When-Does-Reasoning-Matter/general-reasoning-ift-pairs

IFT与推理配对数据集是一个基于Infinity-Instruct的7M_core子集的广泛且高质量的数据集，包含大约90万个查询的查询-答案对。每个查询都生成了两种格式的答案：简洁的直接响应（IFT答案）和启用推理模式下的响应（推理答案）。该数据集用于研究推理对模型性能提升的贡献。

The IFT & Reasoning Paired Dataset is based on the 7M_core subset of the Infinity-Instruct dataset, providing a large-scale and high-quality collection of approximately 900k query-answer pairs. For each query, two types of answers are generated: a concise direct response (IFT Answer) and a response with reasoning mode enabled (Reasoning Answer). This dataset is used to study the contribution of reasoning to model performance improvement.