alea-institute/kl3m-data-uspto

KL3M数据集是ALEA研究所提供的、用于大型语言模型训练的无版权问题训练资源，包含超过1.32亿个文档和数万亿个标记，涵盖16个不同来源。该数据集采用Parquet格式存储文档文本和元数据，并遵循CC BY 4.0许可。它包括获取和处理文档的源代码、原始文档及其元数据、标准化内容、预标记文档以及用于问答、摘要、转换、起草、分类、预测和对话的各种训练资源。

The KL3M dataset is a set of copyright-clean training resources provided by the ALEA Institute for the training of large language models, containing over 132 million documents and trillions of tokens from 16 different sources. The dataset uses the Parquet format to store document text and metadata, and is licensed under CC BY 4.0. It includes source code for acquiring and processing documents, original document formats with metadata, standardized content, pre-tokenized documents, and various training resources for question-answering, summarization, conversion, drafting, classification, prediction, and conversational applications.