Nazonazo Benchmark
收藏资源简介:
Nazonazo Benchmark数据集由日本儿童谜语构建,旨在测试大型语言模型的基于洞察力的推理能力。该数据集包含120个谜语,每个谜语简短(大多数为一句话),无需专业知识,且易于大规模生成。数据集的创建过程包括从Nazonazo Gakuen网站收集谜语,并对模型进行评估。该数据集的应用领域是评估大型语言模型的洞察力推理能力,旨在解决当前大型语言模型评估危机的问题。
The Nazonazo Benchmark dataset is constructed from Japanese children's riddles, aiming to test the insight-based reasoning capabilities of large language models (LLMs). It comprises 120 riddles, each of which is brief (most are single sentences), requires no professional expertise, and is easy to generate at scale. The dataset's creation process includes collecting riddles from the Nazonazo Gakuen website and conducting model assessments. The primary application of this dataset is to evaluate the insight-based reasoning abilities of large language models, with the intention of resolving the current crisis in large language model evaluation.

- 1The NazoNazo Benchmark: A Cost-Effective and Extensible Test of Insight-Based Reasoning in LLMs日本先进科学技术研究所 · 2025年



