Trained model tend to give too high a modified probability for each candidate base

[weir12]

Hi:
I've trained a model.To evaluate the performance of the model, I used unused experimental samples and control samples as test sets.
As a result, I find that the model tends to give higher modified probability for each candidate base.
What's worse,The control sample (without any modified base) was also incorrectly identified with a large number of modified bases.
If this is a hyperparameters problem for the model architecture,I wish someone would give me some advice.