amphora/QwQ-LongCoT-130K

QwQ-LongCoT-130K是一个用于训练类似O1的大型语言模型（LLMs）的SFT数据集，包含约130k个实例。每个实例的响应是通过QwQ-32B-Preview模型生成的。数据集主要由两部分组成：一部分来自NuminaMath-CoT数据集的90k个样本，另一部分是通过Magpie方法生成的约43k个样本。数据集的特点是包含了较长的链式思考（Chain-of-Thought）推理，适用于需要复杂推理的任务。数据集经过了过滤，去除了包含不当内容的响应，并且部分样本可能仍包含中文字符。数据集的使用受到Apache 2.0许可证的保护。

QwQ-LongCoT-130K is an SFT dataset designed for training O1-like large language models, containing about 130k instances. The dataset construction involves sourcing math questions and answers from the NuminaMath-CoT dataset and using the Magpie approach to extract questions from the QwQ-32B-Preview model to generate diverse instructions. The generated responses are filtered to remove instances containing specific phrases or repetitive sentences. The dataset analysis shows its distribution and length distribution, compared with the Magpie-Ultra dataset.