🗨️ Data, Languages and African Model-Building

📍 The Map: Charting the Landscape of African NLP

If you read only one survey paper before going further into the African NLP literature, make it this one.

🌍 The Global Indigenous Data-Sovereignty Arc

Before we look at the models, we look at the framework under which African researchers are choosing to release the data the models are built on. This conversation does not start with African scholarship; it sits inside a longer, global Indigenous data-sovereignty tradition. Naming that lineage matters, because the African work is part of a wider movement that has been theorising community-grounded data governance for considerably longer than the current AI moment has been running.

For policy context at the African continental level: the AU Data Policy Framework (adopted February 2022, published 28 July 2022) sets the stated continental orientation toward transparency, accountability, inclusion, and equity, aimed at an African Digital Single Market by 2030. The AU Continental AI Strategy (July 2024) explicitly invokes data sovereignty across its five focus areas. Both documents are real and worth reading; both, as we noted in 11.4, are policy positions rather than enforced regulatory mechanisms.

🤝 Community Infrastructure: Masakhane, Lanfrica, AfricaNLP, Indaba

Before the foundation models, before the benchmarks, before the licensed datasets — there is the community infrastructure that makes everything else possible. Four organisations or venues are doing the disproportionate share of the work.

🧠 The African Foundation-Model Inventory

What follows is a verified inventory of the foundation models — encoder, from-scratch decoder, adapted decoder, and named-language — that are actually shipping for African languages as of May 2026. Two centrepieces matter most for this course: InkubaLM (the first sovereign African small language model, from Lelapa AI) and MzansiLM (the UCT-built decoder covering all eleven official South African written languages).

🎯 The African Benchmark Stack

Benchmarks are how a field measures progress. The African benchmark stack has expanded substantially in 2024 and 2025, with several important 2026 additions. What follows is the verified inventory you can actually use to evaluate African-language work.

⚖️ Two Strategic Positions

There is a real and currently unresolved disagreement inside the African AI community about what the right strategic move is. The argument is worth understanding clearly because it shapes which kinds of project a postgraduate researcher might join.

The pedagogical point of putting these two positions side by side is not to declare a winner. Both are sovereignty practices in different registers; both have honest arguments behind them; the right answer for a particular project depends on what the project is for. The useful exercise is to be able to read a piece of African-AI work and identify clearly which strand it sits in, because the success criteria for the two are quite different. A 125M-parameter community-licensed model that runs on a feature-phone-class device is a success in the resource-efficient register; it is a partial step in the frontier-aspirant register. A nation-state frontier-scale model that is announced but not yet running is a partial step in the sovereign register; it is mostly absent from the resource-efficient one.

🔍 Where the Gaps Are

The single most useful thing this sub-lesson can do for the postgraduate audience is identify, concretely, what is missing in the current African data-and-models landscape. Six gaps are particularly thesis-shaped: each one is a genuine research opening that could be pursued from a UCT or comparable position.

1. No African-language LegalBench equivalent. Legal NLP benchmarks exist for English; the equivalent for any African language does not. Africa has substantial bodies of customary law, statutory law, and judicial decisions in many of its languages. A LegalBench-for-Swahili or LegalBench-for-isiXhosa would be a benchmark contribution and a step toward AI-supported access to law for people who do not work in English.
2. No African clinical QA benchmark grounded in local guidelines. MedQA exists; an African equivalent that takes local clinical guidelines, local epidemiology, and local pharmacological reality seriously does not. This is the kind of benchmark that would let African researchers evaluate foreign frontier models against local clinical reality before deploying them in healthcare settings.
3. No PRISMA-style reproducibility audit of African NLP. Imam et al.'s 2025 ASR review is the closest equivalent, and found <15% of African ASR studies reproducible. A wider study covering the full African NLP literature would surface the actual size of the reproducibility gap and would, by itself, be a publishable contribution.
4. No published Africa-specific critique-and-adaptation of CARE. A paper arguing what CARE would need to look like as a genuinely African framework rather than a wholesale Indigenous import is missing. The constructive work (Esethu) is there; the critical-theoretical companion is not.
5. No continent-level inventory of endangered-language ML resources. N|uu (Khoisan), Khwedam, and several Cushitic, Nilotic, and Khoisan languages are at risk; the inventory of which of them have any ML data at all does not exist. SADiLaR's University of Pretoria node and the Wits DSFSI lab hold parts of the picture; a unified continental inventory would let endangered-language documentation and AI work coordinate more effectively.
6. Operational uptake of community licensing remains low. Despite Esethu, despite Kaitiakitanga, despite CARE, most African datasets in 2024 and 2025 still ship under CC-BY or research-only licences. There is room for both empirical research on why, and for practical work helping projects move to community-grounded licences when it is appropriate.

If you are a postgraduate looking for a thesis topic in this part of the AI landscape, any of these gaps is a real opening — not in the sense that they are easy, but in the sense that the field will recognise the contribution.

🎯 What This Means for Your Research

• Use the benchmarks. The African benchmark stack is real and broad enough to evaluate work in most of the major NLU and ASR areas. If your project involves African-language NLP and you are not evaluating against at least one of AfroBench / IrokoBench / AfriMTEB / Afri-MCQA / AfriHate / AfriSenti / AfriSpeech-200, you are working blind.
• Use the models that exist. For inference and adaptation work on the languages they cover, InkubaLM, MzansiLM, AfroXLMR, Serengeti, and the named-language models are usable starting points. The frontier-aspirant N-ATLAS-style projects are not yet downloadable; the resource-efficient ones are.
• Consider community licensing when releasing data. If you collect or curate a dataset, especially one with significant cultural or linguistic content, look at Esethu and Kaitiakitanga before defaulting to CC-BY. The community-licensing track is genuinely more aligned with the African sovereignty conversation; defaulting away from it without considering it is a missed opportunity.
• Pick the gaps deliberately. The six gaps above are not the only ones, but they are concrete and verifiable. If you are starting a project, choosing one of them by intention is more likely to produce a publishable contribution than choosing a topic only because foundation-model work is currently fashionable.
• Resource-efficient is a sovereignty practice. The Lelapa argument generalises: matching your work to the compute and data you actually have access to, rather than to the headline numbers of foreign labs, is the operationally honest version of the sovereignty position covered in 11.4.

✏️ A Short Exercise

1. Pick a language and a domain relevant to your current or planned research project. The domain might be medicine, agriculture, law, climate science, education, the humanities, or another.
2. Check what benchmarks and datasets exist for that language × domain combination. Use the table above; check Lanfrica; check the AfricaNLP 2025 proceedings.
3. Identify what is missing. Is there a benchmark? Is there a dataset of the right size and quality? Is there a model that handles the language?
4. Write a one-paragraph proposal for what you would build first if you were going to fill that gap. Be honest about scope — would this be a year of work, a thesis, a community collaboration?
5. Bring it to class. We will pool the proposals across the cohort and look at which gaps the group together could imaginably address.

📚 Sources & Further Reading

Coming up in 11.6: we turn from data and models to the remaining two layers of the stack — policy and talent. The African Union Continental AI Strategy in its operational detail, the national strategies of South Africa / Kenya / Rwanda / Nigeria / Egypt (and the peer-reviewed comparative analysis in Yilma & Wodajo's Science and Public Policy special section), and the institutional landscape of Deep Learning Indaba, AIMS, ARIN, and Lelapa as the talent pipeline. We close 11.6 with the question that opens Week 12 (the integrative capstone): given everything we now know about African AI capacity, what is the most useful thing a postgraduate in this room can actually do?