I would say that is uncertain. Model weights are always going to effectively be a huge collection of statistics about the training corpus. Unless you are envisioning artisanal, hand-crafted, free-range model weights where a person used a non-mathematical method to purposely and creatively choose each one?