Because it
is in their training set but it's unrealistic to expect a 2B or 4B model to be able to perfectly reproduce everything it's seen before.
The training no doubt contributed to their ability to (very) loosely approximate an SVG of pelican on a bicycle, though.
Frankly I'm impressed