about experimental setting

[KennyNH]

Thanks for your wonderful work "Which Examples to Annotate for In-Context Learning? Towards Effective and Efficient Selection".

My question is that should we ensure the class balance for the demonstrations when doing in-context learning on sentiment classification task?

If we just select k-nearest example samples for the test sample, how to ensure GPT knows all possible answers (i.e., sentiment labels) and would there exist class bias issue? e.g. in case that k-nearest example samples are all positive but the test sample is actually negative.