Training a model on poor quality data can cause it to pick up “bad habits” that don’t transfer well to real-world scenarios. So how do we quantify if our synthetic data is relevant and valuable?