Dataset Card for Dataset Name

Dataset Summary

RedPajama is a clean-room, fully open-source implementation of the LLaMa dataset. This HuggingFace repo contains a 1B-token sample of the RedPajama dataset. The full dataset has the following token counts and is available for download :

A full set of scripts to recreate the dataset from scratch can be found here .

Languages

Primarily English, though the Wikipedia slice contains multiple languages.

Dataset Structure

The dataset structure is as follows:

{
    "text": ...,
    "meta": {"url": "...", "timestamp": "...", "source": "...", "language": "...", ...}
}

Dataset Creation

This dataset was created to follow the LLaMa paper as closely as possible to try to reproduce its recipe.

Source Data

We download five dumps from Commoncrawl, and run the dumps through the official cc_net pipeline. We then deduplicate on the paragraph level, and filter out low quality text using a linear classifier trained to classify paragraphs as Wikipedia references or random Commoncrawl samples.

C4 is downloaded from Huggingface. The only preprocessing step is to bring the data into our own format.

The raw GitHub data is downloaded from Google BigQuery. We deduplicate on the file level and filter out low quality files and only keep projects that are distributed under the MIT, BSD, or Apache license.

We use the Wikipedia dataset available on Huggingface, which is based on the Wikipedia dump from 2023-03-20 and contains text in 20 different languages. The dataset comes in preprocessed format, so that hyperlinks, comments and other formatting boilerplate has been removed.

The PG19 subset of the Gutenberg Project and Books3 datasets are downloaded from Huggingface. After downloading, we use simhash to remove near duplicates.

ArXiv data is downloaded from Amazon S3 in the arxiv requester pays bucket. We only keep latex source files and remove preambles, comments, macros and bibliographies.

The Stack Exchange split of the dataset is download from the Internet Archive . Here we only keep the posts from the 28 largest sites, remove html tags, group the posts into question-answer pairs, and order answers by their score.