They don't. Generally we already have trillions tokens worth of classifier datasets so there's no reason to create synthetic data that will be worse than the data it was trained on.