A training sample for a translation task illustrates the sequence-to-sequence structure by distinguishing the source text from the target text. For instance, in a task translating from Chinese to English, a training sample can be expressed as: $[\mathrm{CLS}]\ \textrm{Translate from Chinese to English: 你好！} \to \langle s \rangle\ \textrm{Hello!}$. Here, the arrow ($\to$) separates the source text, which encompasses the task instruction and the input, from the target text, representing the system's response. Additionally, $[\mathrm{CLS}]$ and $\langle s \rangle$ designate the start symbols on the source and target sides, respectively.