CodeFlow
收藏资源简介:
CodeFlow数据集是由清华大学和微软研究院共同构建的,通过迭代调试过程生成的函数级代码片段和测试用例,不断迭代直到代码通过所有单元测试。数据集包含正确版本和之前失败版本的对比,形成偏好对,用于指导代码生成模型学习更精确的错误修正模式。该数据集的构建模仿了人类程序员如何通过不断修改代码来调试程序,旨在帮助代码生成模型学习识别关键错误并提高代码质量。
The CodeFlow dataset was jointly developed by Tsinghua University and Microsoft Research. It is constructed via iterative debugging workflows, generating function-level code snippets and test cases, with the iteration process continuing until the code passes all unit tests. The dataset includes comparisons between correct code versions and their previously failed counterparts, forming preference pairs to guide code generation models in learning more precise error correction patterns. The construction of this dataset mimics how human programmers debug programs through continuous code modifications, aiming to help code generation models learn to identify critical errors and improve code quality.

- 1IterPref: Focal Preference Learning for Code Generation via Iterative Debugging清华大学 · 2025年



