingface
  - peft
  - lora
  - english
  - telugu
datasets:
  - HackHedron/English_Telugu_Parallel_Corpus
language:
  - en
  - te
library_name: peft
inference: false
widget:
  - text: "Hello, how are you?"
---

# 🌍 LoRA-mBART50: English ↔ Telugu Translation (Few-shot)

This model is a parameter-efficient fine-tuned version of [facebook/mbart-large-50-many-to-many-mmt](https://huggingface.co/facebook/mbart-large-50-many-to-many-mmt) using [LoRA (Low-Rank Adaptation)](https://arxiv.org/abs/2106.09685) via the Hugging Face PEFT library.

It is fine-tuned in a **few-shot setting** on the [HackHedron English-Telugu Parallel Corpus](https://huggingface.co/datasets/HackHedron/English_Telugu_Parallel_Corpus) using just **1% of the data (~4.3k pairs)**.

---

## 🧠 Model Details

- **Base model**: `facebook/mbart-large-50-many-to-many-mmt`
- **Languages**: `en_XX` ↔ `te_IN`
- **Technique**: LoRA (r=8, α=32, dropout=0.1)
- **Training regime**: 3 epochs, batch size 8, learning rate 5e-4
- **Library**: 🤗 PEFT (`peft`), `transformers`, `datasets`

---

## 📚 Dataset

- **Source**: [HackHedron/English_Telugu_Parallel_Corpus](https://huggingface.co/datasets/HackHedron/English_Telugu_Parallel_Corpus)
- **Size used**: 4338 sentence pairs (~1%)
- **Format**:
  - `english`: Source text
  - `telugu`: Target translation

---

## 💻 Usage

### Load Adapter with Base mBART
```python
from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
from peft import PeftModel

# Load base model & tokenizer
base_model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")
tokenizer = MBart50TokenizerFast.from_pretrained("your-username/lora-mbart-en-te")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "your-username/lora-mbart-en-te")

# Set source and target languages
tokenizer.src_lang = "en_XX"
tokenizer.tgt_lang = "te_IN"

# Prepare input
inputs = tokenizer("Hello, how are you?", return_tensors="pt")
generated_ids = model.generate(**inputs, forced_bos_token_id=tokenizer.lang_code_to_id["te_IN"])
translation = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
print(translation)
````

---

## 🔧 Training Configuration

| Setting         | Value    |
| --------------- | -------- |
| Base Model      | mBART-50 |
| LoRA r          | 8        |
| LoRA Alpha      | 32       |
| Dropout         | 0.1      |
| Optimizer       | AdamW    |
| Batch Size      | 8        |
| Epochs          | 3        |
| Mixed Precision | fp16     |

---

## 🚀 Applications

* English ↔ Telugu translation for low-resource settings
* Mobile/Edge inference with minimal memory
* Foundation for multilingual LoRA adapters

---

## ⚠️ Limitations

* Trained on limited data (1% subset)
* Translation quality may vary on unseen or complex sentences
* Only supports `en_XX` and `te_IN` (Telugu) at this stage

---

## 📎 Citation

If you use this model, please cite the base model:

```bibtex
@inproceedings{liu2020mbart,
  title={Multilingual Denoising Pre-training for Neural Machine Translation},
  author={Liu, Yinhan and others},
  booktitle={ACL},
  year={2020}
}
```

---

## 🧑‍💻 Author

Fine-tuned by **Koushik Reddy**, ML & DL Enthusiast | NLP | LoRA | mBART | Hugging Face

Connect: [Hugging Face](https://huggingface.co/Koushim)