Welcome to the forum, @rosies!
Are you also trimming those sequences with extract-reads as you have shown? If not, it is not surprising that these are unclassified or misclassified, since you are training the classifier using the "wrong" training data. The query sequences must be trimmed in the same way as the reference (or trimmed to an internal site, e.g., you can use a full-length 18S classifier to classify trimmed 18S reads).