llm-sandbox

A Python sandbox project to explore and experiment with text generation using Large Language Models (LLMs).

Repository Structure

The repository is structured as follows:

pytorch-project/
├── data/                   # Directory to store chat JSON files
├── src/
│   ├── __init__.py         # Shared code
│   ├── *.py                # Entrypoints for model demonstration
├── requirements*.txt       # Python dependencies
├── setup-venv.sh           # Setup script for the .venv

Prerequisites

• Python Version: Ensure Python version 3.10.x or 3.11.x is installed. (Python versions >= 3.13 are currently not compatible with PyTorch.)
• A performant GPU with Cuda-support is desirable, but you can fall back to using CPU if you want.

Installation

1. Clone this repository:

   $ git clone <repository-url>
   

2. Set up a virtual environment and install dependencies:

   $ bash setup-venv.sh
   

3. Create a .env file with Hugging Face credentials.

Hugging Face Credentials

This project uses the transformers python package for downloading and using pre-trained models. To ensure smooth functionality:

1. Obtain an access token from your Hugging Face account.

2. Create a .env file in the root directory with the following content:

   HF_TOKEN=<your access token>
   HF_HOME=.cache/
   

   Ensure that:

   • You have requested gated model access where required (e.g., for models like LLaMA).

How to Use

Once all dependencies are set up, you can run any of the demo scripts located in the src directory:

$ python src/<script-name>.py

Example:

$ python src/recurrentgemma.py

Note: The file __init__.py contains shared code and is not meant to be executed directly.

Data Handling

The data/ directory contains .json5 files, which store structured chat-like message interactions. The following schema is used for these files:

[
    {
        "role": "system" | "user" | "assistant",
        "content": "message text"
    },
    ...
]

Message Schema Details:

1. Roles:

   • system: Defines the model's behavior or initial instructions. Not every model supports system roles (e.g., src/gemma-*.py and src/recurrentgemma.py do not use these entries).
   • user: Represents user-supplied input for the model.
   • assistant: Represents the model-generated output in response to the user.

2. Content: The content field contains the text for each message, whether it's input or output.

Usage Guidelines:

• A chat session may optionally begin with a system role.
• The first non-system message must have the role user.
• Messages should alternate between the user and assistant roles.
• To use a demo, the final entry in the JSON file must be a user message, which serves as the input request for the model.

Chat Templates

Models in this project support chat templates to automatically transform chat-like instructions into structured data for processing.

Additional Notes

• Model Support:

  • Models behind gated access (e.g., LLaMA) require prior approval via the Hugging Face model repository.
  • Refer to the specific demo script for details on model compatibility and features.

• Errors: For Torch not compiled with ... or similar issues, ensure PyTorch is installed for your system configuration using this guide.

Further resources

huggingface model database