Like the other commenter said, I'm sure you could create models that give probabilities of high score vs low score. But how would you interpret a 50:50 probability vs 30:70. I cannot see how to convert these probabilities into scores reliably. For instance compare a model training on 20 rows for maths but those scores are in reality ~80%, compare this to 20 rows for maths with normal distribution above and below 80%. Your model would predict the same scores even though the original scores are vastly different.

I feel like this may help you what-are-the-consequences-of-not-freezing-layers-in-transfer‌​-learning