README.md

---
license: bigscience-openrail-m
language:
- yo
base_model:
- Jacaranda/YorubaLlama
metrics:
- cer
---

# YORI-LLaMA-Quantized

**YORI-LLaMA-Quantized** is a **Yoruba large language model** specialized in **text generation** and **AI-based assistance** in Yoruba.
It was **quantized** from its base model , [`Jacaranda/yorubaLLaMA`](https://huggingface.co/Jacaranda/yorubaLLaMA), to achieve lighter memory usage and faster inference while maintaining strong linguistic performance.

---
## 💬 Description
YORI-LLaMA-Quantized is trained to **generate and understand natural Yoruba text** with contextual fluency and syntactic awareness.
It can be used for:

- Yoruba **AI assistants**
- **fine-tuned for Chatbot**
- **Text completion**, storytelling, and creative generation
- **Language preservation** and computational linguistics research

![yori](https://cdn-uploads.huggingface.co/production/uploads/684c3d70f57a8cc5970ae3a9/-cf7gcdMjYd-j11XLIyEg.png)

## ⚠️ Limitations

- **Code-switching weakness:** The model struggles when Yoruba and English are mixed, leading to misinterpretations or incorrect tokens.
- **Numerical inaccuracies:** Occasionally produces factual errors, e.g., reporting *mẹ́ẹ̀dógún (15)* instead of *mẹ́rìndínlógójì (36)* for the number of Nigerian states.
- **Ambiguous prompts:** May output irrelevant or nonsensical responses when the query lacks clarity.

These limitations suggest the need for **more diverse and updated training data** across dialects and domains.

---

## ⚙️ Inference Notice
A key consideration when running inference is **model precision**.
YORI was quantized for efficiency, but inference should be performed in **FP16 precision** for stability and performance.

Example snippet:
![FP16 Snippet](https://cdn-uploads.huggingface.co/production/uploads/684c3d70f57a8cc5970ae3a9/kbqyQzQ3zhYRmjR0Ha1xz.png)

---
## 🧩 Intended Use

This model is intended for:
- **Academic and research purposes**
- **Educational AI assistants**
- **Yoruba language technology development**

Do not use this model for:
- Disinformation or impersonation
- Generating offensive or harmful content
- Applications violating user consent or privacy

---