MetaMathQA 数学推理数据集

大多数现有的开源 LLM（例如 LLaMA-2）推理过程都比较复杂，在解决数学问题方面仍然不能令人满意。为了弥补这一差距，研究人员提出了 MetaMath，这是一种专门从事数学推理的微调语言模型。为了提升模型的正向和逆向推理能力，剑桥、港科大、华为的研究者基于两个常用的数学数据集（GSM8K 和 MATH）提出了 MetaMathQA 数据集：一个覆盖面广、质量高的数学推理数据集。 MetaMathQA 由 395K 个大语言模型生成的正向逆向数学问答对组成。他们在 MetaMathQA 数据集上基于 LLaMA-2 微调得到专注于数学推理（正向和逆向）的大语言模型 MetaMath，在数学推理数据集上达到了 SOTA 。 MetaMathQA 数据集和不同规模的 MetaMath 模型已开源供研究人员使用。

Most existing open-source LLMs (e.g., LLaMA-2) feature complex inference processes and remain unsatisfactory in solving mathematical problems. To address this gap, researchers proposed MetaMath, a fine-tuned language model specialized in mathematical reasoning. To enhance the model's forward and backward reasoning capabilities, researchers from the University of Cambridge, Hong Kong University of Science and Technology (HKUST), and Huawei proposed the MetaMathQA dataset—a high-quality, comprehensive mathematical reasoning dataset—based on two widely used mathematical datasets, GSM8K and MATH. MetaMathQA consists of 395K forward and backward mathematical question-answer pairs generated by large language models. Using LLaMA-2 as the base model, the researchers fine-tuned it on the MetaMathQA dataset to obtain MetaMath, an LLM specialized in mathematical (forward and backward) reasoning, which achieved state-of-the-art (SOTA) performance on mathematical reasoning benchmarks. The MetaMathQA dataset and MetaMath models of various sizes have been open-sourced for research use.

搜集汇总

背景与挑战

背景概述

MetaMathQA是一个专注于数学推理的高质量数据集，由395K个大语言模型生成的正向逆向问答对组成，基于GSM8K和MATH数据集构建。它采用四种数据增强方法（如答案增强和问题改写）来提升模型的推理能力，旨在微调LLaMA-2等模型以在数学任务上达到最优性能。