How to improve our models more and more without struggling to label a lot of data? it is very expensive and time-consuming to make our own training set to develop perception scores. Also, that is risky since the data labeled ourselves may not be guaranteed for model to get smarter than before.