Hi @clairewill22,
Sorry to hear you're running into issues!
How are you training your own classifier? can you share your command? Make sure you use the min and max length parameters with appropriate thresholds... a common issue with training SILVA classifiers specifically is that junk sequences left in SILVA (e.g., with lots of ambiguous bases) can cause hits to disparate kingdoms, causing the classifier to get confused (search "hot spring metagenome" in the forum archive for some examples!).
Also keep your eyes peeled for when the latest QIIME 2 release comes out this week, and follow-ups... the newest pre-trained classifiers may relieve some of these headaches.
Finally, are the 27F/1492R primers included in your query (sanger) sequences? I'd be surprised if their inclusion is causing this issue, but it's worth looking if the above advice does not lead to a diagnosis.