argilla/ultrafeedback-binarized-preferences

该数据集是Ultrafeedback的二值化版本，通过Argilla工具进行整理。数据集解决了原始UltraFeedback数据集中`overall_score`生成方式的问题，通过计算偏好评分的均值来选择最佳响应，并随机选择评分较低的响应作为被拒绝的响应。数据集包含多个特征，如来源、指令、选择的响应、被拒绝的响应等，并且提供了训练集的分割信息。

This dataset is a binarized version of UltraFeedback, curated using the Argilla tool. It addresses the issue with the `overall_score` generation method in the original UltraFeedback dataset. Specifically, it calculates the mean of preference scores to select the best response, and randomly selects a response with a lower score as the rejected response. The dataset includes multiple features such as source, instruction, chosen response, rejected response, etc., and provides training set split information.

原始信息汇总

数据集概述

数据集来源

• 该数据集由Argilla进行整理和加工。

数据集内容

• 数据集包含Argilla整理工作的成果。