Quick update: I think the problem is the same as the one reported in this issue. In fact, I am trying to train my models with just 72 events. When I try to use TMVA with a dataset containing more events, I do not have this problem.
I am using ROOT 6.26/06 (so it’s not really that old, and I cannot further update due to my OS), and I was wondering if the problem has been really fixed in ROOT 6.28.