Consider moving to a third-party library to handle datasets (e.g. the huggingface datasets library)

So we can drop the custom code from db.py, and also we should have improved performance.