Columbia-NLP/DPO-HelpSteer
收藏Hugging Face2024-07-10 更新2024-07-06 收录
下载链接：
https://hf-mirror.com/datasets/Columbia-NLP/DPO-HelpSteer
下载链接
链接失效反馈官方服务：
资源简介：
我们将nvidia/HelpSteer数据集重新格式化为DPO数据集的通用格式。具体来说，我们将所有分数转换为[1, 10]的尺度，并通过取最佳评分响应作为chosen，随机抽取评分低于最佳响应的响应作为rejected来构建偏好对。我们跳过了所有响应分数相同的提示/数据行。
This dataset is reformatted from nvidia/HelpSteer and is used for training the LION series models. It includes multiple features such as prompt, prompt_id, chosen, rejected, messages, score_chosen, score_rejected, and other_info. The chosen and rejected features contain content and role, while other_info includes multiple scoring dimensions such as coherence, complexity, correctness, helpfulness, and verbosity. The dataset is split into train and validation parts, containing 8722 and 443 samples respectively.
提供机构：
Columbia-NLP
