Introducing OribAI 1.0
An instruction-tuned language model for Hausa and Yoruba speakers — fine-tuned on 27,498 curated conversational pairs to preserve cultural nuance and linguistic integrity.
Model Specifications
Trained on curated Pan-African data. Available in full precision and quantized GGUF for local deployment.
OribAI-14B
Instruction-tuned conversational model fine-tuned on 27,498 unique Hausa and Yoruba pairs. Based on Qwen2.5-14B-Instruct with LoRA adapters trained via Unsloth + TRL.
- LanguagesHausa, Yoruba
- Base ModelQwen2.5-14B-Instruct
- TrainingLoRA r=32, α=64 · 3 epochs
OribAI-14B GGUF
4-bit quantized (Q4_K_M) export for local inference via llama.cpp and Ollama. Run OribAI on consumer hardware — no GPU or cloud required.
- Quantization4-bit NF4 (Q4_K_M)
- Runtimellama.cpp, Ollama
- Est. Memory~10 GB RAM
Measured on CohereForAI/aya_dataset
Perplexity evaluated on 50 held-out samples from the AYA dataset train split. Yoruba performance is strong; Hausa v2 improvements are planned.
| Language / Task | Perplexity (↓ better) | Quality Assessment | Deployment Status |
|---|---|---|---|
| Yoruba — Fluent Generation | 3.22 PPL | Strong | Production Ready |
| Yoruba — Factual Q&A | 3.22 PPL | Strong | Production Ready |
| Hausa — Short Factual Q&A | 62.54 PPL | Limited | Use with Care |
| Hausa — Open-ended Generation | 62.54 PPL | Unreliable | v2 Planned |
Built on Industry Standard Foundations
Build Anywhere. Run Locally.
Seamless integration with existing deep learning stacks and enterprise infrastructure.
Open & Responsible AI for All
OribAI is more than a model — it's a digital public good. We believe technological autonomy is a fundamental right for the next generation of African innovators.
UN SDG Goal 4
Quality Education: Enabling native-language AI tutoring for millions of students.
UN SDG Goal 9
Industry & Innovation: Building resilient infrastructure for regional tech independence.
UN SDG Alignment
Directly contributing to Sustainable Development Goals 4 (Quality Education) and 9 (Industry, Innovation & Infrastructure) through open access LLMs.
Cultural Safety
Dedicated safety layers designed to filter and protect against cultural misrepresentation and regional biases in synthetic data.
Responsible Use
Our cultural safety guardrails are developed in collaboration with linguists and historians to prevent algorithmic bias and ensure dialectal integrity. We prioritize transparency and model traceability.
Start Contributing
The frontier of African AI is open to all developers.
Documentation
Quickstart guides, API reference, deployment instructions, and hardware requirements for OribAI-14B.
Model Weights
Download full-precision weights or the Q4_K_M GGUF for local inference directly from Hugging Face Hub.
Contribute
Join the open-source effort. Contribute training data, evaluation scripts, and dialect coverage for v2.