Anthropic’s Helpful and Harmless (HH)
收藏arXiv2025-09-30 收录
下载链接：
https://huggingface.co/OpenAssistant/reward-model-deberta-v3-large-v2
下载链接
链接失效反馈官方服务：
资源简介：
该数据集包含了161,000个训练样本和8,500个测试样本，每个样本都包括一个提示和一对回应（一个是被选中的，另一个是被拒绝的）。此外，回应是通过使用奖励模型进行评分的，这产生了列表大小为K=8的排名偏好数据。该数据集的规模为161,000个训练样本和8,500个测试样本，所涉及的任务是偏好学习和回应排名。
This dataset comprises 161,000 training samples and 8,500 test samples. Each sample includes a prompt and a pair of responses, with one being selected and the other rejected. Additionally, the responses are scored using a reward model, generating ranked preference data with a list size of K=8. The tasks involved in this dataset are preference learning and response ranking.
提供机构：
Anthropic
