T5 model fine-tuning in the stsb dataset generates wrong outputs

The generate space for the T5forconditionalgeneration is the whole vocabloary. That is, you may want a number, but the number is dealed as the string, i.e. ordinary text. You may get what you want after fine-tuned it well.Otherwise, you may use the T5 model and add a custom layer to convert the lasthiddenstate to the number.