Improving Fine-tuning Pre-trained Models on Small Source Code Datasets via Variational Information Bottleneck.