Tasks should be handled by the model, not dataset.

[HesitantlyHuman]

Inferring tasks is dependent on model architecture in many cases. It would make more sense to allow models to do this step themselves and limit the datasets to returning ids, data and targets.

[HesitantlyHuman]

There are a couple of things that I would like to think about before making this decision. I like the discoverability that the tasks currently give. (.summary() for example)

Maybe we just have the dataset report on the target shapes? Should we also require that targets be the same shape for each target?