Fine-Tuning Llama 2 | Generative AI Series

Fine-Tuning Llama 2 | Generative AI Series

Updated on July 25, 2024
Fine-Tuning Llama 2 | Generative AI Series header image

Introduction

Fine-tuning involves updating an AI model with new data to adapt the model to a specific task. You can fine-tune a model using:

  1. The retrieval augmented generation (RAG) technique. This method involves tapping real-time data sources like MySQL/PostgreSQL database to generate factual results.
  2. The labeled dataset technique: This method uses a smaller labeled task-specific dataset to optimize the model's performance.

Consider the Llama 2 language Models (LLMs) with over 70 billion parameters. The model is trained on vast volumes of data to perform different tasks. However, with time, the model's data can become outdated.

To make the LLama 2 model generate relevant results for queries, you can fine-tune the model with either the RAG or labeled dataset techniques. Each method has its pros and cons, as discussed in this article.

In this guide, you'll auto-tune the LLama 2 model with a new ad generation labeled dataset. You'll then push the new model to your Hugging Face account and use it to perform ad-generation tasks using the Vultr cloud GPU server.

Prerequisites

Before you begin:

Compare RAG With Labelled Datasets

The RAG technique is more applicable when using structured data while fine-tuning a model using labeled datasets works well with unstructured data. The following table highlights the key differences between the two.

RAG Versus Fine-tuning

Install the Required Dependencies

To fine-tune and use the LLama 2 model with Python, follow the steps below to install the required libraries:

  1. Install the huggingface_hub and autotrain-advanced libraries.

    console
    $ pip install huggingface_hub autotrain-advanced
    
  2. Add the libraries to the system path.

    console
    $ echo "export PATH="`python3 -m site --user-base`/bin:\$PATH"" >> ~/.bashrc
    source ~/.bashrc
    

Implement the AutoTrain Tool

The Hugging Face AutoTrain tool allows you to train natural language processing (NLP) models like Llama 2 for specific tasks. The tool accepts different custom parameters in a user-friendly format without spending too much time on the technical details.

In this section, you'll feed an ad copy generation dataset to the AutoTrain tool to fine-tune the meta-llama/Llama-2-7b-chat-hf model. Then, you'll upload the newly trained model to your Hugging Face account and later run the model using a Docker container to retrieve fine-tuned results.

Prepare the New Model's Dataset

The AutoTrain tool accepts the following specific data format for training the Llama 2 model:

xml
<s>[INST] <<SYS>>
 {{system message}}
<</SYS>>
{{message}} [/INST] {{answer}} </s>

Hugging Face provides a hub for various datasets that adhere to the above format. You can use these datasets to fine-tune your model. For instance, for this guide, you'll use the ad copy generation labeled dataset to fine-tune your model. Here are sample rows from the dataset.

xml
<s>[INST] <<SYS>> Create a text ad given the following product and description. <</SYS>> Product: Wide-leg trousers Description: Pants with a loose and flowing fit, creating an elongated and sophisticated silhouette. [/INST] Ad: Introducing our Wide-Leg Trousers: Effortlessly chic and endlessly versatile. Discover a flattering, comfortable fit that creates an elegant, elongated silhouette. Elevate your style today! </s>

...

<s>[INST] <<SYS>> Create a text ad given the following product and description. <</SYS>> Product: Electric Food Chopper Description: Chop ingredients effortlessly with an Electric Food Chopper - designed for quick and efficient meal prep. [/INST] Ad: Chop, prep, and conquer your kitchen tasks! 🌶️🌟 Simplify your meal preparation. Perfect for busy cooks and making cooking a breeze with a touch of electric food chopper efficiency! 🌟🍳🍛 </s>

Fine-tune the Llama 2 Model

Follow the steps below to launch and use the AutoTrain tool:

  1. Log in to your Hugging Face account.

    console
    $ huggingface-cli login
    
  2. Enter your Hugging Face access token and press Enter to proceed. Then, press N and Enter to skip adding your access token as a git credential.

    console
    ...
    Login successful
    
  3. Initialize the following custom environment variables. Replace YOUR_HUGGING_FACE_ACCESS_TOKEN and YOUR_HUGGING_USERNAME with the correct values. The smangrul/ad-copy-generation parameter points to the ad copy generation labeled dataset discussed earlier.

    console
    $ export model=meta-llama/Llama-2-7b-chat-hf
    export data_path=smangrul/ad-copy-generation
    export text_column=content
    export token=YOUR_HUGGING_FACE_ACCESS_TOKEN
    export project_name=demo-ads
    export repoid=YOUR_HUGGING_FACE_USERNAME/ads-llama2-demo
    
  4. Run the following command to train the Llama-2-7b-chat-hf model with the new dataset. The following command takes several minutes to complete.

    console
    $ autotrain llm 
      --train 
      --project_name $project_name 
      --model $model 
      --data_path $data_path 
      --text_column $text_column 
      --use_peft 
      --use_int4 
      --learning_rate 2e-4 
      --train_batch_size 2 
      --num_train_epochs 1 
      --trainer sft 
      --block_size 1024 
      --lora_r 16 
      --lora_alpha 32 
      --merge-adapter 
      --lora_dropout 0.045 
      --push-to-hub 
      --token $token 
      --repo-id $repoid
    

    Output:

    ...
    > INFO    Finished training, saving model...
    > INFO    Merging adapter weights...
    > INFO    Loading adapter...
    ...
  5. Visit your Hugging Face profile page to ensure you've trained and pushed the new model to the repository.

    https://huggingface.co/YOUR_HUGGING_FACE_USERNAME

    Output:

    Custom Llama Model

Serve the Custom Auto-trained Model

You've fine-tuned and pushed the model to your Hugging Face account. Follow the steps below to run the model in a Docker container:

  1. Initialize the following system variables. Replace YOUR_HUGGING_FACE_USERNAME and YOUR_HUGGING_FACE_TOKEN with the correct details.

    console
    $ export model=YOUR_HUGGING_FACE_USERNAME/ads-llama2-demo
    export volume=$PWD/data
    export token=YOUR_HUGGING_FACE_TOKEN
    
  2. Run the following Docker command to run a container that serves the new model.

    console
    $ sudo docker run -d 
    --name hf-tgi 
    --runtime=nvidia 
    --gpus all 
    -e HUGGING_FACE_HUB_TOKEN=$token 
    -p 8080:80 
    -v $volume:/data 
    ghcr.io/huggingface/text-generation-inference:1.1.0 
    --model-id $model 
    --max-input-length 2048 
    --max-total-tokens 4096
    

    Output:

    1.1.0: Pulling from huggingface/text-generation-inference
    ...  
    Status: Downloaded newer image for ghcr.io/huggingface/text-generation-inference:1.1.0   
  3. Check the Docker logs to monitor the container as it loads the new model.

    console
    $ sudo docker logs -f hf-tgi
    

    Output:

    ...
    Warming up model
    Setting max batch total tokens to .....
    ...Connected
    ...Invalid hostname, defaulting to 0.0.0.0
  4. Allow port 8890 through the firewall. You need this port to run a Jupyter lab instance.

    console
    $ sudo ufw allow 8890
    $ sudo ufw reload
    
  5. Run the Jupyter lab instance to retrieve an access token.

    console
    $ jupyter lab --ip 0.0.0.0 --port 8890
    

    Output:

    http://YOUR_SERVER_HOST_NAME:8890/lab?token=b7ab2bdscb366edsddssfsff0faeb5fa68b6b0cf
  6. Access the Jupyter Notebook on your browser.

     http://YOUR_SERVER_HOST_NAME:8890/lab?token=b7ab2bdscb366edsddssfsff0faeb5fa68b6b0cf
  7. Click Python 3 ipykernel under Notebook. Then, enter the following Python code.

    python
    from huggingface_hub import InferenceClient
    
    URI   =  'http://127.0.0.1:8080'
    client = InferenceClient(model = URI)
    
    def chat_completion(system_prompt, user_prompt, length = 1000):
      final_prompt = f"""<s>[INST]<<SYS>>
                      {system_prompt}<</SYS>> {user_prompt} 
                      [/INST]
                      """
    
      return client.text_generation(prompt = final_prompt,max_new_tokens = length).strip()
    
    system_prompt = """
                    Create a text ad given the following product and description.
                    """
    
    user_prompt   = """
                    Product: Lightweight Bicycle Description: Perfect bike for any road conditions. this bike comes with durable tires, LCD display, and modern disc brakes
                    """
    print(chat_completion(system_prompt, user_prompt))
    
  8. Run the above Python code to generate an ad based on your input.

    Output:

    🚴‍♂️🌟 Experience the thrill of cycling with a Lightweight Bicycle! 🌟🚴‍♂️ Perfect for road adventures, this bike boasts durable tires, LCD display, and modern disc brakes. Limited stock - ride in style! 🚴‍♂️🌟

Conclusion

In this guide, you've explored the difference between fine-tuning a model using RAG and labeled datasets to improve the model's accuracy. Then, you've implemented the HuggingFace AutoTrain tool to fine-tune the model with a new dataset. Finally, you've run the fine-tuned model in a Docker container to generate an ad based on a custom product description.