Adding soft prompts into the pre-training stage to obtain a better initialization.can reach or even outperform full-model fine-tuning under both full-data and few-shot settings