Research · Open Source AI

Building AI in the open.

Open Source AI research explores how openly shared models, datasets and tools can help more people across Africa build with artificial intelligence, and how we can contribute back to that work.

Shared openly, improved together.

Focus
Open models, data and tools
Priority
African languages and contexts
Guided by
Safety & Evaluation
Why it matters

AI built for Africa should not depend on a few closed systems built elsewhere.

Open models and datasets let researchers, startups and universities adapt AI to their own languages and needs, study how it behaves, and build on each other’s work instead of starting from scratch.

We are exploring which parts of our work can be shared openly, and how to do it responsibly, so the people who build and use AI across Africa can shape it too.

Research areas

Where the work happens.

Open Source AI is made up of 3 research areas. Each has its own questions, methods and links to our products.

OSA-01

Open Models for African Languages

Adapting openly available models to African languages, including tonal languages and the code-switching people use every day.

Most language models are trained mainly on a few high-resource languages. We study how open models can be adapted to under-resourced languages such as Yoruba, Hausa, Igbo, Swahili and Nigerian Pidgin, in written and spoken form, including mixed speech such as Yoruba–English, without losing what they already do well.

  • Adapting open models
  • Tonal and code-switched language
  • Small, task-specific adaptations
OSA-02

Open Datasets & Documentation

Building language and domain datasets with consent, and documenting them well enough that others can trust and reuse them.

Good open data is more than a download link. We look at how data is gathered and with whose consent, how it is licensed, what it does and does not represent, and how to document it clearly so that researchers and builders can decide whether it fits their needs.

  • Consent and provenance
  • Licensing
  • Clear dataset documentation
OSA-03

Open Evaluation

Shared tests that show how AI systems perform on African languages, tasks and contexts, so anyone can check a model before relying on it.

Popular benchmarks say little about how a model handles a Lagos market conversation or a clinic intake form. We explore evaluation sets grounded in local tasks and languages, and ways to publish methods and results that others can repeat and extend.

  • Locally grounded test sets
  • Repeatable methods
  • Reporting limitations honestly
Our approach

From shared work to better AI.

Open research only works when it is useful to others and safe to share. Each piece of work moves through the same cycle.

  1. 01

    Gather

    Collect data and needs from the communities the work is for, with consent.

  2. 02

    Adapt

    Build on open models and tools rather than starting from scratch.

  3. 03

    Evaluate

    Test against local tasks and our safety and evaluation principles.

  4. 04

    Share

    Release what is ready, with documentation, licences and known limits.

  5. 05

    Improve

    Learn from what others find, fix and build, then start again.

The questions we are asking

What we are exploring.

01

What should be shared openly?

We look at which models, datasets, evaluations and tools are most useful to share, and what needs more care before release.

02

How can open models serve African languages?

Studying how openly available models can be adapted for languages and dialects that most systems underserve.

03

How do we share responsibly?

Working through licensing, documentation and safety checks, so shared work is useful and harder to misuse.

04

How can a community improve AI together?

Exploring how researchers, developers and institutions across the continent can contribute data, testing and improvements.

Where it connects

Related work.