I would like to know if there is a statistical/empirical method to determine the proper number of samples for development/test set used for testing the generalized ability of a machine learning model. Traditionally, the split of 80% vs 20% (or 70% vs 30% for training and test sets respectively) is ok for the purpose. But for the large dataset, say 100,000 samples, it seems the little percent such as 1% is enough to estimate the developed model. Is that true? What is the empirical or theoretical rule to select proper number of sample for evaluating the models?