felekekinfe/NLP-Amharic-to-English-Translator-with-MT5

```markdown # Amharic-English Machine Translation with mT5 & LoRA This repository contains the complete pipeline for fine-tuning a multilingual Text-to-Text Transfer Transformer (mT5) to translate Amharic to English. The project leverages **Low-Rank Adaptation (LoRA)** to achieve high-performance results on consumer-grade hardware (Google Colab T4 GPU). --- ## 🚀 Project Overview Amharic is a "low-resource" language with significant morphological complexity. This project addresses these challenges through a specialized preprocessing pipeline and parameter-efficient fine-tuning (PEFT). ### Key Features * **Base Model**: `google/mt5-small` (Encoder-Decoder architecture). * **Efficiency**: **LoRA** (Rank 16, Alpha 32) updates <1% of total parameters. * **Preprocessing**: Phonetic-aware Unicode Normalization for the Ge'ez script. * **Evaluation**: Multi-angle assessment using **BLEU** and **chrF** scores. --- ## 📂 Project Structure * `preprocessing.py`: Contains the `clean_and_normalize_dataset` function for Unicode mapping and data cleaning. * `main.py`: The fine-tuning script utilizing Hugging Face `Trainer` and `peft` for LoRA training. * `dataset/`: Directory containing dataset files. * Various CSV files: `cleaned_dataset.csv`, `cleaned_master.csv`, `converted_test.csv`, `converted_train_1.csv`, `converted_train.csv`, `preprocessed_dataset.csv` - processed datasets at different stages. --- ## 🛠️ Installation ### Prerequisites - Python 3.8+ - CUDA-compatible GPU (recommended for training, though CPU training is possible) ### Dependencies ```bash pip install pandas torch numpy datasets transformers peft evaluate sacrebleu chrf sentencepiece ``` --- ## 📖 Methodology ### 1. Unicode Normalization Amharic features redundant characters (e.g., ሐ, ኀ, ሀ) that represent the same sound. Our preprocessing script maps these to a canonical form, which: * Reduces vocabulary sparsity. * Prevents the model from treating identical phonetic concepts as different words. * …