The Pile Dataset

https://pile.eleuther.ai/

The Pile is a 825 GiB diverse, open source language modelling data set that consists of 22 smaller, high-quality datasets combined together.

I think this dataset by Eleuther AI is probably the highest quality large text dataset publicly available.

1 Like

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.

Woah, this is great.

1 Like

It really is, I think it’s improved since I last looked at it. Reading the paper it looks like they have pretty high quality data (Hacker News, Project Gutenberg, Wikipedia, etc.) and a lot of it. They’re also thoughtful about dealing with copyright, they’re practical but try to use data that the authors intended to be open.

It’s more text data then I can currently use, so we shouldn’t have any shortage for the foreseeable future. It’s probably very English centric but also should have a lot of multilingual data. Translation data, especially for smaller languages, is still very valuable though.

I’ve done some experiments training models on unstructured text data by splitting sentences and “translating” to recover the second half of the sentence.

For example:

{"q":"I baked a cake ", "source":"auto", "target":"infer"}

↓

{"translatedText": "for my friend's birthday party."}

I haven’t had very good results but I think this would work well with more powerful models. This is similar to the AlexaTM model which used a translation style encoder-decoder architecture for generating text.

1 Like