What if valid_loss is much less than train_loss?

Here is the result of the training:

74.00% [37/50 07:26<02:36]
epoch train_loss valid_loss exp_rmspe time
0 0.011315 0.003927 0.062056 00:12
1 0.017746 0.000746 0.026780 00:12
2 0.014916 0.001486 0.039329 00:12
3 0.015425 0.013706 0.109037 00:12
4 0.024731 0.006519 0.077436 00:11
5 0.026221 0.003306 0.059323 00:12

some epochs gave very low valid_loss but relatively high train_loss. Shall I pick or abandon them? I’m using following to train:

learn.fit_one_cycle(50, max_lr =1e-01,callbacks=[SaveModelCallback(learn,
  name='/content/gdrive/My Drive/goodmdl')])

Can we change the monitor to save an improvement on the combination of both valid_loss and train_loss?

Some data augmentations are not applied to the validation set (and won’t be applied during inference on the real data). It is also the case with layers like dropout. It could be a reason why validation loss is lower than the train loss.
If your validation set is good (big enough and with distribution similar to the real data), it is what you want to monitor and optimize for.