In stage 1, researchers pre-train the cross-lingual MOSS-base model with public text and code corpora. In stage 2, they first perform supervised fine-tuning (SFT) with synthetic conversational data ...