The task of evaluating translation quality can be structured as a sequence-to-sequence text generation problem rather than a standard numerical scoring task. In this format, the source sequence includes a specific task instruction, such as "$[\mathrm{CLS}]$ Score the translation from English to Chinese. English: when in Rome, do as the Romans do. Chinese: 人在罗马就像罗马人一样做事。", and the model maps this to a target sequence containing the predicted score represented as generated text, like "$\langle s \rangle$ ${}0.81$".