Frontman-11/Yoruba-to-english-translation

# 🧠 Yoruba ↔ English Neural Machine Translation This project focuses on building a Neural Machine Translation (NMT) system to translate between Yoruba and English using a Transformer-based architecture inspired by Google's Attention Is All You Need paper. Yorùbá is a low-resource language predominantly spoken in Nigeria and across the diaspora. Existing parallel datasets are limited and largely religious in nature. Despite these challenges, this project successfully trains high-performing translation models using fewer than 600,000 sentence pairs. --- ## 🚀 Project Overview - **Goal**: Build monolingual Transformer models for Yoruba ⇄ English translation. - **Architecture**: Custom implementation of the Transformer model using TensorFlow/Keras. - **Pipeline Phases/Directory Structures**: - `data/train/`: Pretraining on large datasets from scratch. - `data/dev/`: Fine-tuning and experimentation on smaller or domain-specific subsets. - `data/test/`: Final evaluation and inference. - `notebook/`: Contains high-level execution logic and experimental notebooks. - `models/`: Contains the models saved from the notebooks. - `utils/`, `transformer/`, `tokenizer/`: Custom utility modules, Transformer layers and model, and tokenization logic. --- ## 📊 Results - **Training Accuracy**: Peaked around the 60% range (masked accuracy). - **BLEU Score**: Ranged from 70%–78% across ~6,000+ evaluation samples. - **Best Performance**: Achieved when evaluating Yoruba → English translations in batch mode on GPU. --- ## 🖥️ Platform - All training and fine-tuning were done using the free GPU on Kaggle. --- ## 📁 Dataset Sources All data used were manually processed, cleaned, and renamed before training. The names seen in the codebase reflect the final cleaned versions. ### 🔗 Aggregated Public Datasets - Andrews2017/africanlp-public-datasets - Origin of many datasets including: - `train.tsv`, `dev.tsv`, `test.tsv` from menyo-20k - `test_yo_en.tsv`, `GNOME_yo_en.tsv`, `opus_yo_e …