---

Introduction

MAV - Model Activity Visualiser (for LLMs)

Getting started

METHOD 1: If uv is installed:

uv run --with openmav mav

or

uv run --with git+https://github.com/attentionmech/mav mav --model gpt2 --prompt "hello mello"

METHOD 2: Without uv:

1. Set up and activate a virtual environment

2. Install the package:

   pip install openmav
   

   or

   pip install git+https://github.com/attentionmech/mav
   

3. Run:

   mav --model gpt2 --prompt "hello mello"
   

4. or Import

   from openmav.mav import MAV
   
   MAV("gpt2", "Hello")
   

METHOD 3: Locally from scratch

1. git clone https://github.com/attentionmech/mav
2. cd mav
3. Set up and activate a virtual environment
4. Install the package:

   pip install .
   

5. Run:

   mav --model gpt2 --prompt "hello mello"
   

METHOD 4: Inside Jupyter notebook/Colab

You can replace gpt2 with other Hugging Face models for example:

• meta-llama/Llama-3.2-1B
• HuggingFaceTB/SmolLM-135M
• gpt2-medium
• gpt2-large

Tutorials

Writing your custom plugin tutorial in colab

writing custom plugin panel

running MAV with a training loop with a custom model (not pretrained one)

uv run examples/test_vis_train_loop.py

running MAV with custom panel selection and arrangement

uv run --with git+https://github.com/attentionmech/mav mav --model gpt2 --num-grid-rows 3 --selected-panels generated_text attention_entropy top_predictions --max-bar-length 20 --refresh-rate 0 --max-new-tokens 10000

Demos

• Basic plugins
• Entropy Fire plugin
• interactive mode
• limit chars
• sample with temperature
• running with custom model
• panel selection
• running in colab notebook

Note: explore it using the command line help as well, since many sampling params are exposed.

Contributing

Clone the repository and install the package in development mode:

git clone https://github.com/attentionmech/mav
cd mav

# recommended
uv sync

# if you don't use uv
pip install -e .