Garnet-Owl/taura-2.0

# Taura 2.0 > *"Taũra"* is Kikuyu for *"Translate."* **Taura** is an open-source machine translation engine for **Kikuyu ↔ English**, one of the first of its kind for this low-resource Bantu language. The Kikuyu are Kenya's largest ethnic group, with 9 to 10 million speakers making up roughly 17 to 20% of the country's population. Modern translation tools largely ignore Kikuyu. Taura exists to change that, starting with strong linguistic foundations: curated parallel corpora, morphological segmentation tuned to Kikuyu's agglutinative structure, and cross-lingual embedding alignment using Orthogonal Procrustes mapping. ## What We're Building Taura is not just a translation API. It is a full pipeline: - **Corpus curation:** parallel sentence pairs from the Bible (Matthew, Mark, Luke, John, Acts, Romans, 1–2 Corinthians) and agriculture sector data (coffee, dairy, poultry, potato, banana, mango, cabbage, avocado): over 9,700 curated pairs and growing - **Cross-lingual alignment:** monolingual FastText embeddings aligned via iterative Orthogonal Procrustes with three anchor sources (parallel sentence embeddings, seed dictionary, and identical-string vocabulary pairs) and CSLS-based refinement - **Translation API:** FastAPI service with retrieval and word-by-word modes, top-K candidates, and a lightweight web UI The long-term goal is to push translation quality to the point where Taura is genuinely useful to Kikuyu speakers, and to serve as a replicable blueprint for other low-resource African languages. ## Performance Progression > This section tracks the evolution of Taura's cross-lingual alignment. Every row represents a real training run, evaluated on a 100-sentence held-out validation set. > [!NOTE] > **Metrics Key:** > - **BLEU / chrF:** Measures translation similarity to a human reference. Higher is better. > - **Top-1 / Top-5:** The percentage of times the exact correct translation was in the #1 or Top 5 retrieved results. > - **MRR (Mean Reciprocal …