README.md

---
license: apache-2.0
language:
- ar
- en
tags:
- llama3.1
- arabic
- 'pretrained '
- 'lora '
- peft
- AutoPeftModel
library_name: peft
pipeline_tag: text-generation
---

# 🚀 Arabic LLaMa 3.1 Lora Model (Version #1)


This fine-tuned model is based on the newly released LLaMA 3.1 model and has been specifically trained on the Arabic BigScience xP3 dataset. It is designed to respond to various types of questions in Arabic, leveraging the rich linguistic data provided by the [Arabic BigScience xP3](https://huggingface.co/datasets/M-A-D/Mixed-Arabic-Datasets-Repo/viewer/Ara--bigscience--xP3).

## Model Summary

- **Model Type:** Llama3.1 Lora Model
- **Language(s):** Arabic 
- **Base Model:** [unsloth/Meta-Llama-3.1-8B](https://huggingface.co/unsloth/Meta-Llama-3.1-8B)

## Model Details

- The model was fine-tuned in 4-bit precision using [unsloth](https://github.com/unslothai/unsloth) for 16k Step on 1 GPU


## I prepared for you a Gradio App to do inference with the model and compare its results with the base llama3.1 model

## <span style="color:Red">Note</span> 
just run the following code in colab: 

### Gradio APP (Colab T4 GPU is enough to run the app)
```python

!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install --no-deps "xformers<0.0.27" "trl<0.9.0" peft accelerate bitsandbytes
!pip install gradio

import gradio as gr
from unsloth import FastLanguageModel
import torch

# Load base model and tokenizer
base_model, base_tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Meta-Llama-3.1-8B",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(base_model)  # Enable native 2x faster inference

# Load LoRA model and tokenizer
lora_model, lora_tokenizer = FastLanguageModel.from_pretrained(
    model_name="Omartificial-Intelligence-Space/Arabic-llama3.1-lora-FT",  # Replace with your LoRA model path/name
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(lora_model)  # Enable native 2x faster inference

simplified_prompt = """Input: {}
Response: {}"""

def extract_response(text):
    """ Extracts the Response part from the generated text """
    response_marker = "Response:"
    if response_marker in text:
        return text.split(response_marker, 1)[1].strip()
    return text.strip()

def generate_responses(input_text):
    prompt = simplified_prompt.format(input_text, "")
    
    # Tokenize input for base model
    base_inputs = base_tokenizer([prompt], return_tensors="pt").to("cuda")
    # Generate output using base model
    base_outputs = base_model.generate(**base_inputs, max_new_tokens=128, use_cache=True)
    # Decode base model output
    base_decoded_outputs = base_tokenizer.batch_decode(base_outputs, skip_special_tokens=True)[0]
    base_response = extract_response(base_decoded_outputs)
    
    # Tokenize input for LoRA model
    lora_inputs = lora_tokenizer([prompt], return_tensors="pt").to("cuda")
    # Generate output using LoRA model
    lora_outputs = lora_model.generate(**lora_inputs, max_new_tokens=128, use_cache=True)
    # Decode LoRA model output
    lora_decoded_outputs = lora_tokenizer.batch_decode(lora_outputs, skip_special_tokens=True)[0]
    lora_response = extract_response(lora_decoded_outputs)
    
    return base_response, lora_response

# Custom CSS for the interface
css = """
h1 {
    color: #1E90FF;
    font-family: 'Arial', sans-serif;
    text-align: center;
    margin-bottom: 20px;
}

.description {
    color: #4682B4;
    font-family: 'Arial', sans-serif;
    text-align: center;
    font-size: 18px;
    margin-bottom: 20px;
}

.gradio-container {
    background-color: #F0F0F0;
    border-radius: 10px;
    padding: 20px;
}

.gr-button {
    background-color: #FFA500;
    color: white;
    border: none;
    padding: 10px 20px;
    text-align: center;
    display: inline-block;
    font-size: 16px;
    margin: 4px 2px;
    cursor: pointer;
}

.gr-button:hover {
    background-color: #FF8C00;
}

.gr-textbox {
    border: 2px solid #1E90FF;
    border-radius: 5px;
    padding: 10px;
}
"""

# JavaScript for additional functionality (if needed)
js = """
function createGradioAnimation() {
    var container = document.createElement('div');
    container.id = 'gradio-animation';
    container.style.fontSize = '2em';
    container.style.fontWeight = 'bold';
    container.style.textAlign = 'center';
    container.style.marginBottom = '20px';

    var text = 'Omartificial Intelligence Space';
    for (var i = 0; i < text.length; i++) {
        (function(i){
            setTimeout(function(){
                var letter = document.createElement('span');
                letter.style.opacity = '0';
                letter.style.transition = 'opacity 0.5s';
                letter.innerText = text[i];

                container.appendChild(letter);

                setTimeout(function() {
                    letter.style.opacity = '1';
                }, 50);
            }, i * 250);
        })(i);
    }

    var gradioContainer = document.querySelector('.gradio-container');
    gradioContainer.insertBefore(container, gradioContainer.firstChild);

    return 'Animation created';
}
"""

with gr.Blocks(css=css, js=js) as demo:
    gr.Markdown("<h1>Arabic llaMa3.1 Lora Model (Version 1)</h1>")
    gr.Markdown("<p class='description'>This model is the Arabic version of Llama3.1, utilized to answer in Arabic for different types of prompts.</p>")
    
    with gr.Row():
        input_text = gr.Textbox(lines=5, placeholder="Enter input text here...", elem_classes="gr-textbox")
        base_output = gr.Textbox(label="Base Model Output", elem_classes="gr-textbox")
        lora_output = gr.Textbox(label="LoRA Model Output", elem_classes="gr-textbox")

    generate_button = gr.Button("Generate Responses", elem_classes="gr-button")

    generate_button.click(generate_responses, inputs=input_text, outputs=[base_output, lora_output])

demo.launch(debug = True)

```

### Recommendations

- [unsloth](https://github.com/unslothai/unsloth) for finetuning models. You can get a 2x faster finetuned model which can be exported to any format or uploaded to Hugging Face.

## <span style="color:blue">Acknowledgments</span>

The author would like to thank Prince Sultan University for their invaluable support in this project. Their contributions and resources have been instrumental in the development and fine-tuning of these models.


```markdown
## Citation

If you use the Arabic llama3.1 Lora Model, please cite it as follows:

```bibtex
@model{nacar2024,
  author       = {Omer Nacar},
  title        = {Arabic llama3.1 Lora Model},
  year         = 2024,
  url          = {https://huggingface.co/Omartificial-Intelligence-Space/Arabic-llama3.1-Chat-lora},
  version      = {1.0.0},
}