The eval set was in the training data0.92 on the holdout, 0.71 in production. The holdout was the problem.