webgpt_comparisons

WebGPT Comparisons 仓库提供了一个包含19578条对比数据的集合，这些数据来源于WebGPT项目中用于训练奖励模型的反馈。数据集中，每个样本包含针对同一问题的两个模型答案及其元数据，包括问题文本、模型浏览时找到的引文、最终答案以及答案的偏好得分（-1到1之间）。此数据集适用于训练奖励模型，以使问答模型更好地对齐人类偏好。该数据集的授权许可信息可在其引用的论文中找到。

The WebGPT Comparisons repository provides a collection of 19,578 comparative samples sourced from the feedback used to train the reward model in the WebGPT project. Each sample in this dataset contains two model answers to the same question along with their associated metadata, including the question text, citations retrieved by the model during browsing, the final generated answers, and the preference score ranging from -1 to 1. This dataset is intended for training reward models to better align question-answering models with human preferences. The licensing information for this dataset can be found in the papers that cite it.

搜集汇总

背景与挑战

背景概述

WebGPT Comparisons数据集是一个包含19,578条对比样本的集合，源自WebGPT项目，专门用于训练奖励模型以对齐人类偏好。每个样本提供同一问题的两个模型答案、相关引文和偏好得分（范围-1到1），适用于优化问答模型的对齐性能。