Generates synthetic datasets for training and evaluating vision models on object sorting and rule-based reasoning tasks. Each sample contains multiple objects of the same shape but different sizes ...