(논문 요약) Chat Vector: A Simple Approach to Equip LLMs with Instruction Following and Model Alignment in New Languages (Paper)

핵심 내용

• chat 기능을 weight 의 vector addition 으로 구현.
  • CP: continual pretrain
  • SFT: supervised fine-tuning
  • RLHF: Reinforcement learning from human feedback
• prompts

실험 결과

• 0.5 곱한건 heuristic 인 듯함.