WritingPreferenceBench
收藏资源简介:
WritingPreferenceBench是一个包含1800个由人工标注的偏好对的数据集,涵盖了8种创意写作类型,其中1200个英语偏好对,600个中文偏好对。数据集旨在评估主观写作偏好,通过消除了客观质量信号的影响,例如语法错误、事实错误和长度差异。该数据集为评估主观偏好提供了一个新的标准,并揭示了当前偏好学习方法在捕捉主观创造质量方面的局限性。
WritingPreferenceBench is a dataset consisting of 1,800 manually annotated preference pairs, spanning 8 categories of creative writing, including 1,200 English preference pairs and 600 Chinese preference pairs. Aimed at evaluating subjective writing preferences, this dataset eliminates the interference of objective quality signals such as grammatical errors, factual inaccuracies and length discrepancies. It serves as a novel benchmark for subjective preference evaluation and reveals the limitations of current preference learning methods in capturing subjective creative quality.




