Foundational large language models (LLMs) can be instruction-tuned to develop
open-ended question-answering capability, facilitating applications such as the
creation of AI assistants. While such efforts are often carried out in a single
language, building on prior research, we empirically analyze cost-efficient
approaches of monolingual and multilingual tuning, shedding light on the
efficacy of LLMs in responding to queries across monolingual and multilingual
contexts. Our study employs the Alpaca dataset and machine translations of it
to form multilingual training data, which is then used to tune LLMs through
low-rank adaptation and full-parameter training. Comparisons reveal that
multilingual tuning is not crucial for an LLM's English performance, but is key
to its robustness in a multilingual environment. With a fixed budget, a
multilingual instruction-tuned model, merely trained on downsampled data, can
be as powerful as training monolingual models for each language. Our findings
serve as a guide for expanding language support through instruction tuning with
constrained computational resources.

通过使用多语言调优方法研究基础大型语言模型（LLMs）的代价效益，检验了 LLMs 对于单语和多语环境中查询的有效性，并发现多语调优对于 LLMs 在多语环境中的鲁棒性是关键。研究表明，在有限的计算资源情况下，仅使用有限数据集对多语调优模型进行训练，与为每种语言训练单语模型相比具有相同强大的性能。这些发现可作为扩展语言支持的指南，通过使用约束的计算资源进行指令调优。