How do you know it really does not exist? There is this effect called 'over-fitting', where you might expect an exact sample from the training set to show up in the video