Love this kind of experiment. Would the model perform better with word tokens?