mncai/orpo-text-pairs-full

该数据集包含两个版本的偏好对，用于使用ORPO、DPO或类似基于偏好的对齐方法训练语言模型。数据集包含经过精炼/过滤的版本（推荐使用）和完整版本。数据格式为JSONL，语言为英语，任务为纯文本偏好学习（不包含图像）。每个数据行包含prompt（聊天消息）、chosen（首选响应）、rejected（非首选响应）和meta（元数据，包括源数据集、使用的模型、判断信息等）字段。数据集来源于多个源数据集，包括HelpSteer2、MathInstruct、CodeIO-PyEdu-Reasoning和MathV360K，每个源数据集有不同的许可证要求。使用数据集时需要遵守相应的许可证要求，特别是对HelpSteer2和CodeIO-PyEdu-Reasoning的样本需要注明出处。

This dataset contains two versions of preference pairs for training language models using ORPO, DPO, or similar preference-based alignment methods. It includes a refined/filtered version (recommended) and the full dataset before filtering. The format is JSONL, the language is English, and the task is text-only preference learning (no images). Each row contains fields for prompt (chat messages), chosen (preferred response), rejected (non-preferred response), and meta (metadata including source dataset, models used, judge info, etc.). The dataset is derived from multiple source datasets including HelpSteer2, MathInstruct, CodeIO-PyEdu-Reasoning, and MathV360K, each with its own license terms. Attribution is required when using samples derived from HelpSteer2 or CodeIO-PyEdu-Reasoning.