Zhihu Answer Preference Dataset

该数据集是由知乎上公开可获取的问答对构建而成的，旨在训练一个能够辨识人类对答案偏好的奖励模型。为了减少偏差，问答对是基于用户的点赞和发布时间创建的。该数据集的规模包括10万个用于训练的答案偏好对和5,000个用于评估的对。所涉及的任务是答案偏好排序。

This dataset is constructed from publicly accessible question-answer pairs sourced from Zhihu, with the goal of training a reward model that can recognize human preferences for answers. To mitigate bias, the question-answer pairs are curated based on users' upvotes and post timestamps. The dataset comprises 100,000 answer preference pairs for training and 5,000 pairs for evaluation. The core task involved is answer preference ranking.