UWNSL/MATH_training_split_long_cot
收藏Hugging Face2025-02-21 更新2025-04-19 收录
下载链接：
https://hf-mirror.com/datasets/UWNSL/MATH_training_split_long_cot
下载链接
链接失效反馈官方服务：
资源简介：
本数据集用于研究小型模型（参数不超过3B）在学习长链式思维（CoT）或从大型模型中蒸馏知识时的学习差距问题。数据集包含了从MATH数据集的7.5k训练分割中选取的问题，并使用Qwen/QwQ-32B-Preview模型通过拒绝采样生成长链式思维答案。该数据集旨在探究长链式思维数据与短链式思维数据对不同的学生模型的影响。
This dataset is for studying the learning gap of small models (≤3B parameters) when learning from long chain-of-thought (CoT) or distilling knowledge from larger models. The dataset consists of problems selected from the 7.5k training split of the MATH dataset, and long chain-of-thought answers generated using the Qwen/QwQ-32B-Preview model via reject sampling. The dataset aims to investigate the effects of long and short CoT data on different student models.
提供机构：
UWNSL
