summarize_from_feedback

Summarize from Feedback 提供了用于训练奖励模型的由人类反馈组成的数据集，旨在使摘要模型与人类偏好对齐。该数据集包含两部分：`comparisons` 部分通过人工标注比较两个摘要的优劣，`axis` 部分则让人工标注者对摘要质量进行评分。`comparisons` 部分包含训练集和验证集，而 `axis` 部分包含测试集和验证集。该数据集的摘要来源于 TL;DR 数据集、CNN 文章和 Daily Mail 文章。

The Summarize from Feedback dataset is a collection of human feedback data designed for training reward models, with the objective of aligning summarization models with human preferences. This dataset comprises two core components: the `comparisons` split, in which human annotators compare two summaries and determine the better-performing one, and the `axis` split, where human annotators assign quality ratings to individual summaries. The `comparisons` component includes both training and validation subsets, while the `axis` component consists of test and validation subsets. The summaries within this dataset are sourced from the TL;DR dataset, CNN articles, and Daily Mail articles.