True Detective: A Deep Abductive Reasoning Benchmark
收藏资源简介:
True Detective数据集由塔尔图大学计算机科学研究所创建,包含191个平均1200字的侦探谜题,源自“5 Minute Mystery”平台。每个谜题包含4-5个答案选项,旨在通过复杂的线索和观察测试高级推理能力。数据集的创建涉及专业和有抱负的作者编写谜题,每个谜题平均被尝试2000次,平均解决率为47%。该数据集主要用于评估大型语言模型在高级推理任务中的表现,特别是在解决复杂犯罪问题方面的能力。
The True Detective dataset was developed by the Institute of Computer Science at the University of Tartu. It consists of 191 detective puzzles with an average word count of 1200 words, sourced from the "5 Minute Mystery" platform. Each puzzle features 4 to 5 answer options, and is designed to assess advanced reasoning abilities through complex clues and observational skills. The dataset was created using puzzles written by both professional and aspiring writers; each puzzle has been attempted an average of 2,000 times, with an average solve rate of 47%. This dataset is primarily utilized to evaluate the performance of large language models (LLMs) on advanced reasoning tasks, specifically their capacity to resolve complex criminal cases.

- 1True Detective: A Deep Abductive Reasoning Benchmark Undoable for GPT-3 and Challenging for GPT-4塔尔图大学计算机科学研究所 · 2023年



