Why is lr_finder on train instead of test?

If you read Leslie Smith’s paper, he uses the LR Range test with the validation / test loss. Why is it that the lr_finder for learns in fastai plots the training loss? I couldn’t seem to find an answer…

Thanks!