train alpaca based on meta opt-125m

see https://github.com/lunaczp/stanford_alpaca