RealChat

该数据集是由名为“苏格拉底”的用户模拟器生成的以人为中心的合成对话数据集，旨在为训练语言模型提供高质量的会话数据。该数据集用于训练ReaLM模型，重点关注系统代理的输出，同时忽略用户查询，确保在模型之间进行公平实验并保持一致的设置。规模上，我们从不同来源分层随机抽取了1万个数据集。该数据集的任务是对话建模以及语言模型的训练。

This dataset is a human-centric synthetic conversational dataset generated by a user simulator named "Socrates", designed to provide high-quality conversational data for language model training. It is utilized for training the ReaLM model, focusing on the outputs of system agents while disregarding user queries to ensure fair comparative experiments across models and consistent experimental configurations. In terms of scale, we conducted stratified random sampling of 10,000 conversational samples from diverse sources. The core tasks of this dataset cover conversational modeling and language model training.