GSM8K
收藏资源简介:
该研究基于GSM8K数学问题-答案数据集,构建了一个包含256条样本的精简子集,由马里兰大学团队通过语义软自举(SSB)方法自动生成。数据集包含原始数学问题及其对应的分步推理解答,通过模型自生成-筛选-修正的流程产生,无需人工标注。其核心特点在于采用对比学习框架,整合正确与典型错误解答以增强模型推理鲁棒性。该数据集专为提升大语言模型的数学推理能力设计,在MATH500和AIME2024基准测试中实现了10%以上的准确率提升,适用于数学教育、自动解题等AI推理研究领域。
This study develops a compact subset of 256 samples based on the GSM8K math problem-answer dataset, which is automatically generated by the University of Maryland team using the Semantic Soft Bootstrapping (SSB) method. The dataset includes original mathematical problems and their corresponding step-by-step reasoning solutions, produced via a model-driven self-generation, filtering and revision pipeline without manual annotation requirements. Its core characteristic is the adoption of a contrastive learning framework that integrates both correct and typical incorrect solutions to strengthen the model's reasoning robustness. This dataset is specifically designed to enhance the mathematical reasoning abilities of Large Language Models (LLMs), achieving an accuracy improvement of over 10% on the MATH500 and AIME2024 benchmark tests, and is suitable for AI reasoning research areas such as mathematics education and automated problem-solving.

- 1Semantic Soft Bootstrapping: Long Context Reasoning in LLMs without Reinforcement Learning马里兰大学 · 2025年



