upiter/d3
收藏数据链接:
官方服务:
资源简介:
D3是一个大规模的数据集,包含用于训练语言模型按差分逐一生成和编辑Python代码的**指令+文件状态+差分序列**轨迹。每个轨迹将自然语言目标与源文件的初始内容和实现目标的有序差分序列配对。
D3 is a large-scale dataset of **instruction + file-state + diff-sequence** trajectories for training LMs to synthesize and edit Python code **diff-by-diff**. Each trajectory pairs a natural-language goal with the **initial contents of a source file** and an ordered sequence of atomic **file diffs** that realize the goal.
提供机构:
upiter


