I'm wondering whether using the en-target language parallel data or just the target language data when training the language-sft?