Hi @tengfey,
This error is occurring because you only have 3 samples per category but the default cross-validation is 5-fold (meaning that the input samples are split into 5 subsets for model building and validation), so this can’t be done. You can change this number by using the option --p-cv.
It looks like you have a total of 6 samples. That is going to be entirely insufficient for any kind of learning model. The methods in q2-sample-classifier will not give meaningful results if you have so few samples. As a rule of thumb I would recommend at least ~50 samples before trying to use any of these methods.