Multilabel t5 generates labels not in input

[NienkeMekkes]

I have a training set consisting of sentences, and each sentence has one or more labels. Very similar to the "toxic tweet" dataset. When I follow the instructions, some of my predictions are very strange new labels. An example would be a sentence getting the label 'rude', while 'rude' is not in my label options at all.

How do I prevent this behavior? I want the model to only pick from the label set I feed it, since the new ones are very strange.