Basic principles

DataLad is designed to be used both as a command-line tool, and as a Python module. The sections Command line reference and Python module reference provide detailed description of the commands and functions of the two interfaces. This section presents common concepts. Although examples will frequently be presented using command line interface commands, all functionality with identically named functions and options are available through Python API as well.


A DataLad dataset is a Git repository that may or may not have a data annex that is used to manage data referenced in a dataset. In practice, most DataLad datasets will come with an annex.

Datasets can contain other datasets (subdatasets), which can in turn contain subdatasets, and so on. There is no limit to the depth of nesting datasets. Each dataset in such a hierarchy has its own annex and its own history. The parent or superdataset only tracks the specific state of a subdataset, and information on where it can be obtained. This is a powerful yet lightweight mechanism for combining multiple individual datasets for a specific purpose, such as the combination of source code repositories with other resources for a tailored application. In many cases DataLad can work with a hierarchy of datasets just as if it were a single dataset.

A superdataset can also be seen as a curated collection of datasets, for example, for a certain data modality, a field of science, a certain author, or some all from one project (maybe the resource for a movie production). The lightweight coupling between super and subdatasets enables scenarios where individual datasets are maintained by a disjoint set of people, and the dataset collection itself can be curated by a completely independent entity. Any individual dataset can be part of any number of such collections.

Benefitting from Git’s support for workflows based on decentralized “clones” of a repository, DataLad’s datasets can be (re-)published to a new location without loosing the connection between the “original” and the new “copy”. This is extremely useful for collaborative work, but also in more mundane scenarios such as data backup, or temporary deployment fo a dataset on a compute cluster, or in the cloud. Using git-annex, data can also get synchronized across different location of a dataset (siblings in DataLad terminology). Using metadata tags, it is even possible to configure different levels of desired data redundancy across the network of dataset, or to prevent publication of sensitive data to publicly accessible repositories. Individual datasets in a hierarchy of (sub)datasets need not be stored at the same location. Continuing with an earlier example, it is possible to post a curated collection of datasets, as a superdataset, on Github, while the actual datasets live on different servers all around the world.

API principles

You can use DataLad’s install command to download datasets. The command accepts URLs of different protocols (http, ssh) as an argument. Nevertheless, the easiest way to obtain a first dataset is downloading the canonical superdataset from using a shortcut.

Downloading DataLad’s canonical superdataset

DataLad’s canonical superdataset provides an automated collection of datasets from various portals and sites (see Automatic creation and maintenance of datasets by crawling external resources). The argument /// can be used as a shortcut that points to the superdataset located at Here are three common examples in command line notation:

datalad install ///
installs the canonical superdataset (metadata without subdatasets) in a subdirectory under the current directory
datalad install -r ///openfmri
installs the openfmri superdataset into an openfmri/ subdirectory. Additionally, the -r flag recursively downloads all metadata of datasets available from as subdatasets into the openfmri/ subdirectory
datalad install -g -J3 -r ///labs/haxby
installs the superdataset of datasets released by the lab of Dr. James V. Haxby and all subdatasets’ metadata. The -g flag indicates getting the actual data, too. It does so by using 3 parallel download processes (-J3 flag).

Downloading datasets via http

In most places where DataLad accepts URLs as arguments these URLs can be regular http or https protocol URLs. For example:

datalad install

Downloading datasets via ssh

DataLad also supports SSH URLs, such as ssh://me@localhost/path.

datalad install ssh://me@localhost/path

Finally, DataLad supports SSH login style resource identifiers, such as me@localhost:/path.

datalad install me@localhost:/path

–dataset argument

All commands which operate with/on datasets (practically all commands) have a dataset argument (-d or --dataset in command line) which takes a path to the dataset that the command should operate on. If a dataset is identified this way then any relative path that is provided as an argument to the command will be interpreted as being relative to the topmost directory of that dataset. If no dataset argument is provided, relative paths are considered to be relative to the current directory.

There are also some useful pre-defined “shortcut” values for dataset arguments:

refers to the “canonical” dataset located under $HOME/datalad/. So running datalad install -d/// crcns will install the crcns subdataset under $HOME/datalad/crcns. This is the same as running datalad install $HOME/datalad/crcns.
topmost superdataset containing the dataset the current directory is part of. For example, if you are in $HOME/datalad/openfmri/ds000001/sub-01 and want to search metadata of the entire superdataset you are under (in this case ///), run datalad search -d^ [something to search].

Commands install vs get

The install and get commands might seem confusingly similar at first. Both of them could be used to install any number of subdatasets, and fetch content of the data files. Differences lie primarily in their default behaviour and outputs, and thus intended use. Both install and get take local paths as their arguments, but their default behavior and output might differ;

  • install primarily operates and reports at the level of datasets, and returns as a result dataset(s) which either were just installed, or were installed previously already under specified locations. So result should be the same if the same install command ran twice on the same datasets. It does not fetch data files by default
  • get primarily operates at the level of paths (datasets, directories, and/or files). As a result it returns only what was installed (datasets) or fetched (files). So result of rerunning the same get command should report that nothing new was installed or fetched. It fetches data files by default.

In how both commands operate on provided paths, it could be said that install == get -n, and install -g == get. But install also has ability to install new datasets from remote locations given their URLs (e.g., for our super-dataset) and SSH targets (e.g., [login@]host:path) if they are provided as the argument to its call or explicitly as --source option. If datalad install --source URL DESTINATION (command line example) is used, then dataset from URL gets installed under PATH. In case of datalad install URL invocation, PATH is taken from the last name within URL similar to how git clone does it. If former specification allows to specify only a single URL and a PATH at a time, later one can take multiple remote locations from which datasets could be installed.

So, as a rule of thumb – if you want to install from external URL or fetch a sub-dataset without downloading data files stored under annex – use install. In Python API install is also to be used when you want to receive in output the corresponding Dataset object to operate on, and be able to use it even if you rerun the script. If you would like to fetch data (possibly while installing any necessary to be installed sub-dataset to get to the file) – use get.