about the accuracy in BFCL irrelevance test

Hi, thanks for your great work!

While Recall has sufficiently improved the multi-turn accuracy of our model, the irrelevance test score is close to zero and the overall score has dropped. Is there any way to mitigate this problem?