3D model dissection
Follow one saved forward pass: open each layer, check residual additions, and inspect vocabulary scores.
8×8 only wraps 64 coordinates for display. It is not a semantic space.
8×8 only wraps 64 coordinates for display. It is not a semantic space.
Hatched future positions have exactly zero weight. Pick a bar to inspect one read.
This weighted V value is before head merging and output projection. Attention weights do not directly explain what the model “thinks.”
Softmax uses all 79 candidates. The five highest probabilities are shown here.
Reading guide and data source
Fixed pretrained weights · 92,352 parameters · Inspection only
8×8 only wraps 64 coordinates for display. It is not a semantic space. 79 vocabulary candidates, wrapped by ID. This is a different axis from the 64 internal coordinates.
Color shows sign and relative magnitude. No coordinate is assigned a meaning such as “fruit” or “grammar.” Each layer has its own color scale. The view stays fixed until you enable free rotation.
Saved pretrained forward trace · View the static lesson figure
Parameter digest
Preparing the 3D view…