Thanks for the links !
Maybe someone can comment if this seems reasonable ? (It is going to take a while to digest those papers)...
I downloaded the AcousticBrainz data (including increments) and plotted the distribution of probabilities for "relaxed" songs and there was a sharp peak for the value 0.8. Like, way more than any other value.
Does that mean that for most songs the classifier provides the same probability value for having assigned the correct label ?