AletheiaAING/gemma-3-270m-african-diacritizer
Text Generation • 0.3B • Updated • 9
AI & ML interests
# Building AI that speaks people's languages We are rooted in innovation and inclusivity. We bridge the gap between technology and humanity, creating AI solutions tailored to diverse cultural, linguistic, business and industrial landscapes. Most AI is built in a handful of languages, for a handful of contexts. We start from the other direction: the languages, accents, scripts and everyday realities of the people who are under-served, then we make the technology fit them, small enough to run where they are. ## What we build **Language models for African languages.** The SabiYarn family: compact models pretrained from scratch on English, Nigerian Pidgin and 21 African languages, including Yorùbá, Hausa, Igbo, Twi, Ewe, Fon, Efik, Ibibio and Urhobo. We favour small, efficient, sparse architectures (a 125M dense model upcycled to a 306M mixture-of-experts with ~207M active parameters) because capable AI should be affordable to run. **Translation.** English-to-African-language models fine-tuned and reinforced with quality estimation (AfriCOMET), so output is judged on meaning, not only on word overlap. **Assistants you can trust.** A structured conversation format with an explicit "unsure" answer, and tool use and retrieval for what a small model should not memorise, instead of confident guessing. **Data and evaluation infrastructure.** Licence-aware multilingual corpora, a language-identification and deduplication pipeline, instruction data in 20 languages, and a data-triage classifier that decides what belongs in a model's weights, what needs a tool, how technical it is, how hard it is to translate, and how much mathematics it teaches. ## How we work - **Inclusive by design**: low-resource languages are the starting point, not an afterthought. - **Small and efficient**: models that fit real budgets, devices and latency needs. - **Honest about uncertainty**: calibrated confidence, documented limits, no overclaiming. - **Open about data**: every source is licence-checked and recorded. - **Built with communities**: native speakers shape evaluation and review. ## Work with us We partner with organisations that want AI in the languages and workflows of their customers, in education, agriculture, health information, finance, public services and customer support. Reach us at <contact>. Collaborations, evaluations and native-speaker review are welcome.