As the current implementation stores all predictions and targets in RAM we need to explicitly detail such behaviour and note that this could lead to a memory error depending on the size of the input data.