Sweaterdog/GRaPE-Thinking-Mix

该数据集是基于200万段对话构建的，包含代码、数学和一般推理类型的对话，以及少量不必要的非思考示例。这些示例旨在教模型如GRaPE在思考方面的最佳实践。数据集中最长token长度为33,000。对于代码评估，包含两种消息，一种用于生成代码，另一种用于验证。这能教会模型如何审查自己的工作，并在代码运行之前发现问题的能力。数据集分为不同推理难度级别，包括中等、高、低、最小和自动，样本数量分别为1,201,021、483,332、261,802、92,854和1,018,704。

Built with 2M conversations with Code, Math, and General reasoning, as well as a small sample of non-thinking examples where it is not necessary, this teaches models such as GRaPE the best practices when it comes to thinking. The longest token length in this dataset is 33,000 tokens. For Code evaluations, there are two messaages, one asking to generate the code, and one asking for validation. This will teach the model how to review its own work, and figure out problems before code is ever run. The dataset is split into different reasoning difficulty levels, including Medium, High, Low, Minimal, and Auto, with 1,201,021, 483,332, 261,802, 92,854, and 1,018,704 examples respectively.