thnhan3/vietnamese-summarization-dpo

该数据集包含为直接偏好优化（DPO）训练设计的越南语文本摘要对。它是在现有越南语摘要数据集的基础上构建的，并添加了低质量摘要以进行偏好学习。数据集包含2000个偏好对，涵盖新闻文章、健康指南和生活方式提示等领域。每个样本包括一个提示、一个高质量选择的摘要和一个带有故意错误的低质量拒绝摘要。数据集采用JSON格式，并可以使用Python中的datasets库加载。它旨在用于DPO训练、模型开发、偏好学习研究以及摘要质量评估。该数据集根据MIT许可证授权，并欢迎改进的贡献。

This dataset contains Vietnamese text summarization pairs designed for Direct Preference Optimization (DPO) training. It is an extension of an existing Vietnamese summarization dataset with added low-quality summaries for preference learning. The dataset includes 2000 preference pairs across domains such as news articles, health guides, and lifestyle tips. Each sample consists of a prompt, a high-quality chosen summary, and a rejected low-quality summary with intentional errors. The dataset is structured in JSON format and can be loaded using the datasets library in Python. It is intended for use in DPO training, model development, preference learning research, and summarization quality evaluation. The dataset is licensed under the MIT License and contributions for improvement are welcomed.