I have fine-tuned a BERT model for a binary text classification task. The model achieves an F1 Score of ~0.9, but when plotting the Train and the Validation-Loss, the plot looks weird and I am not sure how to interpret this.
Here is the plot
What is this plot saying? Is there overfitting happening?
The best way to check overfitting is to also plot your classification performance on your validation set and see what happens. If it goes down, then it overfits. Otherwise, you may conclude that the NLL is not a very useful predictor of task-specific performance.
Assuming the validation set is large enough and well constructed, you may conclude that you are not overfitting on the training set (otherwise your classifier would get worse as you train) despite the loss indicating so.