The Pile: An 800GB Dataset of Diverse Text for Language Modeling

TL;DR

The Pile is a large, diverse 825GB text dataset designed to enhance training of large-scale language models, improving their generalization across various domains.

Contribution

We introduce the Pile, a novel large-scale, diverse dataset from multiple sources, aimed at advancing language model training and evaluation.

Findings

Models trained on the Pile outperform those trained on raw Common Crawl data.

GPT-2 and GPT-3 struggle with some Pile components, highlighting dataset challenges.

The dataset reveals potential biases and issues for future research.

Abstract

Recent work has demonstrated that increased training dataset diversity improves general cross-domain knowledge and downstream generalization capability for large-scale language models. With this in mind, we present \textit{the Pile}: an 825 GiB English text corpus targeted at training large-scale language models. The Pile is constructed from 22 diverse high-quality subsets -- both existing and newly constructed -- many of which derive from academic or professional sources. Our evaluation of the untuned performance of GPT-2 and GPT-3 on the Pile shows that these models struggle on many of its components, such as academic writing. Conversely, models trained on the Pile improve significantly over both Raw CC and CC-100 on all components of the Pile, while improving performance on downstream evaluations. Through an in-depth exploratory analysis, we document potentially concerning aspects of the data for prospective users. We make publicly available the code used in its construction.