Hi Ravi and thanks for the reply. Basically, this is just a standard data science project, where I have a pipeline for preprocessing. At the end of this pipeline, I get numpy arrays x and y for the actual model training / cross validation, detached from the file format of the original input data. I thought this was a quite common approach, therefore I was surprised to not see something as basic as numpy included in kedro-datasets. While pickle works, .npy and .npz files would be the recommended way to store numpy arrays I guess. I implemented a very barebones custom dataset, but it does only support npy, not npz files (see snippet).