rl-rag/combined-sft-training-data-v20250901
收藏Hugging Face2025-09-01 更新2025-09-13 收录
下载链接：
https://hf-mirror.com/datasets/rl-rag/combined-sft-training-data-v20250901
下载链接
链接失效反馈官方服务：
资源简介：
SFT数据集是一个包含了OpenScholar和MiroSFT响应的数据集。最新版本v20250901增加了1500个OpenScholar响应和由GPT4 mini生成的摘要片段。数据集对Miro子集采用了严格的`boxed{}`格式化，并进行了拒绝采样。Miro的系统提示也进行了更新。数据来源分布包括openscholar_deep_research_gpt5_v2_answers等五个部分，整体平均响应长度为4500字，OpenScholar平均长度为7200字，MiroSFT平均长度为1600字。
The SFT dataset is a collection of OpenScholar and MiroSFT responses. The latest version v20250901 includes 1.5k additional OpenScholar responses and snippet summarizations generated by GPT4 mini. The dataset has implemented strict `boxed{}` formatting and rejection sampling on Miro subsets, and system prompts for Miro have been updated. The source distribution includes five parts: openscholar_deep_research_gpt5_v2_answers, rl-rag/MiroVerse-QA-Expert-Multi-Hop-V1.0_XML, rl-rag/MiroVerse-TaskCraft_XML, rl-rag/openscholar_deep_research_gpt5_v2_answers, and rl-rag/MiroVerse-MegaScience_XML. The overall average response length is 4.5k words, with OpenScholar averaging 7.2k words and MiroSFT averaging 1.6k words.
提供机构：
rl-rag
