I created a training, a validation and a test set for an image classification task. Then, I did training using the training and did evaluation on validation set. So, the next step is to evaluate the test set, basically it is for inference. To select the best epoch model, I normally check the validation loss. However, I’m not sure if this way is correct.
I add the results as a picture and I marked the best train/val accuracy and the lowest train/val loss. The lost function is the Cross Entrophy.

The later steps are clearly overfitting, but still the training loss gets lower and lower till the 25th epoch. If I would only check the val loss, I might select the 13th epoch. Or the highest validation accuracy could also be important.

I read some articles or forum questions like this, this, and this, and several others, but there is no real conclusion on this topic.