That is why I suggested this approach was better ( note the that this is where()[0] instead of just where() as it was in my first example):

x,i=numpy.unique(y, return_inverse=True)
f=[numpy.where(i==ind)[0] for ind in range(len(x))]

type(f)     --> list

type(f[0]) --> ndarray

type(f[0][0]) is meaningless since it is just a single element in an array.  It must be an int type of some sort of since indices have to be int types.  x will be the same dtype as your input array.

You could conceivably change the type of f[0] to a list, but why would you want to?  One of the big advantages of python is that usually it doesn't matter what the type is.  In this case, a numpy ndarray will work the same as a list in most cases where you would want to use these sorts of indices. It is possibly to change the ndarray to a list, but unless there is a specific reason you need to use lists so then it is better not to. 

You cannot change the list to an ndarray because the elements of the list are different lengths.  ndarray doesn't support that.