openai/mrcr

OpenAI MRCR是一个长上下文多轮指代消解基准数据集，旨在评估大型语言模型在区分长上下文中隐藏的多个相似信息的能力。数据集包含模拟对话，对话中包含用户请求撰写关于某一主题的多个写作内容，模型需要返回特定顺序的写作内容。数据集难度随相似请求的数量和上下文长度的增加而提高。

OpenAI MRCR is a long context multi-round co-reference resolution benchmark designed to evaluate the ability of large language models to distinguish among multiple similar pieces of information hidden in the context. The dataset consists of simulated conversations containing multiple writing requests on a certain topic, and the model is required to return the writing content in a specific order. The difficulty of the dataset increases with the number of similar requests and the length of the context.