Relative encodings enable models to become evaluated for lengthier sequences than Those people on which it absolutely was trained.