rachel

There are two main reasons for the limitations of African language machine translation: the training data and the expensive compute requirements of frontier open-models. In our latest work, we are addressing both by shipping compact small translation language models as well as our high-quality training data to support future work on what is a challenging long-term goal. The large open datasets for African languages are noisy, full of duplicates and imbalanced over languages. The few carefully curated sets are clean but cover too few languages, and human-translated sets are excellent and too small to make an impact. Every option trades away scale, coverage or quality. At the same time, the frontier open-source models are too large for the performance they provide, something we tackle with our SLMs.