Attention pattern atlas

What it does

Mechanistic interpretability research has produced a growing taxonomy of attention head “circuits” — patterns of behavior that recur across models. This atlas catalogues them with concrete examples you can inspect: which head, which layer, which input, and what the pattern means.

What you’ll see

• Browsable gallery of attention patterns, grouped by circuit type
• Side-by-side comparison across Llama, Mistral, Qwen, and Gemma at matched scales
• Layer-by-layer evolution: watch a circuit emerge across depth
• Direct links to the source papers and reproducible code

Status

Planned — data collection is underway. If you want to contribute patterns from a specific model, open an issue.