OpenMathInstruct-2
收藏资源简介：
OpenMathInstruct-2是一个数学指令调优数据集，包含1400万个问题-解决方案对。这些数据对是通过使用Llama3.1-405B-Instruct模型生成的。数据集的构建基于GSM8K和MATH训练集的问题，通过解决方案增强和问题-解决方案增强两种方式生成新的问题和解决方案。数据集包含四个主要字段：问题、生成的解决方案、预期答案和问题来源。预期答案对于训练集中的问题是从数据集中提供的真实答案，而对于增强的问题则是通过多数投票得出的答案。数据集还包括一个污染探索器，用于查找与GSM8K、MATH、AMC 2023、AIME 2024和Omni-MATH测试集问题相似的问题。此外，还发布了一系列基于此数据集训练的OpenMath2模型，并提供了代码、模型和数据集的链接，以及如何重现结果的详细说明。
OpenMathInstruct-2 is a mathematical instruction tuning dataset containing 14 million question-solution pairs. These pairs are generated using the Llama3.1-405B-Instruct model. The dataset is constructed based on the questions from the GSM8K and MATH training sets, with new questions and solutions generated via two approaches: solution augmentation and question-solution augmentation. It includes four core fields: question, generated solution, expected answer, and question source. For questions from the original training sets, the expected answer is the ground-truth answer provided in the source dataset, while for augmented questions, the expected answer is derived via majority voting. The dataset also features a contamination explorer designed to identify questions similar to those in the GSM8K, MATH, AMC 2023, AIME 2024, and Omni-MATH test sets. Additionally, a series of OpenMath2 models trained on this dataset have been released, along with links to the code, models, and dataset, as well as detailed instructions on how to reproduce the experimental results.
OpenMathInstruct-2 数据集概述
基本信息
- 数据集名称: OpenMathInstruct-2
- 许可证: CC BY 4.0
- 语言: 英语 (en)
- 标签: 数学 (math), NVIDIA
- 大小类别: 10M < n < 100M
- 任务类别: 问答 (question-answering), 文本生成 (text-generation)
数据集结构
- 特征:
- problem: 字符串类型，原始问题或增强问题。
- generated_solution: 字符串类型，合成生成的解决方案。
- expected_answer: 字符串类型，原始问题的真实答案或增强问题的多数投票答案。
- problem_source: 字符串类型，问题来源（GSM8K、MATH或增强版本）。
数据生成
- 生成方式:
- Solution augmentation: 为GSM8K和MATH训练集中的问题生成链式思维解决方案。
- Problem-Solution augmentation: 生成新问题，并为其生成解决方案。
数据集内容
- 问题来源:
- 直接来自GSM8K或MATH训练集的问题。
- 从GSM8K或MATH训练集增强生成的新问题。
附加资源
- 污染探索器: 用于查看OpenMathInstruct-2数据集中与GSM8K、MATH、AMC 2023、AIME 2024和Omni-MATH测试集问题相似的问题。
- 相关论文: 详细信息参见 arXiv:2410.01560。
模型表现
- 模型:
- Llama3.1-8B-Instruct
- OpenMath2-Llama3.1-8B
- Llama3.1-70B-Instruct
- OpenMath2-Llama3.1-70B
- 性能:
- 在GSM8K、MATH、AMC 2023、AIME 2024和Omni-MATH测试集上的表现。
开源资源
- 代码: NeMo-Skills
- 模型: NVIDIA OpenMath 2 模型集合
- 数据集: OpenMathInstruct-2
引用
bibtex @article{toshniwal2024openmath2, title = {OpenMathInstruct-2: Accelerating AI for Math with Massive Open-Source Instruction Data}, author = {Shubham Toshniwal and Wei Du and Ivan Moshkov and Branislav Kisacanin and Alexan Ayrapetyan and Igor Gitman}, year = {2024}, journal = {arXiv preprint arXiv:2410.01560} }
