LIMO 数学推理基准数据集

LIMO (Less Is More for Reasoning) 是一个数学推理数据集，旨在通过精心挑选高质量的训练样本，训练和评估大模型的数学推理能力，相关论文成果为「LIMO: Less is More for Reasoning」。该数据集主要用于训练大模型的数学解题能力，提升其在数学考试、竞赛题目（如 AIME 、 MATH-500 等）上的表现。

LIMO (Less Is More for Reasoning) is a mathematical reasoning dataset developed to train and evaluate the mathematical reasoning capabilities of large language models (LLMs) via curated high-quality training samples. Its associated scholarly paper is titled "LIMO: Less is More for Reasoning". This dataset is primarily intended to enhance the mathematical problem-solving abilities of LLMs and boost their performance on mathematical examinations and competitive problems including AIME, MATH-500, and other similar tasks.

搜集汇总

背景与挑战

背景概述

LIMO数学推理基准数据集专注于通过精心挑选的高质量样本，训练和评估大模型的数学推理能力。该数据集仅包含817个样本，规模小但高效，在AIME 2025评测中取得了接近大规模训练模型的成绩。