The Pile Replication Code

The Pile is a large, diverse, open source language modelling data set that consists of many smaller datasets combined together. The objective is to obtain text from as many modalities as possible to ensure that models trained using The Pile will have much broader generalization abilities. We are currently developing Version 1, with an ultimate goal of 1 TiB of English text. After the completion of Version 1, our next goal is a fully-multilingual, 10TiB text dataset.

This repository is for replicating or making variants of the Pile. If you would like to use the Pile, an official release is planned for the near future. If you would like to be an early user of the Pile before the official release, please contact us through our discord server.

The Pile is currently under heavy development. Breaking changes may be introduced rapidly and without warning.

Component	Raw Size	Weight	Epochs	Effective Size	Mean Document Size
Pile-CC	227.12 GiB	18.11%	1.0	227.12 GiB	4.33 KiB
PubMed Central	90.27 GiB	14.40%	2.0	180.55 GiB	30.55 KiB
Books3	100.96 GiB	12.07%	1.5	151.44 GiB	538.36 KiB
OpenWebText2	62.77 GiB	10.01%	2.0	125.54 GiB	3.85 KiB
ArXiv	56.21 GiB	8.96%	2.0	112.42 GiB	46.61 KiB
Github	95.16 GiB	7.59%	1.0	95.16 GiB	5.25 KiB
FreeLaw	51.15 GiB	6.12%	1.5	76.73 GiB	15.06 KiB
StackExchange	32.20 GiB	5.13%	2.0	64.39 GiB	2.16 KiB
USPTO Backgrounds	22.90 GiB	3.65%	2.0	45.81 GiB	4.08 KiB
PubMed Abstracts	19.26 GiB	3.07%	2.0	38.53 GiB	1.30 KiB
Gutenberg (PG-19)	10.88 GiB	2.17%	2.5	27.19 GiB	398.73 KiB
OpenSubtitles	12.98 GiB	1.55%	1.5	19.47 GiB	30.48 KiB
Wikipedia (en)	6.38 GiB	1.53%	3.0	19.13 GiB	1.11 KiB
DM Mathematics	7.75 GiB	1.24%	2.0	15.49 GiB	8.00 KiB
Ubuntu IRC	5.52 GiB	0.88%	2.0	11.03 GiB	545.48 KiB
BookCorpus2	6.30 GiB	0.75%	1.5	9.45 GiB	369.87 KiB
EuroParl	4.59 GiB	0.73%	2.0	9.17 GiB	68.87 KiB
HackerNews	3.90 GiB	0.62%	2.0	7.80 GiB	4.92 KiB
YoutubeSubtitles	3.73 GiB	0.60%	2.0	7.47 GiB	22.55 KiB
PhilPapers	2.38 GiB	0.38%	2.0	4.76 GiB	73.37 KiB
NIH ExPorter	1.89 GiB	0.30%	2.0	3.79 GiB	2.11 KiB
Enron Emails	0.88 GiB	0.14%	2.0	1.76 GiB	1.78 KiB
Total				1254.20 GiB	5.91 KiB

(Epochs refers to the number of epochs elapsed after 1.2TB)

Usage

Install:

pip install -e .

To replicate pile

python the_pile/pile.py --interleave_output 30 --using pile_reprod

Use the pass 2 script here to complete shuffling.

Other

To force download all data:

python the_pile/pile.py --force_download

To generate fasttext training data for CC filtering (OWT2 only):

sudo apt install build-essential
python the_pile/pile.py --using owt2 --make_fasttext

Manual Download Components

The following components need manual downloading. Either download them or comment out from pile.py.

Bibliotik: books3.tar.gz needs to be in the current directory. Download temporarily unavailable.

Workflow

To propose a new dataset be added to the Pile, open an issue. Your issue should include a description of the dataset, its size, what language(s) it is in, a link to the data, and any other relevant information. If a project manger approves your proposal, they will change its label to and add it to . Datasets that we elect to not include in the current version of the Pile will receive a or label. While we welcome multilingual datasets and plan on including non-English datasets in the future, the initial release of the Pile will be English-only and all submissions of non-English datasets will be deferred.

To claim responsibility for implementing an unclaimed dataset, leave a comment on one of our unassigned issues. Once an dataset has been assigned to you, make the necessary changes to datsets.py and pile.py in a fork and submit a pull request. If you require, you can also submit a script for processing the data as shown here.

To raise an issue that is not proposing a new dataset, open an issue with the tag or as appropriate.

Data ready for final implementation should meet the following criteria:

The data must be in lm_dataformat format.
The data must be shuffled.