Why Not Transform Chat Large Language Models to Non-English?

• DOI: 10.48550/arxiv.2405.13923
• Publication Date: 2024-05-22
• Authors: Xiang Geng; Ming Zhu; Jiahuan Li; Zhejian Lai; Wei Zou; Shuaijie She; Jiaxin Guo; Xiaofeng Zhao; Yinglu Li; Yuang Li; Chang Su; Yanqing Zhao; Min Zhang; Yang Hao; Xinglin Lyu; Jiajun Chen; Shujian Huang
• Keywords: Online chat

The scarcity of non-English data limits the development large language models (LLMs). Transforming English-centric LLMs to has been identified as an effective and resource-efficient method. Previous works start from base perform knowledge distillation (KD) with generated by stronger LLMs, e.g. GPT-4. Compared chat are further optimized for advanced abilities, multi-turn conversation human preference alignment, thus more powerful in both helpfulness safety. However, transforming a LLM involves two critical issues: (1) How can we effectively transfer abilities without their supervised data? (2) prevent original catastrophic forgetting during transformation? We target these issues introducing simple framework called TransLLM. For first issue, TransLLM divides problem into some common sub-tasks translation chain-of-thought, which uses bridge between English step-by-step. enhance performance publicly available data. second propose method comprising synergistic components: low-rank adaptation training maintain parameters, recovery KD, utilizes itself recover frozen parameters. In experiments, transform LLaMA-2-chat-7B Thai language. Our method, using only single-turn data, outperforms strong baselines ChatGPT on benchmark MT-bench. Furthermore, our safety rejects harmful queries AdvBench than