Last week we had the honor of participating at MSR’18, where two of the members of our team, Vadim Markovtsev and Waren Long, presented the research paper they wrote on our latest dataset: Public Git Archive.

Public Git Archive is the result of months of effort curating a dataset suitable for training Machine Learning on Source Code (aka MLonCode) models.

Dataset contents

The dataset contains 3TB of repositories from GitHub ready to download. This includes all of the contents (git metadata and file contents) for all of the repositories on GitHub with 50 or more stars. The list of repositories was obtained from the GHTorrent project, specifically from the snapshot of January 1st 2018. GHTorrent provides information about GitHub repositories but unfortunately doesn’t include the file contents.

Rooted Repositories

To avoid unnecessary repeated files we unified all of the repositories by merging all of the git trees which contained a common root, according to their commit hashes. There is still file repetition when two or more repositories contain copies of a single file, but have no common root.

The figure below shows visually two repositories being merged into a single one, or rooted.

Just to give you an idea of the size of the dataset, let’s compare it to previous datasets:

- Qualitas Corpus Sourcerer GitHub Java Corpus Public Git Archive
N Projects 111 19,233 14,807 182,014
Year of Release 2013 2014 2013 2018
Languages 1 (Java) 1 (Java) 1 (Java) 455
Repeatable No No No Yes
N Files 177K 1.9M 1.5M 54.5M (HEAD)
Lines of Code 37M 320M 352M 15,941M (HEAD)
Storage Size 1.3GB 19GB 14GB 3.0TB

And since an image is worth a thousand tables, here you can see the size of Public Git Archive compared to other datasets. Note that the scale is logarithmic!

Similarly, the number of lines of code show the size of Public Git Archive.

Comparing PGA to GHTorrent and the BigQuery GitHub dataset

Many people asked us already how this dataset compares to two related projects: GHTorrent and BigQuery.

The main difference with GHTorrent is that while GHTorrent only provides metadata about the repositories, Public Git Archive also includes the contents and metadata of the git repository. This means you can actually analyze source code, rather than just metadata about it.

On the other hand, the GitHub Data on BigQuery does contain source code, but only the latest release for each file. While incredibly useful, this is not enough for some analyses where the full history of a repository is needed.

Public Git Archive schema

Public Git Archive provides an index file with information that can be used to select which parts of the dataset one might want to download. This includes:

  • repository URL,
  • name of the siva with the repository contents
  • number of total files, lines of code, and bytes,
  • number of total commits and commits in the HEAD path to the root,
  • list of languages detected in the repository with enry,
  • license detected in the repository with go-license-detector.

You can find the documented schema of the dataset on GitHub.

Each siva file contains one or more GitHub repositories, compressed in an append-friendly format. The repositories contain all the commits, references, branches, and file contents of each repository, similarly to what one would obtain by cloning the repository directly.

Download it now!

You can download the index file, find the siva files corresponding to the repositories of your interest, and download one by one.

To make this task easier we’ve also released a tool that we call pga. pga is able to filter and download sections of the datasets with an easy command line interface. You can download it from the releases page at Alternatively, if you already have Go installed on your machine, you can simply run go get

Once installed listing all of the repositories in the dataset is as simple as running pga list. To list only repositories that contained some Java:

▶️ pga list -l java

, or to show only the repositories under the src-d organization:

▶️ pga list -u /src-d/

To download the data, rather than just listing the repositories, replace list with get.

Finally, you can also use pga to present the index in different formats, such as CSV or JSON:

This can be useful when a more complex filter of the index is necessary. For instance, if you wanted to download all of the source{d} repositories with Apache v2 licenses, you can run the following command:

▶️ pga list -u /src-d/ -f json | grep "Apache-2.0" | jq -r ".sivaFilenames[]" | pga get -i
downloading siva files by name from stdin
filter flags will be ignored
 0 / 5 [------------------------------------------------------]   0.00%

We first list all of the source{d} repositories as JSON, grep those using Apache v2 licenses, then use jq to extract the corresponding filenames, and download them with pga get -i.


The Public Git Archive index is available only for research purposes and under a dual license Open Database License (ODBL) and CC-BY-SA-NC 4.0.

We’d love to hear if you decide to use it. You can get in touch with us via email to or joining our Slack community following the instructions below.


We understand that some GitHub projects may become private or deleted with time. Previous dataset snapshots will continue to include such dead code. If you are the author and want to remove your project from all present and future public snapshots, please send a request to

This post was written by Francesc Campoy. Follow him on Twitter: @francesc.