The scary part isn't necessarily this dataset, but that unlabeled data causes a silent decrease in model performance -- which can be esp important for underrepresented classes.