jtliang/decode
收藏资源简介:
DECODE是一个包含53.6K个真实IDE代码编辑的数据集,这些编辑是针对AI生成代码的修改,来自1000多名开发者在Python、TypeScript和JavaScript上的操作。与主要基于Git提交训练代码LLM不同,DECODE捕获了开发者对AI生成代码的中间手动编辑,这些编辑反映了Git历史完全忽略的现实编辑行为。这些编辑是对AI完成中不完美的响应,提供了关于开发者何时、为什么以及如何修改AI生成代码的细粒度信号。数据集中的每一行代表对AI完成的一个代码编辑,包含原始AI完成ID、编辑事件ID、时间戳、开发者ID、AI模型信息以及代码在不同粒度上的状态和编辑内容。
DECODE is a dataset of 53.6K real-world, in-IDE code edits of AI-generated code, collected from 1,000+ developers writing Python, TypeScript, and JavaScript. Unlike LLMs for code which are predominantly trained on Git commits, DECODE captures the intermediate manual edits developers make to AI-generated code, reflecting realistic editing behavior that Git history misses entirely. These edits occur in response to imperfections in AI completions and provide fine-grained signal on when, why, and how developers modify AI-generated code. Each row in the dataset represents a code edit made to an AI completion, containing fields such as the original AI completion ID, edit event ID, timestamps, developer ID, AI model information, and the state of the code at various granularities along with the edit content.




