loading the model
The answer
The six likeliest intents
The words, and the arguments it pulled out of them
What it looked at
24 fields for this sentence, one per layer and head. Every row sums to one, so a bright row is a token that made up its mind and a flat row is one that did not. The first position is the cls token, and the intent is read from it and from the average of every position together, so every row here is part of the answer. All twenty four share one scale, so a faint field really is a faint one, and the colour is the square root of the value so the weak ones stay readable. The exact figure is on every label.