Building a model eval to learn what meaningful training data takes

The user argues that building an eval can teach what it takes to generate truly meaningful training data, rather than leaving someone with only a surface-level understanding of model performance.