引用wsxiaoys · 2023-03
“There’re recent works show a potential of beating Copilot’s performance (e.g RepoCoder) with much smaller models (500M vs 10B+).”
引用它:小得多的模型(500M 对 10B+)有望赶超 Copilot 的补全表现
检索与生成迭代交替,利用整个代码仓库的信息做代码补全。
针对仓库级代码补全中有用信息分散在多个文件的难题,提出 RepoCoder 框架:把基于相似度的检索器与预训练代码语言模型放进迭代的“检索—生成”流程,支持不同粒度的补全。论文同时提出 RepoEval 基准,涵盖行级、API 调用和函数体补全。实验显示它在各设置下比文件内补全基线提高超过 10%,并稳定优于普通的检索增强补全。
评论者把它当作“代码检索加 LLM”做仓库级补全的相关研究,并有人引它说明借助仓库级检索,小得多的模型也可能追平 Copilot 的补全表现。
“There’re recent works show a potential of beating Copilot’s performance (e.g RepoCoder) with much smaller models (500M vs 10B+).”
“Related research works include [1]. (Hints: combine code search and LLM).”
“For instance, RepoCoder [2] focuses on repository-level code completion by integrating code generation and retrieval models in an iterative paradigm.”