Masakhane began in 2018 as an insight by two South African researchers and launched formally in 2019 in Nairobi. In under eight years its evaluation datasets — MasakhaNER, MasakhaNER 2.0, MasakhaPOS, IrokoBench — became the yardstick for measuring language models across dozens of African languages. But the “grassroots” label now tells half the story: Masakhane is a registered Kenyan foundation with multi-million dollar funding from IDRC, the Gates Foundation and Google.org.
In 2021 a group of native-speaker volunteers sat down with articles from local newspapers in Nigeria, Ethiopia, Rwanda and Kenya and marked them word by word: which tokens were person names, which were organisations, which were places. The tool was called ELISA. There were ten languages: Amharic, Hausa, Igbo, Kinyarwanda, Luganda, Luo, Nigerian Pidgin, Swahili, Wolof, Yoruba. The result — MasakhaNER, published in the journal TACL with David Ifeoluwa Adelani as first author — was the first large public named entity recognition dataset for African languages. It won the Best Paper Award at that year’s AfricaNLP Workshop.
This is where the artificial intelligence enters the story, and it is worth being precise because it is the part most coverage skips. Masakhane did not primarily build models. It built the ground on which models get measured. A named entity recognition system is a machine learning model trained to classify every token in a sentence; a large language model is a different beast entirely. But both, in order to be evaluated, need text annotated by humans who actually speak the language. Without that text, saying a model “works in Swahili” is a claim with no numbers behind it. Evaluation datasets are the piece of infrastructure nobody writes about and without which everything else is vapour.
Before this, there was JW300
To understand the break, you need to know the starting point. For many African languages, before community-built datasets the only open corpus available was JW300: a collection of religious content translated for missionary purposes. A translation model trained on that learns a language with a very particular vocabulary, register and universe of reference. Try asking it to translate a public health bulletin, a commercial contract, or a conversation between teenagers.
Masakhane — the word means “we build together” in isiZulu — started in 2018 from an insight by two South African machine learning researchers, Jade Abbott and Laura Martinus, who identified the research gap and decided it would not close by waiting for someone else to close it. The formal launch came in 2019 at the Deep Learning Indaba in Nairobi, the conference that has served for years as the gathering point for the African machine learning community. Among the founding members is Vukosi Marivate of the University of Pretoria, one of the Indaba’s own engines.
The method came before the results
Masakhane’s most formally decorated contribution is not a dataset. It is a paper about method. Participatory Research for Low-resourced Machine Translation: A Case Study in African Languages, by Nekoto et al. in 2020, with an extremely long author list — the list is the argument — won the first Wikimedia Foundation Research Award of the Year in 2021.
The claim is simple and uncomfortable: in low-resource language research, the bottleneck is not the model, it is the relationship with the people who speak the language. The standard approach — a Global North lab collects data, publishes, speakers maybe appear in the acknowledgements — produces brittle results and leaves nothing behind locally. The participatory approach puts annotators in the author list and in the framing of the problem. It costs more in coordination and it produces communities that are still there the following year.
Everything else follows from that. MasakhaNER 2.0, presented at EMNLP in 2022, extended the framework to twenty African languages with person, organisation, location and date entities, and training sets ranging from roughly a thousand to roughly eight thousand sentences per language. Then MasakhaPOS, for part-of-speech tagging. Then, in 2024, IrokoBench: a benchmark built explicitly to evaluate large language models on African languages — that is, to answer with numbers a question that press releases usually settle with an adjective.
The people
The collective portrait is more interesting than any single biography. Jade Abbott, co-founder, is now CTO and co-founder of Lelapa AI, a company doing applied research on African languages: the move from informal network to company is one available trajectory. David Adelani, who led MasakhaNER, is at McGill and Mila in Canada with a Canada CIFAR AI Chair: the academic trajectory, with its unresolved knot — a share of the expertise built in Africa ends up housed in North American and European institutions.
Kathleen Siminyu chairs the board of the Masakhane Research Foundation. Hatem Haddad sits on the board. Everlyn Asiko Chimoto started from eight thousand Luhya–English sentence pairs and now organises AfricaNLP and KenyaNLP: the organiser’s trajectory, the one that keeps standing the conferences where a community recognises itself.
On the size of the network the numbers are deliberately imprecise, and it is worth saying so rather than picking the showiest figure. A 2022 academic survey counts over a thousand members; journalistic sources write “thousands”. On published models and datasets, one preprint cites over four hundred models and over twenty datasets; a 2025 survey counts 359 models and 25 datasets on the organisation’s Hugging Face account. None of these figures is officially confirmed by Masakhane. Even MasakhaNER’s citation count shifts with the counter: roughly 341 on Google Scholar, between 90 and 105 on Dimensions. Anyone writing about infrastructure should always name the instrument.
What the “grassroots” label no longer covers
There is a convenient narrative about Masakhane — African volunteers building for free what big labs then take without thanks — and it needs careful handling, because the available documentation tells a different and in some ways more complicated story.
IrokoBench, the 2024 benchmark, was funded by the Lacuna Fund, co-founded by the Rockefeller Foundation, Google.org and Canada’s IDRC. It used Oracle compute credits, a Cohere For AI Research Grant, and API credits granted by OpenAI so the team could evaluate GPT-3.5 and GPT-4. All of them appear in the acknowledgements. The Masakhane Consortium had four of six proposals funded by the Lacuna Fund, with Knowledge4All on operations, Strathmore CIPIT on legal and Bantucracy on ethics.
Today the Masakhane Research Foundation is based in Kilifi, Kenya. The project The Masakhane AI languages hub is funded by IDRC with 5,360,000 Canadian dollars over thirty-seven months. In September 2025 a Gates Foundation grant went to the foundation to develop language models and tools. The African Languages Hub opened a call for datasets in fifty African languages with support from Google.org, the UK’s FCDO, IDRC and the Gates Foundation.
None of this cancels the argument about data extractivism, which remains central to how Masakhane talks about its own work. It relocates it. The question is no longer “who takes without citing”, because the credits are there and they are traceable. The question is what happens to a participatory research community when it becomes an entity with millions on the balance sheet, reporting deadlines, a board, and institutional donors with priorities of their own. The method Wikimedia honoured in 2021 worked partly because nobody had to answer to a grant cycle. Now somebody does.
One thing the wider debate about linguistic sovereignty in AI rarely notices: Masakhane did not build the first translation system for Yoruba, Swahili or Amharic, and does not claim to have. Its documented firsts concern the datasets used to evaluate. That is a strategically stronger position than it sounds. Whoever controls the benchmark decides what counts as “it works”, and that is the one lever a community without its own GPUs can hold onto while everyone else trains.