Using a codified pipeline for synthetic data generation helps to avoid biases in the training data