Oh, i think this is quite relatable

You're most likely overfitting on the train data provided

let's say you probably removed all the missing values in the data, you might be doing well locally on CV, but on the provided test data you might have lost most of the correlated patterns you needed to make good predictions

Your score is from your training data, the public score is generated from the label of the test data they don't provide, it is used to check how well your model generalize on unseen data, what the score is telling you is that your model is over fitting on the training data