JohnsonPedia/yoruba_llama_8B_v2-lora-adapter

Yoruba LLaMA 8B v2 - LoRA Adapter

Fine-tuned Yoruba language adapter for LLaMA 3 8B by Johnson Pedia (Oduduwa AI).

🚀 Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3-8B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto"
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct")

# Load Yoruba adapter
model = PeftModel.from_pretrained(base_model, "JohnsonPedia/yoruba_llama_8B_v2-lora-adapter")

# Chat!
prompt = "Bawo ni o ṣe wa?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

💻 Works on CPU!

# CPU inference (no GPU needed)
base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3-8B-Instruct",
    torch_dtype=torch.float16,
    device_map="cpu"  # ← CPU only
)

📊 Model Details

• Base: LLaMA 3 8B Instruct
• Language: Yoruba (yo)
• Training: LoRA (r=16, alpha=16)
• Size: ~50MB adapter
• Use: Yoruba chatbot, Q&A, translation

🎯 What It Does

• Answers questions in Yoruba
• Uses proper diacritics (à, ẹ, ọ, ṣ, gb)
• Culturally appropriate responses
• Natural Yoruba conversations

💾 Requirements

• RAM: 16GB (full precision) or 10GB (8-bit)
• Disk: 20GB
• GPU: Optional (works on CPU!)

🙏 Credits

Built with Unsloth • Based on Meta LLaMA 3 • By Johnson Pedia