More than two million Kenyans speak Ekegusii. None of the major translation systems handle it, so public notices reach those speakers in English and Kiswahili or not at all. A team at USIU-Africa built software that translates into Ekegusii, and it works.
Ekegusii is spoken by more than two million people in Kisii and Nyamira. It is not one of the 200 languages carried by NLLB-200, and the commercial translation engines do not serve it either.
That gap has a practical cost. When a government agency issues a public service announcement about a health campaign, a land registration deadline or a school funding programme, it goes out in English and Kiswahili. Reaching Ekegusii speakers in their first language means paying for human translation, every time, for every notice. Most of the time it does not happen.
The USIU team set out to change the economics of that, with a model that translates real announcements and is scored on a test set of announcements it never saw during training.
Training a language into a model from scratch needs data nobody has for Ekegusii. The team instead started the new language from what the model already knows about its neighbours.
| Corpus source | Sentence pairs | Share | What it teaches |
|---|---|---|---|
| Ekegusii scripture, aligned to English and Kiswahili | 56,866 | 90.7% | General grammar and vocabulary |
| Kenyan public service announcements | 5,692 | 9.1% | The target domain, upsampled four times |
| Everyday contemporary sentences | 111 | 0.2% | Modern register |
A further 2,993 scripture pairs, 944 announcements and 200 everyday sentences were held back and never trained on. Every score on this page comes from that held-out material.
Full fine-tune of NLLB-200-distilled-600M. Ekegusii added to the tokenizer and its embedding initialised from Kikuyu with 1% noise. Effective batch 48, learning rate 5e-5, three epochs, bf16 with gradient checkpointing, 8-bit AdamW, on a single NVIDIA A100 80GB. Announcement data upsampled four times so the domain is not drowned out by scripture. Scored with BLEU and chrF against held-out references.
Translations are scored by comparing them against a human translation of the same sentence. BLEU and chrF are the two standard ways of doing it, and higher is better. Under 2, the output is not recognisably the right sentence. In the teens, the meaning comes through and the grammar is rough. That is the move this project made.
Everyday contemporary speech is weak. BLEU moved from 0.07 to 1.59, which is real movement but not a usable translation. The reason is visible in the corpus: the model has 56,866 scripture pairs to learn from and 111 everyday ones. That is the next data problem to solve, and the team says so in their own model card rather than leaving it out.
The team also tested a two-stage curriculum, general text first and announcements second, against a single pass over everything mixed together. The two-stage version had more training updates and still lost on every test set. Both versions are published, so the comparison can be checked by anyone.
This model was trained on Navon compute during the USIU programme, on a cluster shared between sixteen research teams who filled it. The work has since been given a machine of its own, and the next phase is more than twice the data across three different kinds of model.
The team has collected far more material since. The next version learns from roughly 150,000 example sentences and pits three different kinds of model against each other, so the choice is settled by testing rather than by opinion.
Purpose-built for translation and already proven on this language pair.
A different multilingual pretraining recipe at a comparable size. Tests whether the gains are specific to NLLB.
A general purpose large language model, ten times the size. Tests whether scale beats specialisation here.
Alongside the translation work the team is training a domain classifier, so incoming announcements are routed and handled by category. The whole system is being built to be handed over and run in production, rather than to sit in a paper.
The point of this project is not one language pair. It is that the recipe is within reach of any university team with a corpus and a GPU.
62,669 sentence pairs is within reach of a motivated university team. It is not a dataset that needs a large company to assemble.
One GPU, hours rather than weeks. The barrier to a Kenyan language getting machine translation is data collection and expertise, not a data centre.
Starting a new language from a related one the model already knows is repeatable across the other Kenyan Bantu languages.
The corpus, the model and the people are Kenyan. The compute is Navon’s own rather than rented from a hyperscaler, which is why a project like this can be given a machine to itself.
This project worked for one reason. The team held something nobody else held, a body of Ekegusii that had never been collected before. The model was ordinary. The data was not.
If your organisation has built up a large store of its own records, documents, transactions, images or sensor readings, we want to build a reinforcement learning environment on top of it. In plain words, a practice ground where a model attempts a real task from your world, is scored on whether it got it right, and tries again until it is good at it.
What comes out is not a general chatbot. It is a narrow intelligence that is genuinely good at one thing your industry cares about, trained on ground nobody else can reach, and it stays yours.
At a university with work held up by hardware? The invitation on university compute still stands.