Surprisingly, however, the overall time to train the XML-based neural

neural-network:train

XML-based:
average time (net/ms) = 1711.572
total time (net/ms) = 169,445.644

map-based:
average time (net/ms) = 3633.811
total time (net/ms) = 359,747.295

I don't understand how this could possibly happen.

Does the profile not reveal where the map based solution spends most of
the time?

It might matter to judge that if the internal implementation of maps in
Saxon has changed in different versions.