Certain tokens in the vocabulary cannot be decoded to valid utf-8, I'm actually not sure if this is because they represent partial utf codepoints, but in any case they cause generation to fail.