GBM early stopping is computed on training set logloss
The early stopping metric logloss seems to be computed on the training frame and not on the validation frame
Here is a snipped of the code
I noticed that looking at the logloss plot in the H2O flow interface the decease in the validation set (orange line) is less than the threshold but the tree generation does not stop after the chosen number of rounds, see picture attached.
Please, don't mind the clearly overfitting model