njerikahoro/Kikuyu-English-Machine-Translation

# Kikuyu-English-Machine-Translation # Thiomi Multilingual Machine Translation (MMT): Kikuyu ⇄ English 👉 **[Live Demo on Hugging Face Spaces] huggingface.co ** 👉 **[Model Checkpoint on Hugging Face Hub] huggingface.co ** An end-to-end Machine Translation (MT) pipeline featuring a fine-tuned, bidirectional... # Thiomi Multilingual Machine Translation (MMT): Kikuyu ⇄ English An end-to-end Machine Translation (MT) pipeline featuring a fine-tuned, bidirectional **NLLB-200 (No Language Left Behind)** model optimized for high-quality translation between **English (eng_Latn)** and **Gĩkũyũ / Kikuyu (kik_Latn)**. This repository also contains an interactive deployment interface powered by **Gradio**. --- ## Project Overview Low-resource language technology often relies on massive datasets or brittle parallel architectures. This project adapts Meta's `nllb-200-distilled-600M` via targeted sequence-to-sequence fine-tuning on localized language pairs. ### Key Features * **True Bidirectional Performance:** A single model checkpoint handling both `English ➔ Kikuyu` and `Kikuyu ➔ English` without code changes, using exact target prefix token controls. * **Balanced Shuffling:** The text preprocessing architecture flattens parallel rows into explicit language directions and interleaves them perfectly ($50\%$ train/test per path direction) to prevent model steering or catastrophic forgetting. * **Optimized VRAM Profile:** Implemented using 8-bit Paged AdamW optimizers, gradient checkpointing, and expandable segment allocation configs, enabling stable training loops on single-GPU instances. --- ## 📊 Dataset & Preprocessing The model is fine-tuned using the parallel text repository: `NjeriKahoro/thiomi-multilingual-text`. * **Train Samples:** 10,906 total rows (5,453 base parallel pairs flattened bidirectionally) * **Test Samples:** 2,730 total rows (1,365 base parallel p …