The LLM Transparency Tool (LLM-TT) provides an interactive, web-based interface for dissecting the internal mechanisms of Transformer-based language models. It is designed for researchers and practitioners seeking to understand model behavior, attention patterns, and neuron activations.

How It Works

LLM-TT leverages TransformerLens to create hooks into model layers, enabling detailed analysis of token contributions and representations. Users can visualize attention head contributions, explore neuron activations within Feed-Forward Networks (FFNs), and trace information flow through the model's layers. This approach allows for granular inspection of how specific tokens influence model outputs.

Quick Start & Requirements

• Docker:

  docker build -t llm_transparency_tool .
  docker run --rm -p 7860:7860 llm_transparency_tool
  

• Local Installation:

  git clone git@github.com:facebookresearch/llm-transparency-tool.git
  cd llm-transparency-tool
  conda env create --name llmtt -f env.yaml
  pip install -e .
  cd llm_transparency_tool/components/frontend
  yarn install
  yarn build
  

• Launch:

  streamlit run llm_transparency_tool/server/app.py -- config/local.json
  

• Prerequisites: Python 3.x, Conda, Yarn. Supports models already integrated with TransformerLens.
• Demo: https://huggingface.co/spaces/facebook/llm-transparency-tool-demo

Highlighted Details

• Interactive contribution graphs visualize token influence.
• Detailed views of attention head and FFN neuron activity.
• Supports models integrated with TransformerLens.
• Includes citation for related research papers.

Maintenance & Community

Developed by facebookresearch. Links to relevant research papers are provided for citation.

Licensing & Compatibility

Licensed under CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International). This license restricts commercial use and redistribution.

Limitations & Caveats

Adding support for models not already integrated with TransformerLens requires custom implementation of the TransparentLlm class and modifications to the Streamlit application. The CC BY-NC 4.0 license prohibits commercial use.