At Hugging Face, we are on a mission to democratize good Machine Learning and we believe in the value of open source. That’s why we designed 🤗 Datasets so that anyone can share a dataset with the greater ML community. There are currently thousands of datasets in over 100 languages in the Hugging Face Hub, and the Hugging Face team always welcomes new contributions!
Dataset repositories offer features such as:
- Free dataset hosting
- Dataset versioning
- Commit history and diffs
- Metadata for discoverability
- Dataset cards for documentation, licensing, limitations, etc.
This guide will show you how to share a dataset that can be easily accessed by anyone.
You can share your dataset with the community with a dataset repository on the Hugging Face Hub. It can also be a private dataset if you want to control who has access to it.
In a dataset repository, you can either host all your data files and/or use a dataset script.
The dataset script is optional if your dataset is in one of the following formats: CSV, JSON, JSON lines, text or Parquet.
The script also supports many kinds of compressed file types such as: GZ, BZ2, LZ4, LZMA or ZSTD.
For example, your dataset can be made of
On the other hand, if your dataset is not in a supported format or if you want more control over how your dataset is loaded, you can write your own dataset script.
When loading a dataset from the Hub:
- If there’s no dataset script, all the files in the supported formats are loaded.
- If there’s a dataset script, it is downloaded and executed to download and prepare the dataset.
For more information on how to load a dataset from the Hub, see how to load from the load-from-the-hub.
Sharing a community dataset will require you to create an account on hf.co if you don’t have one yet. You can directly create a new dataset repository from your account on the Hugging Face Hub, but this guide will show you how to upload a dataset from the terminal.
- Make sure you are in the virtual environment where you installed Datasets, and run the following command:
- Login using your Hugging Face Hub credentials, and create a new dataset repository:
huggingface-cli repo create your_dataset_name --type dataset
-organization flag to create a repository under a specific organization:
huggingface-cli repo create your_dataset_name --type dataset --organization your-org-name
- Install Git LFS and clone your repository:
# Make sure you have git-lfs installed # (https://git-lfs.github.com/) git lfs install git clone https://huggingface.co/datasets/namespace/your_dataset_name
namespace is either your username or your organization name.
- Now is a good time to check your directory to ensure the only files you’re uploading are:
README.mdis a Dataset card that describes the datasets contents, creation, and usage. To write a Dataset card, see the dataset card page.
The raw data files of the dataset (optional, if they are hosted elsewhere you can specify the URLs in the dataset script). If you don’t need a dataset script, you can take a look at how to structure your dataset repository for your data files.
your_dataset_name.pyis your dataset loading script (optional if your data files are already in the supported formats csv/jsonl/json/parquet/txt). To create a dataset script, see the dataset script page.
dataset_infos.jsoncontains metadata about the dataset (required only if you have a dataset script, or if you want to specify custom feature types).
You can directly upload your files from your repository on the Hugging Face Hub, but this guide will show you how to upload the files from the terminal.
- It is important to add the large data files first with
git lfs trackor else you will encounter an error later when you push your files:
cp /somewhere/data/*.json . git lfs track *.json git add .gitattributes git add *.json git commit -m "add json files"
- Add the dataset loading script and metadata file:
cp /somewhere/data/dataset_infos.json . cp /somewhere/data/load_script.py . git add --all
- Verify the files have been correctly staged. Then you can commit and push your files:
git status git commit -m "First version of the your_dataset_name dataset." git push
Congratulations, your dataset has now been uploaded to the Hugging Face Hub where anyone can load it in a single line of code! 🥳
dataset = load_dataset("namespace/your_dataset_name")
Datasets used to be hosted on our GitHub repository, but all datasets have now been migrated to the Hugging Face Hub. The legacy GitHub datasets were added originally on our GitHub repository and therefore don’t have a namespace: “squad”, “glue”, etc. unlike the other datasets that are named “username/dataset_name” or “org/dataset_name”. Those datasets are still maintained, and if you’d like to edit them, please open a Pull Request on the huggingface/datasets repository. Sharing your dataset to the Hub is the recommended way of adding a dataset.
The distinction between a Hub dataset and a dataset from GitHub only comes from the legacy sharing workflow. It does not involve any ranking, decisioning, or opinion regarding the contents of the dataset itself.
The code of these datasets are reviewed by the Hugging Face team, and they require test data in order to be regularly tested.
In some cases it makes more sense to open a PR on GitHub:
- when you need the dataset to be reviewed
- when you need long-term maintenance from the Hugging Face team
- when there’s no clear org name / namespace that you can put the dataset under
For more info, please take a look at the documentation on How to add a new dataset in the huggingface/datasets repository.