s1K 推理问题数据集

s1K 数据集是一个由李飞飞团队于 2025 年发布的高质量推理数据集，包含 1k 个问题及其详细的推理轨迹和答案，这些答案来源于 Google 的 Gemini Thinking Experimental 蒸馏结果。数据集涵盖 50 个不同领域，包括概率论、量化面试题、奥赛题等，确保模型能够处理多种类型的推理任务。相关论文成果为「s1: Simple test-time scaling」。

The s1K dataset is a high-quality reasoning dataset published in 2025 by Li Fei-Fei's team. It comprises 1,000 questions alongside their detailed reasoning traces and answers, with the answers derived from the distillation outputs of Google's Gemini Thinking Experimental. The dataset spans 50 distinct domains including probability theory, quantitative interview questions, Olympiad-level problems and others, enabling models to tackle a wide variety of reasoning tasks. The associated research paper is titled "s1: Simple Test-Time Scaling".

搜集汇总

背景与挑战

背景概述

s1K推理问题数据集由李飞飞团队于2025年发布，包含1000个问题及其推理轨迹和答案，答案基于Google Gemini Thinking Experimental蒸馏。该数据集覆盖50个不同领域，旨在通过简洁的数据工程优化模型在推理任务中的表现。