The paper (improving language understanding with unsupervised learning) uses auxiliary objective in fine-tuning stage, which consists of two objective functions with labeled dat
