How to evaluate the prompt score, not at the time of attack, but only?

[zl-comment]

Because the accuracy of my experimental results does not match the paper. Because I want to directly evaluate the prompt after being attacked by text to see if the score is wrong?