MT-Redundant

该数据集名为MT-Redundant，它将MT-Bench的第二个回合转变为对第一个回合答案变体的请求，模拟现实世界中用户对先前回应提出修改要求的场景。该数据集涵盖了各种实际用例，这些用例被分为八个类别：编程、提取、人文、数学、推理、角色扮演、STEM领域和写作。在规模上，该数据集对GSM-8K和CNN/DM限制在100个样本；其他数据集的大小与之相当。任务涉及多回合互动，包含重复信息。

This dataset, named MT-Redundant, recasts the second turn of MT-Bench as a request for variant versions of the first-round model responses, simulating real-world scenarios where users request revisions to prior model responses. It comprises a diverse set of real-world use cases categorized into eight categories: Programming, Extraction, Humanities, Mathematics, Reasoning, Role-playing, STEM, and Writing. In terms of scale, the dataset restricts the sample count of GSM-8K and CNN/DM to 100 instances, while the sizes of other included datasets are comparable. The tasks involve multi-turn interactions with repeated information.