SakanaAI/FishMath-SFT-Data
收藏资源简介:
FishMath SFT Data是一个用于数学推理的合成监督微调(SFT)数据集,专为Kaggle AI数学奥林匹克3 - 进展奖3(AIMO 3)项目设计。该数据集包含23,257条由多个前沿开源大型语言模型(LLM)生成的正确解题轨迹,涵盖了来自不同来源的竞赛级数学问题。数据集的特点之一是包含工具集成推理(TIR)轨迹,其中约29%的轨迹涉及Python代码执行步骤。每条记录包含数学问题陈述、正确答案以及一个完整的对话轨迹(多轮消息),该轨迹最终得出正确答案。数据集的问题来源于两个精选数据集:nvidia/Nemotron-Math-v2和ycchen/Crystal-Math-Preview,并经过严格的正确性过滤和处理。
FishMath SFT Data is a synthetic supervised fine-tuning (SFT) dataset for mathematical reasoning, specifically designed for the Kaggle AI Math Olympiad 3 - Progress Prize 3 (AIMO 3) project. This dataset contains 23,257 correct problem-solving trajectories generated by multiple cutting-edge open-source large language models (LLMs), covering competition-level mathematical problems from diverse sources. One notable feature of this dataset is the inclusion of Tool-Integrated Reasoning (TIR) trajectories, where approximately 29% of the trajectories involve Python code execution steps. Each record includes a mathematical problem statement, the correct answer, and a complete dialogue trajectory (multi-turn messages) that ultimately leads to the correct answer. The problems in this dataset are sourced from two curated datasets: nvidia/Nemotron-Math-v2 and ycchen/Crystal-Math-Preview, and the dataset has undergone rigorous correctness filtering and processing.




