A guide to using the Mechanistic Interpretability Visualization Tool
🔍 Exploring the Architecture
The visualization shows transformer layers stacked vertically. Each layer contains attention heads
(circles) and an MLP block. Click any component to select it.
📝 Adding Annotations
Select a component and use the right panel to add notes, tags, and importance ratings. Annotations
are saved to your browser's localStorage automatically.
🎨 Color Coding
Red = High importance, Amber = Medium, Green = Low, Gray = Unknown. Components with notes show a
white dot indicator.
⚙️ Configuration
Click the gear icon to change model architecture. Choose from presets (GPT-2 Small/Medium/Large/XL)
or configure custom layer and head counts.
📤 Export/Import
Export annotations as JSON for backup or sharing. Import previously exported annotations to restore
your work.