allenai/tulu-3-sft-prompts-ultrafeedback

该数据集名为Llama 3.1 Tulu 3 SFT Ultrafeedback，包含了基于Tulu 3 SFT提示和多个模型生成的响应的多个子数据集。数据集的特征包括id、prompt、chosen和rejected，其中chosen和rejected是包含content和role的列表。数据集分为多个split，每个split有不同的字节数和样本数。数据集的生成方法结合了on-policy和off-policy数据，并使用Ultrafeedback模板和LLM judge进行偏好标注。数据集的使用受ODC-BY许可证约束，部分数据为非商业用途。

The dataset includes multiple parts, each with unique features such as id, prompt, chosen, and rejected. The chosen and rejected features further contain sub-features like content and role. The dataset is divided into several splits, each with its specific name, number of bytes, and number of examples. The datasets generation involves multiple models, including Mistral, Tulu, Yi, MPT, Google Gemma, InternLM, Falcon, Qwen, Llama, and GPT-4, among others. The completion generation method combines on-policy and off-policy data and uses the Ultrafeedback template and LLM judge for preference annotation. The dataset is licensed under ODC-BY, suitable for research and educational purposes, but some data may be subject to third-party model usage terms.