The following vizualization shows the F1 scores for each data set and percentage of labeled data used during fine-tune training. It is important to note that distillation training is self-supervised and did not use any labelled data.