Attention Viz
An interactive tool that lays out a transformer model's attention patterns by layer and head so they can be compared side by side. Because query and key vectors are projected into the same space, you can see at a glance that different heads produce quite different shapes. You come away seeing that attention heads each track different relationships — note that it assumes you already know how transformers are structured.
How to try it
In the matrix view, layers run down and heads run across; click one plot to zoom into that head, and switch between 2D and 3D or change the projection method and model. A search box highlights tokens, dot size and colour can be adjusted, and Documentation and Paper under 'about' explain how to read the plots. The interface is English only and, as the page warns, the data runs to many megabytes and may take up to a minute to render the first time.