Hi @SarahM,
I am wondering if the sequence orientations you are expecting is somehow wrong, and that may explain the primer matches you are seeing as well as why the sk-learn classify fails in your case!
What happen if you try to assign with blast (classify-consensus-blast: BLAST+ consensus taxonomy classifier — QIIME 2 2020.6.0 documentation), instead (or maybe just blast a couple of sequences to get the feeling)?
An alternative option it may be to try a close-reference approach (cluster-features-closed-reference: Closed-reference clustering of features. — QIIME 2 2020.6.0 documentation), to get the feeling of what is in your sample.
As an aside, you may want to look at the itsxpress plug in to trim off ITS flanking regions (Fungal ITSx and qiime2)!
Hope it helps!