A question about the performance of opt-13b in SST2, Copa and Wino datasets.

[Corsilent]

Thank you for your open-source contributions! I have a few questions about the performance reported in the paper. Is the q value always set to 1? If so, it is not intuitive that opt-13B model can reach such high accuracy in those tasks by using ZO optimizers. Could you please provide the specific experimental parameters in Fig. 1?