GPQA
收藏资源简介:
GPQA是一个由生物学、物理学和化学领域的专家编写的448个多选题数据集。该数据集的特点是问题质量高且极其困难,即使是具有博士学位或在读博士的专家也仅能达到65%的准确率,而高技能的非专家验证者仅有34%的准确率。数据集旨在用于研究未来AI系统在帮助解答非常困难问题时的可扩展监督方法,特别是在开发新科学知识时。
GPQA is a multiple-choice dataset consisting of 448 questions compiled by experts in the fields of biology, physics, and chemistry. This dataset is characterized by high-quality and exceptionally challenging questions: even experts holding doctoral degrees or pursuing doctoral studies only achieve an accuracy of 65%, while highly skilled non-expert validators reach merely 34% accuracy. The dataset is designed for research into scalable supervision methodologies for future AI systems that assist in resolving highly challenging problems, especially during the advancement of new scientific knowledge.




