The model lacks end-to-end training data, and its generated poems and style-transfer dataset can be poorly matched.