Fine-tune it for multiclass or multilabel text classification

[lpschaub]

I have medical reports, and I try to predict the disease associated with each report :

1. both medical reports and disease to predict are written by humans -> mistakes, inconsistency in label names (same disease different ways to write it and reverse)
2. Should I use RobertaForSequenceClassifier or AutomodelForSequenceClassifier ?
3. what-s the best way to handle imperfect labels ? Embed them in the same roberta tokenized space, and predict the mean of the vector, or predict the whole vector (it becomes then a multilabel task).
   best