Evaluating Instruction-Tuned Large Language Models on Code Comprehension and Generation
Authors:
Zhiqiang Yuan,
Junwei Liu,
Qiancheng Zi,
Mingwei Liu,
Xin Peng,
Yiling Lou
Abstract:
In this work, we evaluate 10 open-source instructed LLMs on four representative code comprehension and generation tasks. We have the following main findings. First, for the zero-shot setting, instructed LLMs are very competitive on code comprehension and generation tasks and sometimes even better than small SOTA models specifically fine-tuned on each downstream task. We also find that larger instr…
▽ More
In this work, we evaluate 10 open-source instructed LLMs on four representative code comprehension and generation tasks. We have the following main findings. First, for the zero-shot setting, instructed LLMs are very competitive on code comprehension and generation tasks and sometimes even better than small SOTA models specifically fine-tuned on each downstream task. We also find that larger instructed LLMs are not always better on code-related tasks. Second, for the few-shot setting, we find that adding demonstration examples substantially helps instructed LLMs perform better on most code comprehension and generation tasks; however, the examples would sometimes induce unstable or even worse performance. Furthermore, we find widely-used BM25-based shot selection strategy significantly outperforms the basic random selection or fixed selection only on generation problems. Third, for the fine-tuning setting, we find that fine-tuning could further improve the model performance on downstream code comprehension and generation tasks compared to the zero-shot/one-shot performance. In addition, after being fine-tuned on the same downstream task dataset, instructed LLMs outperform both the small SOTA models and similar-scaled LLMs without instruction tuning. Based on our findings, we further present practical implications on model and usage recommendation, performance and cost trade-offs, and future direction.
△ Less
Submitted 2 August, 2023;
originally announced August 2023.
