🌍 Fine-tuning NLLB on Menyo-20k

This project fine-tunes Meta’s NLLB (No Language Left Behind) model using Parameter Efficient Fine-Tuning on the Menyo-20k dataset to improve translation quality for African languages. Training was conducted until early stopping to prevent overfitting.

---

📘 Overview

• Model: facebook/nllb-200-distilled-600M
• Dataset: Menyo-20k — parallel corpus for English–African language translation
• Goal: Enhance NLLB performance on low-resource African languages (Yoruba, Igbo, Hausa, etc.)
• Framework: Hugging Face Transformers

---

⚙️ Training Setup

Setting	Value
Batch Size	8
Learning Rate	1e-4
Gradient Accumulation Step	8
Scheduler	Cosine
Epochs	50
Early Stopping	Patience = 10
Environment	Kaggle (T4 GPU x 2)

---

🧩 Method

1. Preprocess and tokenize Menyo-20k using NLLB tokenizer
2. Fine-tune the model using LoRA and with Huggingface Trainer
3. Apply early stopping based on validation loss
4. Evaluate with BLEU and qualitative translation tests

---