Currently, get_dataset_splits() in our datasets is a static method (@staticmethod), but it would be more appropriate to have it marked as a class method (@classmethod).

The following example shows the difference between these two:

class A:
    @staticmethod
    def x():
         A.y()

    @staticmethod
    def y():
        print('Ay')

class B(A):
    @staticmethod
    def y():
        print('By')

B.x() prints Ay

class A:
    @classmethod
    def x(cls):
         cls.y()

    @staticmethod
    def y():
        print('Ay')

class B(A):
    @staticmethod
    def y():
        print('By')

B.x() prints By

In our datasets, x corresponds to get_dataset_splits and y corresponds to get_default_fields. The issue is that our current pattern doesn't support overriding get_default_fields by a subclass as can be seen in the example above (replace print with return fields), without overriding get_dataset_splits as well. We have this scenario in the SNLI dataset so this is the main reason why I'm opening this issue.