LLM Visualizer is a web-based tool that visualizes the internal attention mechanisms of large language models (LLMs). This interactive interface allows users to explore how transformer models process and focus on different tokens during inference.

• Attention Head Visualization: Explore how each attention head operates across layers.
• Token Interactions: Visualize the attention distribution between different input tokens.
• Layer-wise Exploration: Analyze attention patterns at each layer of the transformer.
• Custom Models: Supports custom transformer models using Hugging Face’s transformers library.

To install and run the project locally:

1. Clone the repository:

   git clone https://github.com/pierizvi/llm-visualizer.git
   cd llm-visualizer

2. Create and activate a virtual environment:

   python3 -m venv venv
   source venv/bin/activate  # For Linux/macOS
   venv\Scripts\activate  # For Windows

3. Install the required dependencies:

   pip install -r requirements.txt

4. Run the application:

   python app.py

5. Access the application by llm-visulaizer:

   npm start

1. Load a Pre-trained Model: Select any Hugging Face transformer model or load your own fine-tuned model.
2. Input Text: Enter a sentence or a passage to visualize the attention heads and layers for each token.
3. Interactive Exploration: Click through layers and heads to explore attention patterns and token dependencies.

• app.py: Flask-based web server for running the application.
• static/: Frontend assets such as JavaScript, CSS, and images.
• templates/: HTML templates for the web interface.
• visualization.py: Core logic for handling model inference and generating visualizations.
• models/: Code related to loading and processing transformer models.

• Python 3.7+
• Flask
• Hugging Face transformers
• Matplotlib
• Plotly

To install additional dependencies:

pip install transformers flask torch matplotlib plotly

Contributions are welcome! Feel free to submit issues and pull requests to enhance the project.

This project is licensed under the MIT License.