gp02-mcgill/ultrafeedback_binarised_all_max
收藏Hugging Face2025-01-31 更新2025-02-15 收录
下载链接：
https://hf-mirror.com/datasets/gp02-mcgill/ultrafeedback_binarised_all_max
下载链接
链接失效反馈官方服务：
资源简介：
ultrafeedback_binarised_max_min是一个从UltraFeedback数据集衍生出的成对偏好二进制标签数据集，用于训练需要二元偏好标签的模型，如强化学习中的基于人类反馈的模型和基于偏好的排名模型。数据集通过计算每个响应的平均评分，将评分最高的响应作为首选答案，并将其与其他评分较低的响应配对，形成偏好对比。
The ultrafeedback_binarised_max_min dataset is a pairwise preference binary labeled dataset derived from the UltraFeedback dataset, designed for training models that require binary preference labels, such as reinforcement learning models with human feedback and preference-based ranking models. The dataset is constructed by computing the average rating for each response, labeling the highest-rated response as the chosen answer, and pairing it with all other lower-rated responses for preference comparison.
提供机构：
gp02-mcgill
