The process of creating an admissible set of functions that is invariant involves looking at properties of the training data, identifying general and special types of predicates and ensuring their values coincide with the expected values of the model.