African languages barely exist in the data that AI learns from.
Of every 1,754 pages on the open web, one is in an African language.
Africa has more than 2,000 languages. The 28 that Common Crawl tracks account for 0.057% of its pages. Models learn from the data they are given, so a language missing from the data is missing from the model.
What Afirka LLM does
We collect, clean and verify African web text. We generate instruction data grounded in African contexts. We adapt Google's Gemma models to these languages and to the low-resource settings they are used in. Alongside that we build the evaluation benchmarks, the data infrastructure and the review platform the work runs on.
Everything we produce is released openly: the datasets, the models, the tools and the platform itself. We also run a public seminar series on large language models that regularly draws more than 50 researchers.
Growing Africa’s share of the open web
African languages are missing from the web itself, so we started upstream. With Common Crawl we ran a community seeding campaign, then a targeted extraction of our own. The result improves the raw material for anyone who builds on the open web.
A community seeding campaign
85 contributors across 19 countries curated 700 seed URLs covering 34 African languages. Common Crawl used them in its late-2025 crawl, which collected 28.75 million pages from 15,561 domains. African domains grew 38-fold, adding 343,000 African-language pages and lifting Africa's share of Common Crawl to its highest recorded level.
Languages that grew the most, in a single crawl
Our targeted extraction
From that enlarged crawl we ran AfriCC, our own extraction pipeline. It sifted an 892,000-page shortlist down to 109,062 confirmed African pages, each language agreed on by two independent detectors, then merged them with a re-verified public web dataset into the 10.2M-document collection.
From the open web to a fine-tuned model
Raised Africa's share of the web itself
A seeding campaign with Common Crawl: 85 contributors across 19 countries curated 700 seed URLs in 34 languages. The late-2025 crawl collected 28.75M pages from 15,561 domains, a 38-fold expansion in African domains. Tswana grew 279%, Sango 259%, Luganda 160%. Anyone who builds on Common Crawl gets the benefit.
Built and cleaned a large African text collection
Our targeted crawl, AfriCC, recovered 109,062 African pages from an 892,000-page shortlist. We re-verified a public web dataset of 10,146,834 documents and merged the two, de-duplicated, into one collection of about 10.2M documents.
Made sure we know what every document is
Two independent language detectors must agree on every document. We built CommonLID, a benchmark of 373,230 human-annotated web lines in 109 varieties. On African languages, frontier models trail specialised tools by about 24 F1, which suggests much of the "missing" African data is mislabelled rather than absent.
Understood what the collection contains
No tool could sort African web pages by topic, so we trained one. It is 91% accurate and we ran it over all 10.2M documents. Most of the collection turns out to be everyday web writing rather than news.
Audited the public instruction data
We catalogued 355 public instruction datasets and standardised 164 of them, about 367M examples across 108 languages. Mislabelled tags were the biggest distortion: text tagged "unknown" fell from 29% to 0.2% after the audit. Every fix is a re-runnable script.
Generated instruction data grounded in Africa
58,634 examples across 40 languages from six pipelines: African-entity Q&A, translation with round-trip checks, self-instruction, code-switching, cross-lingual pivoting and functional-genre writing. Native-speaker review runs alongside.
Started AfriGuard, a safety corpus
422,375 review items across 40 languages, stratified and classified: the first safety corpus for African languages. Fewer than 0.1% of model outputs currently show safe-refusal behaviour in these languages. Reviewers work under consent gates.
Chose the base model by measurement
We measured how efficiently four tokenizers handle our 40 languages. Gemma 4 is best at 2.36 tokens per word, and the only one that keeps Amharic and Tigrinya workable: Ge'ez script costs over 8 tokens per word elsewhere, against 2.69 to 3.14 under Gemma 4.
Built a platform that runs the whole pipeline
Dataset ingestion, language coverage, tokenizer statistics, training launches with full provenance, and native-speaker review. Every step is recorded and auditable.
Mapped the African speech-data landscape
Catalogued 77 speech datasets across 394 dataset-language pairs, recording licences, consent, speaker counts, and recorded against transcribed hours, before committing compute.
The data, in numbers
Pretraining collection
Instruction and task data
By task type
| Task | Datasets | Examples |
|---|---|---|
| loading… |
Safety and speech
What the collection is about
Most African web text is everyday life rather than headlines. Pages that fit no clear topic are labelled as such instead of being forced into one.
First fine-tuned Gemma models
We settled on NVIDIA's NeMo AutoModel and fine-tuned Gemma with lightweight adapters on our own instruction data. The pilot covered six languages on a small budget and still produced large gains, including 53 F1 points on Igbo question answering.
| Language | Q&A (F1) | Translation (ChrF++) | Topic acc. |
|---|---|---|---|
| Hausa | +37.1 | +2.9 | +10.5 |
| Igbo | +53.3 | +6.8 | +40.3 |
| Kinyarwanda | +31.5 | +0.4 | +28.8 |
| Swahili | +17.3 | +12.7 | +24.5 |
| Yoruba | +4.4 | +3.4 | +30.6 |
| Zulu | +45.9 | +7.7 | +21.2 |
Gains over the base Gemma model. Pilot: Gemma E2B, LoRA (rank 16), two epochs, AfriInstruct.
Scaling carefully before the big run
We are scaling in steps, with 6-, 10- and 14-language runs under a fixed budget, to find a pipeline that holds before the main pretraining pass. In parallel we test ways to keep the model's English reasoning intact as it gains African languages, starting on small models where mistakes are cheap.
Nothing of unknown origin gets into the model.
- Every document's language is confirmed by two independent detectors that must agree.
- Each document is marked African or not, and non-African text is excluded. 503 documents were removed from the training set as a governance check.
- Mislabelled data is the real bottleneck. Text tagged "unknown" fell from 29% to 0.2% after our audit, and every fix is a reviewable, re-runnable script.
- Native-speaker review is built into data generation. Early Yoruba review caught real errors, including a bird term used for insects and inconsistent tonal diacritics, which reshaped the next round.
- Reviewers work under consent gates and opt-outs. Licences and provenance are tracked for every dataset.
- The data, the tools and the methods are openly documented and reproducible.
What comes next
- Benchmark the 14-language and per-language-budget models, and release the first fully fine-tuned Gemma as a public baseline.
- Run pretraining data-mix experiments, three mixes at 10 billion tokens each, to pick the recipe.
- Run the main pretraining pass at 100–200 billion tokens on the winning mix, then fine-tune it for the headline model.
- Extend the tokenizer's vocabulary for Amharic and Tigrinya in parallel.
- Scale native-speaker review into a trusted gold evaluation set of about 400 records and use it to sharpen automatic filtering.
- Publish the AfriCC dataset paper, and design the speech pathway toward a speech-capable model.
Support and infrastructure
Local hardware carries the fast-iteration work: pipeline development, adapter prototypes, the language-ID and topic classifiers, evaluation, and the review platform. Cloud credits then go only to runs already de-risked locally. The exception is the 100–200 billion token pretraining pass, which needs accelerator capacity only the cloud provides.
On-premise
A DGX Spark (2025 NVIDIA grant), two A6000 GPUs, and an RTX 4090 node, brought under shared Slurm scheduling with a common storage pool.
Cloud
Google Cloud compute for gathering, preparing, and training data; Vertex AI and Gemma credits for fine-tuning and evaluation across all 40 languages.
Partners we draw on
- Common Crawl. The community seeding shared task and ongoing collaboration on African web data.
- NVIDIA academic grant (2025). Access to a DGX Spark for local training and inference.
- AI4D African Languages Lab. Data collection and human-capacity development.
- African Next Voices. The largest human-validated African speech collection, feeding the speech workstream.
We need native speakers of African languages.
We are building a dataset across 40 African languages to improve language technology and support linguistic research. Judging whether the text is accurate and natural takes people who speak these languages.
What you'll do
- Verify data. Review samples and say whether they sound natural and correct.
- Check the automatic systems. Tell us where they disagree with your judgement.
- Improve coverage. Decide what good output looks like in your language.
- Flag unsafe output. Mark text that is harmful or culturally wrong.
Benefits for contributors
- Co-authorship. Be added as an author on the resulting dataset research paper.
- Research opportunities. Potential to coordinate further studies based on the insights gathered.
- Long-term collaboration. Explore future participatory research initiatives.