How many data points are necessary to build a reliable machine learning model (with mlc)? Is there a point where adding more data no longer significantly improves the model's performance?