Not just for training data, but for eval data. If you can spend a few grand on really good labels for benchmarking your attempts at making something feasible work, that’s also super handy.