Some experimental Implementations of LLama3 8B designed to leverage personal Research in RAG, Reinforcement Learning and collection of Conversational Data.
Please find below a brief description of the current implementations:
-
Basic Chatbot
Quantized (4bit) Checkpoint of Llama3 8B model built with Llama.cpp. This checkpoint is not finetuned in any specific dataset, and doesn't perform RAG. It excels in interactive general domain conversations. -
Book Shop Assistant -> Chatbot + LLMrouter + RAG
Here I built a RAG on top of the above example, using the same model as a Router to decide on either performing a RAG (Search in database, Retrieve Documents, Generating with the help of additional information) or a normal reply. For the Search, I'm generating a Search Query with the same LLama3 model using the conversational context and performing the search with the library usearch. The database that this chatbot has access is a subset (books) of the Amazon Reviews Dataset (2023), so it is designed to provide books suggestions and help an user in the process of buying a book by delivering grounded information about reviews, plots, pricing, images of covers and etc. -
Chatbot with Internet Access (In Dev)
An internet-enabled chatbot capable of searching the web for information and reference it's answers. -
Chatbot that improves with time RL (In Dev)
An implementation of RLHF tailored with personalized rewards.
To help with the experiments, follows a basic User Interface, a multi-page streamlit app containing all sample chatbot use cases.
You can access this app through this link: link
# Clone this repository using your terminal
$ git clone https://github.com/Joaoluislins/llm-agents.git
# Branch your development
$ git checkout -b <branch_name>
# Navigate to the repository
$ cd demo
# Create a .env file with your environment paths and variables
# Include your Huggingface api token to access some models
$ echo "HF_TOKEN='your_huggingface_api_token'" > .env
# Define a default home path for the repository, follows an example, please replace the your_user_folder with your server username
$ echo "DEMOBOT_HOME='/home/your_user_folder/demo'" >> .env
# Add more variables as needed
# Create a conda environment
$ conda create -n your_new_env_name python=3.10 -y
$ conda activate your_new_env_name
# Install the llama.cpp python package or compile it from source, I would suggest the second option to have control over the compilation flags, gpu support and other settable options.
$ pip install llama-cpp-python
# Install additional required packages
$ pip install -r requirements.txt
# Finally, you can run the streamlit from your local and start the application
$ cd demo # if your are not here already
$ streamlit run Home.py --server.port same_available_port_number
# You've done it! Go to your browser and test the application in http://localhost:available_port_number/