Welcome to the forum @victoriamesa!
The 50% unassigned is a pretty clear indicator: your reads are probably in mixed orientations. The classify-sklearn method currently only supports classification of sequences that are in a single orientation. However, the classify-consensus-vsearch method can classify mixed orientation reads just fine. Could you give that method a try and let us know what you see? Thanks!
But @sixvable is also correct:
Even though this is probably not what is causing this classification issue, it is probably leading to suboptimal performance with q2-dada2, which assumes that reads are not yet merged. So demultiplex and denoise on the unpaired reads and let q2-dada2 do the merging for you.