“Training a model entirely on a specific task's data yields the best performance for that task, provided there is a virtually infinite amount of high-quality data available.”