Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Cloud storage

🤗 Datasets supports access to cloud storage providers through a fsspec FileSystem implementations. You can save and load datasets from any cloud storage in a Pythonic way. Take a look at the following table for some example of supported cloud storage providers:

Storage provider Filesystem implementation
Amazon S3 s3fs
Google Cloud Storage gcsfs
Azure Blob/DataLake adlfs
Dropbox dropboxdrivefs
Google Drive gdrivefs

This guide will show you how to save and load datasets with any cloud storage. Here are examples for S3, Google Cloud Storage and Azure Blob Storage.

Set up your cloud storage FileSystem

Amazon S3

  1. Install the S3 dependency with 🤗 Datasets:
>>> pip install datasets[s3]
  1. Define your credentials

To use an anonymous connection, use anon=True. Otherwise, include your aws_access_key_id and aws_secret_access_key whenever you are interacting with a private S3 bucket.

>>> storage_options = {"anon": True}  # for anonymous connection
# or use your credentials
>>> storage_options = {"key": aws_access_key_id, "secret": aws_secret_access_key}  # for private buckets
# or use a botocore session
>>> import botocore
>>> s3_session = botocore.session.Session(profile="my_profile_name")
>>> storage_options = {"session": s3_session}
  1. Create your FileSystem instance
>>> import s3fs
>>> fs = s3fs.S3FileSystem(**storage_options)

Google Cloud Storage

  1. Install the Google Cloud Storage implementation:
>>> conda install -c conda-forge gcsfs
# or install with pip
>>> pip install gcsfs
  1. Define your credentials
>>> storage_options={"token": "anon"}  # for anonymous connection
# or use your credentials of your default gcloud credentials or from the google metadata service
>>> storage_options={"project": "my-google-project"}
# or use your credentials from elsewhere, see the documentation at
>>> storage_options={"project": "my-google-project", "token": TOKEN}
  1. Create your FileSystem instance
>>> import gcsfs
>>> fs = gcsfs.GCSFileSystem(**storage_options)

Azure Blob Storage

  1. Install the Azure Blob Storage implementation:
>>> conda install -c conda-forge adlfs
# or install with pip
>>> pip install adlfs
  1. Define your credentials
>>> storage_options = {"anon": True}  # for anonymous connection
# or use your credentials
>>> storage_options = {"account_name": ACCOUNT_NAME, "account_key": ACCOUNT_KEY)  # gen 2 filesystem
# or use your credentials with the gen 1 filesystem
>>> storage_options={"tenant_id": TENANT_ID, "client_id": CLIENT_ID, "client_secret": CLIENT_SECRET}
  1. Create your FileSystem instance
>>> import adlfs
>>> fs = adlfs.AzureBlobFileSystem(**storage_options)

Load and Save your datasets using your cloud storage FileSystem

Download and prepare a dataset into a cloud storage

You can download and prepare a dataset into your cloud storage by specifying a remote output_dir in download_and_prepare. Don’t forget to use the previously defined storage_options containing your credentials to write into a private cloud storage.

The download_and_prepare method works in two steps:

  1. it first downloads the raw data files (if any) in your local cache. You can set your cache directory by passing cache_dir to load_dataset_builder()
  2. then it generates the dataset in Arrow or Parquet format in your cloud storage by iterating over the raw data files.

Load a dataset builder from the Hugging Face Hub (see how to load from the Hugging Face Hub):

>>> output_dir = "s3://my-bucket/imdb"
>>> builder = load_dataset_builder("imdb")
>>> builder.download_and_prepare(output_dir, storage_options=storage_options, file_format="parquet")

Load a dataset builder using a loading script (see how to load a local loading script):

>>> output_dir = "s3://my-bucket/imdb"
>>> builder = load_dataset_builder("path/to/local/loading_script/")
>>> builder.download_and_prepare(output_dir, storage_options=storage_options, file_format="parquet")

Use your own data files (see how to load local and remote files):

>>> data_files = {"train": ["path/to/train.csv"]}
>>> output_dir = "s3://my-bucket/imdb"
>>> builder = load_dataset_builder("csv", data_files=data_files)
>>> builder.download_and_prepare(output_dir, storage_options=storage_options, file_format="parquet")

It is highly recommended to save the files as compressed Parquet files to optimize I/O by specifying file_format="parquet". Otherwise the dataset is saved as an uncompressed Arrow file.

You can also specify the size of the Parquet shard using max_shard_size (default is 500MB):

>>> builder.download_and_prepare(output_dir, storage_options=storage_options, file_format="parquet", max_shard_size="1GB")


Dask is a parallel computing library and it has a pandas-like API for working with larger than memory Parquet datasets in parallel. Dask can use multiple threads or processes on a single machine, or a cluster of machines to process data in parallel. Dask supports local data but also data from a cloud storage.

Therefore you can load a dataset saved as sharded Parquet files in Dask with

import dask.dataframe as dd

df = dd.read_parquet(output_dir, storage_options=storage_options)

# or if your dataset is split into train/valid/test
df_train = dd.read_parquet(output_dir + f"/{}-train-*.parquet", storage_options=storage_options)
df_valid = dd.read_parquet(output_dir + f"/{}-validation-*.parquet", storage_options=storage_options)
df_test = dd.read_parquet(output_dir + f"/{}-test-*.parquet", storage_options=storage_options)

You can find more about dask dataframes in their documentation.

Saving serialized datasets

After you have processed your dataset, you can save it to your cloud storage with Dataset.save_to_disk():

# saves encoded_dataset to amazon s3
>>> encoded_dataset.save_to_disk("s3://my-private-datasets/imdb/train", fs=fs)
# saves encoded_dataset to google cloud storage
>>> encoded_dataset.save_to_disk("gcs://my-private-datasets/imdb/train", fs=fs)
# saves encoded_dataset to microsoft azure blob/datalake
>>> encoded_dataset.save_to_disk("adl://my-private-datasets/imdb/train", fs=fs)

Remember to define your credentials in your FileSystem instance fs whenever you are interacting with a private cloud storage.

Listing serialized datasets

List files from a cloud storage with your FileSystem instance fs, using


Load serialized datasets

When you are ready to use your dataset again, reload it with Dataset.load_from_disk():

>>> from datasets import load_from_disk
# load encoded_dataset from cloud storage
>>> dataset = load_from_disk("s3://a-public-datasets/imdb/train", fs=fs)  
>>> print(len(dataset))