Text this: Loss formulations for assumption-free neural inference of SDE coefficient functions