AutoML-org/SyntheticCode-800K

该数据集包含近80万个结构化的四元组，每个四元组包含编程任务、推理轨迹、有效的Python解决方案和单元测试。这些数据使机器学习模型能够学习如何解决问题，而不仅仅是解决方案的内容。数据集通过多阶段合成生成管道创建，确保了问题的多样性和代码的正确性。所有代码均为Python编写，且与标准基准（HumanEval, MBPP）无重叠，适合用于推理感知代码生成和微调的研究。

The dataset consists of nearly 800k structured quadruplets, each containing a programming task, reasoning trace, functionally correct Python solution, and unit tests. These data enable machine learning models to learn how to solve problems, not just what the solution is. The dataset was created through a multi-stage synthetic generation pipeline, ensuring diversity in problem-solving and correctness of code. All code is written in Python, and there is no overlap with standard benchmarks (HumanEval, MBPP), making it suitable for research on reasoning-aware code generation and fine-tuning.