Datasets documentation

Hugging Face Hub

You are viewing v2.0.0 version. A newer version v2.20.0 is available.
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Hugging Face Hub

Now that you are all setup, the first step is to load a dataset. The easiest way to load a dataset is from the Hugging Face Hub. There are already over 900 datasets in over 100 languages on the Hub. Choose from a wide category of datasets to use for NLP tasks like question answering, summarization, machine translation, and language modeling. For a more in-depth look inside a dataset, use the live Datasets Viewer.

Load a dataset

Before you take the time to download a dataset, it is often helpful to quickly get all the relevant information about a dataset. The datasets.load_dataset_builder() method allows you to inspect the attributes of a dataset without downloading it.

>>> from datasets import load_dataset_builder
>>> dataset_builder = load_dataset_builder('imdb')
>>> print(dataset_builder.cache_dir)
>>> print(
{'text': Value(dtype='string', id=None), 'label': ClassLabel(num_classes=2, names=['neg', 'pos'], names_file=None, id=None)}
>>> print(
{'train': SplitInfo(name='train', num_bytes=33432835, num_examples=25000, dataset_name='imdb'), 'test': SplitInfo(name='test', num_bytes=32650697, num_examples=25000, dataset_name='imdb'), 'unsupervised': SplitInfo(name='unsupervised', num_bytes=67106814, num_examples=50000, dataset_name='imdb')}

Take a look at datasets.DatasetInfo for a full list of attributes you can use with dataset_builder.

Once you are happy with the dataset you want, load it in a single line with datasets.load_dataset():

>>> from datasets import load_dataset
>>> dataset = load_dataset('glue', 'mrpc', split='train')

Select a configuration

Some datasets, like the General Language Understanding Evaluation (GLUE) benchmark, are actually made up of several datasets. These sub-datasets are called configurations, and you must explicitly select one when you load the dataset. If you don’t provide a configuration name, πŸ€— Datasets will raise a ValueError and remind you to select a configuration.

Use get_dataset_config_names to retrieve a list of all the possible configurations available to your dataset:

from datasets import get_dataset_config_names

configs = get_dataset_config_names("glue")
# ['cola', 'sst2', 'mrpc', 'qqp', 'stsb', 'mnli', 'mnli_mismatched', 'mnli_matched', 'qnli', 'rte', 'wnli', 'ax']

❌ Incorrect way to load a configuration:

>>> from datasets import load_dataset
>>> dataset = load_dataset('glue')
ValueError: Config name is missing.
Please pick one among the available configs: ['cola', 'sst2', 'mrpc', 'qqp', 'stsb', 'mnli', 'mnli_mismatched', 'mnli_matched', 'qnli', 'rte', 'wnli', 'ax']
Example of usage:
        *load_dataset('glue', 'cola')*

βœ… Correct way to load a configuration:

>>> dataset = load_dataset('glue', 'sst2')
Downloading and preparing dataset glue/sst2 (download: 7.09 MiB, generated: 4.81 MiB, total: 11.90 MiB) to /Users/thomwolf/.cache/huggingface/datasets/glue/sst2/1.0.0...
Downloading: 100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 7.44M/7.44M [00:01<00:00, 7.03MB/s]
Dataset glue downloaded and prepared to /Users/thomwolf/.cache/huggingface/datasets/glue/sst2/1.0.0. Subsequent calls will reuse this data.
>>> print(dataset)
{'train': Dataset(schema: {'sentence': 'string', 'label': 'int64', 'idx': 'int32'}, num_rows: 67349),
    'validation': Dataset(schema: {'sentence': 'string', 'label': 'int64', 'idx': 'int32'}, num_rows: 872),
    'test': Dataset(schema: {'sentence': 'string', 'label': 'int64', 'idx': 'int32'}, num_rows: 1821)

Select a split

A split is a specific subset of the dataset like train and test. Make sure you select a split when you load a dataset. If you don’t supply a split argument, πŸ€— Datasets will only return a dictionary containing the subsets of the dataset.

>>> from datasets import load_dataset
>>> datasets = load_dataset('glue', 'mrpc')
>>> print(datasets)
{train: Dataset({
    features: ['idx', 'label', 'sentence1', 'sentence2'],
    num_rows: 3668
validation: Dataset({
    features: ['idx', 'label', 'sentence1', 'sentence2'],
    num_rows: 408
test: Dataset({
    features: ['idx', 'label', 'sentence1', 'sentence2'],
    num_rows: 1725

You can list the split names for a dataset, or a specific configuration, with the datasets.get_dataset_split_names() method:

>>> from datasets import get_dataset_split_names
>>> get_dataset_split_names('sent_comp')
['validation', 'train']
>>> get_dataset_split_names('glue', 'cola')
['test', 'train', 'validation']