A single book for pretraining boosts model performance by 'less than 0.06%.'