---
language:
- yo
license: apache-2.0
tags:
- sovereign-ai
- yoruba
- peft
- lora
- tiny-aya
- kubernetes
base_model: CohereLabs/tiny-aya-earth
pipeline_tag: text-generation
datasets:
- masakhane/african-ultrachat
---

# TinyAya-Earth Yoruba LoRA Adapter

This repository contains a **QLoRA fine-tuned adapter** for [CohereLabs/tiny-aya-earth](https://huggingface.co/CohereLabs/tiny-aya-earth) (3.35B parameters), specialized on the **Yorùbá (`yo`)** instruction-following split of `masakhane/african-ultrachat`.

## Model Details
- **Developer:** Hussein Alamutu
- **Base Architecture:** Dense decoder-only transformer with Grouped Query Attention (GQA), 36 layers, 8k context window.
- **Base Model:** `CohereLabs/tiny-aya-earth`
- **Fine-Tuning Dataset:** `masakhane/african-ultrachat` (Yoruba split)
- **Training Method:** QLoRA (4-bit NF4 base, rank 16, alpha 32)
- **Primary Language:** Yoruba (`yo`)
- **License:** Apache 2.0
- **GitHub Repository:** [https://github.com/husseinalamutu/sovereign-k8s-llm](https://github.com/husseinalamutu/sovereign-k8s-llm)

## Intended Use
Designed for offline, sovereign inference on resource-constrained Kubernetes infrastructure when quantized to GGUF format and served via `llama.cpp`.