FineInstructions
收藏资源简介:
FineInstructions是由宾夕法尼亚大学等机构构建的大规模指令微调数据集,其核心创新在于通过18M真实用户查询生成的模板,将3000亿Token的无监督预训练语料转化为10亿级高质量指令-答案对。该数据集采用自动化流水线实现文档与指令模板的语义匹配,并基于高斯池化等技术创新确保数据多样性。其应用聚焦于大语言模型的预训练优化,通过指令驱动的监督式学习替代传统无监督预训练,显著提升模型在自由格式应答任务中的表现,为LLM的知识吸收效率提供了新范式。
FineInstructions is a large-scale instruction tuning dataset constructed by institutions such as the University of Pennsylvania and others. Its core innovation lies in converting 300 billion tokens of unsupervised pre-training corpus into 1 billion high-quality instruction-response pairs using templates generated from 18 million real user queries. This dataset adopts an automated pipeline to achieve semantic matching between documents and instruction templates, and leverages technical innovations such as Gaussian pooling to ensure data diversity. Its applications focus on pre-training optimization for large language models (LLMs). It replaces traditional unsupervised pre-training with instruction-driven supervised learning, significantly improving the model's performance on free-form response tasks, and provides a new paradigm for enhancing the knowledge absorption efficiency of LLMs.




