When creating a (GBM or other algo) stage for a spark pipeline, the prediction col should be ignored by default.
This is fixed as part of . H2O correctly ignores response column if it's part of the features, however we in the pipelines, we were adding the response column to the list of existing features in all cases. That lead to behavior that H2O created a new column(because 2 columns can exists with the same name) and was used normally for training purposes which is unwanted behavior