RV-Syn
收藏资源简介:
RV-Syn是一个创新的数学推理数据合成方法,它通过构建一个基于初始种子问题的结构化数学操作函数库,生成计算图作为解决方案,然后将这些图反向翻译成复杂的问题。这种方法以解决方案为导向,利用计算图的执行性确保了求解过程的可验证性。通过这种方法,可以合成具有丰富语义和控制流的高质量推理数据集,从而推动大型语言模型在数学推理能力方面的进步。
RV-Syn is an innovative mathematical reasoning data synthesis method. It constructs a structured library of mathematical operation functions based on initial seed problems, generates computation graphs as solutions, and then reversely translates these graphs into complex problems. This solution-oriented method leverages the executability of computation graphs to ensure the verifiability of the solution process. Through this approach, high-quality reasoning datasets with rich semantics and control flow can be synthesized, thereby advancing the mathematical reasoning capabilities of large language models.




