togethercomputer/RedPajama-Data-1T
收藏资源简介:
RedPajama是一个用于文本生成任务的大型数据集,包含2084个jsonl文件,总token数达到1.2万亿。数据集主要使用英语,但也包含多语言的Wikipedia部分。数据集结构包括文本内容、元数据(如URL、时间戳、来源和语言)以及标识数据子集的字段。创建过程涉及从多个源(如Commoncrawl、C4、GitHub等)下载和预处理数据,以尽可能接近LLaMa论文的描述。
RedPajama is a large-scale dataset designed for text generation tasks, comprising 2084 JSONL files with a total of 1.2 trillion tokens. The dataset is predominantly in English, but also includes multilingual Wikipedia subsets. Its structure encompasses text content, metadata (such as URLs, timestamps, sources, and language tags), as well as fields for identifying specific data subsets. The creation process involved downloading and preprocessing data from multiple sources (e.g., Common Crawl, C4, GitHub, etc.), with the goal of closely aligning with the descriptions outlined in the LLaMa paper.
数据集概述
名称: RedPajama-Data-1T
任务类别: 文本生成
语言: 主要为英语,Wikipedia部分包含多语言
数据集结构: json { "text": ..., "meta": {"url": "...", "timestamp": "...", "source": "...", "language": "...", ...}, "red_pajama_subset": "common_crawl" | "c4" | "github" | "books" | "arxiv" | "wikipedia" | "stackexchange" }
数据集组成:
- 包含2084个jsonl文件
- 可通过HuggingFace或直接下载获取
数据集来源及处理:
- Commoncrawl: 下载自Commoncrawl,经过去重和质量过滤
- C4: 下载自Huggingface,格式转换
- GitHub: 下载自Google BigQuery,去重和质量过滤,仅保留特定开源许可项目
- Wikipedia: 下载自Huggingface,已预处理
- ArXiv: 下载自Amazon S3,仅保留LaTeX源文件并去除不必要部分
- StackExchange: 下载自Internet Archive,仅保留28个最大站点内容,去除HTML标签,组织为问答对
数据集大小:
- 总计1.2万亿token
- 各部分token数量如下:
- Commoncrawl: 878亿
- C4: 175亿
- GitHub: 59亿
- Books: 26亿
- ArXiv: 28亿
- Wikipedia: 24亿
- StackExchange: 20亿
许可证:
- 请参考各数据子集的许可证
引用信息:
@software{together2023redpajama, author = {Together Computer}, title = {RedPajama: An Open Source Recipe to Reproduce LLaMA training dataset}, month = April, year = 2023, url = {https://github.com/togethercomputer/RedPajama-Data} }




