A small dataset has 200 samples. A practitioner uses a 50/50 train/test split. What is the main concern?