We are told in the first video of course 2 that the dev and test set should get their data from the same place. Does that mean they strictly have to get their data from the same place regardless of where the train gets its data, or that all three sets must have the same source of their data?
All three sets must come from the same distribution for best performance of the model.