If I have a set of data and I am not sure about the data at all or the meaning I would try to investigate the dataset a bit first. I understand the importance of having a training-validation-test split when it comes to evaluating model performance but should this also be applied before when I am performing EDA for example?
I assume no, because I want to know that the data in my respective splits is correct, doesn't contain errors etc. Is there a point however when I am performing these checks, i.e. distribtion checks of features etc. that I can inadverntanly include bias into my model by not having the split until the data is prepared?