Language really only exists at the input and output surfaces of the models. In the middle it's all numerical values. Which you might be quick in relating to just being a numeric cypher of the words, which while not totally false, it misses that it is also a numeric cypher of anything. You can train a transformer on anything that you can assign tokens to.