Hi @bsen2018,
Please check out this thread. The issue may be how you are training your classifier, and how your sequences have been processed. A very common issue with ITS classification is if sequencing adapters are still present in the reads.
The reference sequences are not being filtered out, so this does not affect the classifier training step. You only want to filter out query sequences with exclude-seqs.
This tutorial covers the commands that you want. exclude-seqs outputs both a hits.qza and misses.qza file... you will want to use the "hits".
Another option would be to not use exclude-seqs, and instead use the classify-consensus-blast or classify-consensus-vsearch methods in q2-feature-classifier with a perc-identity setting of 0.8 (or whatever you want to use). Any sequences with a lower % identity will be unclassified. That way you can first visualize what proportion of a sample is unclassified, before removing these.
I hope that helps!