pheonix-delta/WiredBrain-Math-CoT

WiredBrain数学推理与链式思维（CoT）数据集是一个经过精心策划、去重和统一处理的集合，包含219,647个数学指令-响应对，旨在训练语言模型进行逐步问题解决。该数据集通过清理、格式化和去重多个来源，提供了高质量的资源，用于链式思维微调（帮助较小规模的语言模型生成自然语言思考轨迹）、复杂多领域推理（涵盖简单算术、符号代数、微积分和物理级控制问题）以及民主化AI推理（为社区提供基准级训练数据以匹敌专有推理模型）。数据集合并了多个高影响力来源，包括NuminaMath-CoT、MetaMathQA、OpenMathInstruct-1和Advanced Math arXiv Abstracts，每个条目包含指令、响应和来源信息。The WiredBrain Math Reasoning & Chain-of-Thought (CoT) Dataset is a curated, deduplicated, and unified collection of 219,647 mathematical instruction-response pairs designed to train language models on step-by-step problem-solving. By cleaning, formatting, and deduplicating these sources, this dataset provides a high-quality resource for Chain-of-Thought (CoT) Tuning (teaching smaller LLMs to generate natural language thinking traces), Complex Multi-Domain Reasoning (bridging gaps between simple arithmetic, symbolic algebra, calculus, and physics-level control problems), and Democratic AI Reasoning (providing benchmark-grade training data to match proprietary reasoning models). It merges and standardizes several high-impact sources including NuminaMath-CoT, MetaMathQA, OpenMathInstruct-1, and Advanced Math arXiv Abstracts, with each entry containing instruction, response, and source fields.