Reviewing PR #539 which tests dataset uploading, I found four things which we could improve to make dataset uploading nicer for the user:

1. There should be no necessity to provide the format as this will always be 'arff'.
2. The current example encodes the class label as a REAL, although it should actually be a categorical.
3. We could add an example for uploading a list of lists, which could then also contain strings and categories.
4. The function create_dataset should be imported into the openml.datasets namespace.

@ArlindKadra @glemaitre