Attention begins with dot-product scores between query and key vectors. This lesson shows the displayed integer Q and K vectors, then the exact score grid they generate before any softmax appears.
highlighted = computed this step
Attention scores
Each query vector is compared with each key vector by a dot product. The displayed Q and K vectors are integer pairs, so every score is an exact integer. This is the exact part of attention: before masking or softmax, the score grid is just Q times K transposed by visible dot products.
Sij=Qi⋅Kj
Exact score rows
The recomputed score rows are [1,0,1], [0,1,1], and [1,1,2]. Each entry is recomputed from one query row and one key row, not copied from a hidden score table.
S=101011112
Summary
The score grid is still only integer dot products. The next step changes which scores are allowed to contribute, but it does not change the already-computed integer scores.