NextWealth/Python-DPO

Python-DPO数据集是Python-DPO-Large数据集的较小版本，使用Argilla创建。数据集包含指令、选择的代码、被拒绝的代码和偏好信息。数据集的创建过程涉及先进的语言模型、严格的评估标准、人工参与和一系列工具，以确保代码的简洁性、边缘案例测试、错误处理和命名规范。源数据来自MBPP数据集和代码生成模型如Codegemma-7b-it和Magicoder-S-DS-6.7B。注释过程包括安全性、功能性和性能、相关性和完整性、风格和格式的评估，使用了pylama、pytest、bandit和性能分析工具。

The Python-DPO dataset is a smaller version of the Python-DPO-Large dataset, created using the Argilla tool. It contains instances of Python programming problems, each including an instruction, a chosen optimal code solution, two suboptimal rejected solutions, and a preference ranking. The dataset addresses key limitations in existing crowdsourced coding datasets by involving a human-in-the-loop team to evaluate code quality in terms of verbosity, edge cases, error handling, and naming conventions. The source data includes programming problem descriptions from the MBPP dataset and Python code solutions generated by models like Codegemma-7b-it and Magicoder-S-DS-6.7B. The annotation process uses a set of rubrics to evaluate code quality across security, functionality, relevance, and style, aided by tools like pylama, pytest, bandit, and profiling tools.