One thing to consider is the option to save various types of predictions as well as, or instead of the model. I think think of a few use cases: - [ ] Save out-of-fold predictions for ensembles. - [ ] Save predictions for Kaggle test set.