Methodology
Data Disclosure
The activation traces and circuit visualizations in this tool are synthetic. They are procedurally generated to mirror the topology and statistical properties of real transformer models—specifically GPT-2 small (12 layers, 12 attention heads, 768 hidden dimensions).
This means the visualizations demonstrate the structure and interaction patterns of mechanistic interpretability analysis, but do not reflect actual model activations from inference.
Why Synthetic Data?
- 01Running real mechanistic interpretability analyses requires significant compute (GPU inference) and model weights (~500MB for GPT-2 small).
- 02This demo focuses on the visualization and interaction patterns rather than live model inference.
- 03Synthetic data allows for a fully static deployment with no backend infrastructure.
What's Real
The Instruments
Circuit Diagram
Visualizes the flow of information through transformer layers as a 3D graph. Nodes represent attention heads and MLP neurons; edges represent weight connections. The ablation (patching) feature demonstrates causal analysis: toggling a node "off" shows how downstream predictions change.
Attention Explorer
Shows attention patterns between input tokens for a selected layer and head. The heatmap view displays the full source-target attention matrix; the beam view renders high-weight connections as curved tubes in 3D space.
Logit Lens
Applies the unembedding matrix at each intermediate layer to see what the model would predict if forced to output at that point. Demonstrates how predictions evolve from diffuse (early layers) to confident (late layers), with a "commit point" where the target token first dominates.