The best way to learn programming is to build from scratch on top of a very simple hello world program. The best way to learn machine learning is to start from one of the most simple toy datasets which is similar to the real world problem. I organize the directories of this repository by dataset names. Each datasets may represents more than one problem, while each problem may have multiple approaches to solve, therefore each directory may have more than one Jupyter Notebooks files.
Pupular machine learning Python libraries usually provide API calls to download toy datasets. Here are some libraries containing useful toy-datasets that I found useful: