cross-validation score is always lower than manual validatoin score

Question

I am trying to train 1.5k size training data with around 50ish features. After data preprocessing, I tried to train with various regressors and found GBR has the best score,