Kanzoet97/Japan

这是一个专为知识蒸馏任务设计的指令微调数据集，基于大规模推理语料库 facebook/natural_reasoning 的前10万个问题构建。数据集通过监督式微调（SFT）范式，将教师模型 gpt-oss-120-high 的高级、多步骤推理能力迁移至学生模型。数据集包含复杂推理问题及其参考答案，覆盖STEM、经济学、社会科学等多个领域，以其高质量、高难度和多样性著称。数据集还提供了教师模型生成的详细思维链（Chain-of-Thought）过程，支持思维链蒸馏，旨在提升模型在复杂问题上的推理能力。

This is a meticulously curated instruction fine-tuning dataset designed specifically for efficient knowledge distillation tasks. Built upon the first 100,000 questions from the large-scale reasoning corpus facebook/natural_reasoning, it aims to transfer the advanced, multi-step reasoning capabilities of the teacher model gpt-oss-120-high to a student model with high fidelity through a Supervised Fine-Tuning (SFT) paradigm. The dataset includes challenging reasoning questions and their reference answers, covering multiple domains such as STEM, economics, and social sciences, renowned for its high quality, difficulty, and diversity. It also provides the detailed Chain-of-Thought (CoT) reasoning process generated by the teacher model, enabling Chain-of-Thought distillation to enhance the models reasoning abilities on complex problems.