OribAI
AI & NLP
Technological Sovereignty for African Languages: A 14B-parameter LLM fine-tuned on Hausa and Yoruba.
ClientInternal Research
Date2025
ServiceAfrican Language AI
CategoryAI & NLP
The Challenge
The vast majority of large language models are trained on English and Western-centric data. Hausa and Yoruba — spoken by over 150 million people combined — remain severely underrepresented, leaving hundreds of millions without AI systems that understand their language, culture, or context.
Our Solution
OribAI is an instruction-tuned 14B language model built on Qwen2.5-14B-Instruct and fine-tuned on 27,498 curated Hausa and Yoruba conversational pairs using LoRA adapters (r=32, α=64) via Unsloth + TRL. A 4-bit GGUF export enables local inference on consumer hardware with ~10 GB RAM via llama.cpp or Ollama — no cloud required.
Key Impact
- First high-quality instruction-tuned LLM for Hausa and Yoruba
- Runs offline on consumer hardware — no GPU or cloud dependency
- Open-sourced on Hugging Face for the global research community
- Preserves cultural nuance and linguistic integrity at scale
Have a similar project in mind?
Let's collaborate to build resilient, verifiable infrastructure for your community.
Visit WebsiteStart a Project