JinaLeejnl/AlignX

AlignX数据集是一个包含大量Reddit数据和现有对齐数据集的集合，旨在维持普遍的价值对齐能力。该数据集总共包含1,311,622个样本，涵盖了90个自定义偏好维度，以及安全性、帮助性/诚实性/指令遵循性/真实性、帮助性/正确性/连贯性/复杂性/冗余性等多个维度。数据集格式包括发帖、用户偏好的回复、较不偏好的回复、90个维度的偏好方向、用户人口统计信息、同一用户在其他帖子上的评论以及同一用户对其他帖子下评论的偏好对。

The AlignX dataset is a collection of a large amount of Reddit data and existing alignment datasets aimed at maintaining universal value alignment capabilities. The dataset contains a total of 1,311,622 samples, covering 90 self-defined preference dimensions, as well as dimensions such as safety, helpfulness/honesty/instruction-following/truthfulness, helpfulness/correctness/coherence/complexity/verbosity. The dataset format includes the post, user-preferred response, less preferred response, preference direction for 90 dimensions, demographic information of the user, comments written by the same user on other posts, and the users preference pairs for comments under other posts.