Only Connect Wall (OCW) dataset

Only Connect Wall (OCW)数据集是由风筝研究所等机构创建，包含618个来自英国问答节目《Only Connect》的连接墙问题。每个问题要求将16个混乱的线索词分组为4个连接组，并确定正确的连接。数据集设计包含误导性线索，即红鲱鱼，用以模拟认知神经科学中的固定效应和Einstellung范式，评估大型语言模型在创造性问题解决任务中的表现。该数据集旨在解决评估语言模型创造性解决问题能力的问题，特别是在面对误导性信息时的表现。

The Only Connect Wall (OCW) dataset was developed by institutions including the Kite Research Institute, and comprises 618 Connect Wall questions sourced from the British quiz show *Only Connect*. Each question requires models to sort 16 scrambled clue words into 4 semantically connected groups and identify the correct underlying connections for each group. The dataset incorporates misleading clues, referred to as red herrings, to simulate the fixed effect and Einstellung paradigm in cognitive neuroscience, and evaluate the performance of large language models (LLMs) on creative problem-solving tasks. This dataset is designed to address the critical gap in assessing the creative problem-solving capabilities of language models, particularly their performance when confronted with misleading information.