exlaw/tis-dpo-data
收藏Hugging Face2025-04-06 更新2025-04-12 收录
下载链接：
https://hf-mirror.com/datasets/exlaw/tis-dpo-data
下载链接
链接失效反馈官方服务：
资源简介：
TIS-DPO数据集仓库包含了ICLR 2025论文中使用的实验数据集，用于验证针对直接偏好优化的token-level importance sampling方法。数据集包括：1. Anthropics HH数据集，专注于有助于和无害的AI响应，用于训练和评估AI的安全性和对齐；2. PKU-Safety数据集，来自北京大学的安全导向数据集，关注于AI的安全和负责任行为；3. TL-DR数据集，文本摘要数据集，包含长文本及其摘要对；4. Ultra-feedback数据集，包含对模型输出的反馈，用于偏好学习和优化。
The TIS-DPO Dataset Repository contains experimental datasets used in the ICLR 2025 paper, which are for validating the token-level importance sampling approach for direct preference optimization. The datasets included are: 1. Anthropic HH Dataset, focusing on helpful and harmless AI responses, used for training and evaluating AI safety and alignment; 2. PKU-Safety Dataset, a comprehensive safety-oriented dataset from Peking University, focusing on safe and responsible AI behavior; 3. TL-DR Dataset, a text summarization dataset containing pairs of long texts and their summaries; 4. Ultra-feedback Dataset, containing human feedback on model outputs, used for preference learning and optimization.
提供机构：
exlaw
