Telephone Game LLM Dataset

该数据集包含了不同大型语言模型在一系列“传话游戏”实验中通过迭代传递生成的文本。它追踪了毒性、积极性、难度和长度等文本属性在多代传递中的演变。此外，该数据集支持研究文本属性在代际间的演变，并分析大型语言模型交互中的吸引态。该数据集的规模包括5个模型、3项任务、20篇初始文本，以及每条包含50个步骤的5条传递链。其任务是通过迭代传递进行文本生成。

This dataset contains text generated via iterative transmission in a series of "telephone game" experiments conducted by different large language models (LLMs). It tracks the evolution of key textual attributes including toxicity, positivity, difficulty, and length across multi-generational transmission processes. Furthermore, this dataset supports research on the inter-generational evolution of textual attributes and the analysis of attractor states in large language model interactions. The dataset encompasses 5 models, 3 tasks, 20 initial text samples, and 5 transmission chains each containing 50 steps. Its targeted task is text generation through iterative transmission.