Create a dataset loading script

The dataset script is likely not needed if your dataset is in one of the following formats: CSV, JSON, JSON lines, text or Parquet. With those formats, you should be able to load your dataset automatically with load_dataset(), as long as your dataset repository has a required structure.

Write a dataset script to load and share datasets that consist of data files in unsupported formats or require more complex data preparation. This is a more advanced way to define a dataset than using YAML metadata in the dataset card. A dataset script is a Python file that defines the different configurations and splits of your dataset, as well as how to download and process the data.

The script can download data files from any website, or from the same dataset repository.

A dataset loading script should have the same name as a dataset repository or directory. For example, a repository named my_dataset should contain my_dataset.py script. This way it can be loaded with:

my_dataset/
├── README.md
└── my_dataset.py

>>> from datasets import load_dataset
>>> load_dataset("path/to/my_dataset")

The following guide includes instructions for dataset scripts for how to:

Add dataset metadata.
Download data files.
Generate samples.
Generate dataset metadata.
Upload a dataset to the Hub.

Open the SQuAD dataset loading script template to follow along on how to share a dataset.

To help you get started, try beginning with the dataset loading script template!

Add dataset attributes

The first step is to add some information, or attributes, about your dataset in DatasetBuilder._info(). The most important attributes you should specify are:

DatasetInfo.description provides a concise description of your dataset. The description informs the user what’s in the dataset, how it was collected, and how it can be used for a NLP task.
DatasetInfo.features defines the name and type of each column in your dataset. This will also provide the structure for each example, so it is possible to create nested subfields in a column if you want. Take a look at Features for a full list of feature types you can use.

datasets.Features(
    {
        "id": datasets.Value("string"),
        "title": datasets.Value("string"),
        "context": datasets.Value("string"),
        "question": datasets.Value("string"),
        "answers": datasets.Sequence(
            {
                "text": datasets.Value("string"),
                "answer_start": datasets.Value("int32"),
            }
        ),
    }
)

DatasetInfo.homepage contains the URL to the dataset homepage so users can find more details about the dataset.
DatasetInfo.citation contains a BibTeX citation for the dataset.

After you’ve filled out all these fields in the template, it should look like the following example from the SQuAD loading script:

def _info(self):
    return datasets.DatasetInfo(
        description=_DESCRIPTION,
        features=datasets.Features(
            {
                "id": datasets.Value("string"),
                "title": datasets.Value("string"),
                "context": datasets.Value("string"),
                "question": datasets.Value("string"),
                "answers": datasets.features.Sequence(
                    {"text": datasets.Value("string"), "answer_start": datasets.Value("int32"),}
                ),
            }
        ),
        # No default supervised_keys (as we have to pass both question
        # and context as input).
        supervised_keys=None,
        homepage="https://rajpurkar.github.io/SQuAD-explorer/",
        citation=_CITATION,
    )

Multiple configurations

In some cases, your dataset may have multiple configurations. For example, the SuperGLUE dataset is a collection of 5 datasets designed to evaluate language understanding tasks. 🤗 Datasets provides BuilderConfig which allows you to create different configurations for the user to select from.

Let’s study the SuperGLUE loading script to see how you can define several configurations.

Create a BuilderConfig subclass with attributes about your dataset. These attributes can be the features of your dataset, label classes, and a URL to the data files.

class SuperGlueConfig(datasets.BuilderConfig):
    """BuilderConfig for SuperGLUE."""

    def __init__(self, features, data_url, citation, url, label_classes=("False", "True"), **kwargs):
        """BuilderConfig for SuperGLUE.

        Args:
        features: *list[string]*, list of the features that will appear in the
            feature dict. Should not include "label".
        data_url: *string*, url to download the zip file from.
        citation: *string*, citation for the data set.
        url: *string*, url for information about the data set.
        label_classes: *list[string]*, the list of classes for the label if the
            label is present as a string. Non-string labels will be cast to either
            'False' or 'True'.
        **kwargs: keyword arguments forwarded to super.
        """
        # Version history:
        # 1.0.2: Fixed non-nondeterminism in ReCoRD.
        # 1.0.1: Change from the pre-release trial version of SuperGLUE (v1.9) to
        #        the full release (v2.0).
        # 1.0.0: S3 (new shuffling, sharding and slicing mechanism).
        # 0.0.2: Initial version.
        super().__init__(version=datasets.Version("1.0.2"), **kwargs)
        self.features = features
        self.label_classes = label_classes
        self.data_url = data_url
        self.citation = citation
        self.url = url

Create instances of your config to specify the values of the attributes of each configuration. This gives you the flexibility to specify all the name and description of each configuration. These sub-class instances should be listed under DatasetBuilder.BUILDER_CONFIGS:

class SuperGlue(datasets.GeneratorBasedBuilder):
    """The SuperGLUE benchmark."""

    BUILDER_CONFIG_CLASS = SuperGlueConfig

    BUILDER_CONFIGS = [
        SuperGlueConfig(
            name="boolq",
            description=_BOOLQ_DESCRIPTION,
            features=["question", "passage"],
            data_url="https://dl.fbaipublicfiles.com/glue/superglue/data/v2/BoolQ.zip",
            citation=_BOOLQ_CITATION,
            url="https://github.com/google-research-datasets/boolean-questions",
        ),
        ...
        ...
        SuperGlueConfig(
            name="axg",
            description=_AXG_DESCRIPTION,
            features=["premise", "hypothesis"],
            label_classes=["entailment", "not_entailment"],
            data_url="https://dl.fbaipublicfiles.com/glue/superglue/data/v2/AX-g.zip",
            citation=_AXG_CITATION,
            url="https://github.com/rudinger/winogender-schemas",
        ),

Now, users can load a specific configuration of the dataset with the configuration name:

>>> from datasets import load_dataset
>>> dataset = load_dataset('super_glue', 'boolq')

Additionally, users can instantiate a custom builder configuration by passing the builder configuration arguments to load_dataset():

>>> from datasets import load_dataset
>>> dataset = load_dataset('super_glue', data_url="https://custom_url")

Default configurations

Users must specify a configuration name when they load a dataset with multiple configurations. Otherwise, 🤗 Datasets will raise a ValueError, and prompt the user to select a configuration name. You can avoid this by setting a default dataset configuration with the DEFAULT_CONFIG_NAME attribute:

class NewDataset(datasets.GeneratorBasedBuilder):

VERSION = datasets.Version("1.1.0")

BUILDER_CONFIGS = [
    datasets.BuilderConfig(name="first_domain", version=VERSION, description="This part of my dataset covers a first domain"),
    datasets.BuilderConfig(name="second_domain", version=VERSION, description="This part of my dataset covers a second domain"),
]

DEFAULT_CONFIG_NAME = "first_domain"

Only use a default configuration when it makes sense. Don’t set one because it may be more convenient for the user to not specify a configuration when they load your dataset. For example, multi-lingual datasets often have a separate configuration for each language. An appropriate default may be an aggregated configuration that loads all the languages of the dataset if the user doesn’t request a particular one.

Download data files and organize splits

After you’ve defined the attributes of your dataset, the next step is to download the data files and organize them according to their splits.

Create a dictionary of URLs in the loading script that point to the original SQuAD data files:

_URL = "https://rajpurkar.github.io/SQuAD-explorer/dataset/"
_URLS = {
    "train": _URL + "train-v1.1.json",
    "dev": _URL + "dev-v1.1.json",
}

If the data files live in the same folder or repository of the dataset script, you can just pass the relative paths to the files instead of URLs.

DownloadManager.download_and_extract() takes this dictionary and downloads the data files. Once the files are downloaded, use SplitGenerator to organize each split in the dataset. This is a simple class that contains:
- The name of each split. You should use the standard split names: Split.TRAIN, Split.TEST, and Split.VALIDATION.
- gen_kwargs provides the file paths to the data files to load for each split.

Your DatasetBuilder._split_generator() should look like this now:

def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
    urls_to_download = self._URLS
    downloaded_files = dl_manager.download_and_extract(urls_to_download)

    return [
        datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"filepath": downloaded_files["train"]}),
        datasets.SplitGenerator(name=datasets.Split.VALIDATION, gen_kwargs={"filepath": downloaded_files["dev"]}),
    ]

Generate samples

At this point, you have:

Added the dataset attributes.
Provided instructions for how to download the data files.
Organized the splits.

The next step is to actually generate the samples in each split.

DatasetBuilder._generate_examples takes the file path provided by gen_kwargs to read and parse the data files. You need to write a function that loads the data files and extracts the columns.
Your function should yield a tuple of an id_, and an example from the dataset.

def _generate_examples(self, filepath):
    """This function returns the examples in the raw (text) form."""
    logger.info("generating examples from = %s", filepath)
    with open(filepath) as f:
        squad = json.load(f)
        for article in squad["data"]:
            title = article.get("title", "").strip()
            for paragraph in article["paragraphs"]:
                context = paragraph["context"].strip()
                for qa in paragraph["qas"]:
                    question = qa["question"].strip()
                    id_ = qa["id"]

                    answer_starts = [answer["answer_start"] for answer in qa["answers"]]
                    answers = [answer["text"].strip() for answer in qa["answers"]]

                    # Features currently used are "context", "question", and "answers".
                    # Others are extracted here for the ease of future expansions.
                    yield id_, {
                        "title": title,
                        "context": context,
                        "question": question,
                        "id": id_,
                        "answers": {"answer_start": answer_starts, "text": answers,},
                    }

(Optional) Generate dataset metadata

Adding dataset metadata is a great way to include information about your dataset. The metadata is stored in the dataset card README.md in YAML. It includes information like the number of examples required to confirm the dataset was correctly generated, and information about the dataset like its features.

Run the following command to generate your dataset metadata in README.md and make sure your new dataset loading script works correctly:

datasets-cli test path/to/<your-dataset-loading-script> --save_info --all_configs

If your dataset loading script passed the test, you should now have a README.md file in your dataset folder containing a dataset_info field with some metadata.

Upload to the Hub

Once your script is ready, create a dataset card and upload it to the Hub.

Congratulations, you can now load your dataset from the Hub! 🥳

>>> from datasets import load_dataset
>>> load_dataset("<username>/my_dataset")

Advanced features

Sharding

If your dataset is made of many big files, 🤗 Datasets automatically runs your script in parallel to make it super fast! It can help if you have hundreds or thousands of TAR archives, or JSONL files like oscar for example.

To make it work, we consider lists of files in gen_kwargs to be shards. Therefore 🤗 Datasets can automatically spawn several workers to run _generate_examples in parallel, and each worker is given a subset of shards to process.


class MyShardedDataset(datasets.GeneratorBasedBuilder):

    def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
        downloaded_files = dl_manager.download([f"data/shard_{i}.jsonl" for i in range(1024)])
        return [
            datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"filepaths": downloaded_files}),
        ]

    def _generate_examples(self, filepaths):
        # Each worker can be given a slice of the original `filepaths` list defined in the `gen_kwargs`
        # so that this code can run in parallel on several shards at the same time
        for filepath in filepaths:
            ...

Users can also specify num_proc= in load_dataset() to specify the number of processes to use as workers.

ArrowBasedBuilder

For some datasets it can be much faster to yield batches of data rather than examples one by one. You can speed up the dataset generation by yielding Arrow tables directly, instead of examples. This is especially useful if your data comes from Pandas DataFrames for example, since the conversion from Pandas to Arrow is as simple as:

import pyarrow as pa
pa_table = pa.Table.from_pandas(df)

To yield Arrow tables instead of single examples, make your dataset builder inherit from ArrowBasedBuilder instead of GeneratorBasedBuilder, and use _generate_tables instead of _generate_examples:

class MySuperFastDataset(datasets.ArrowBasedBuilder):

    def _generate_tables(self, filepaths):
        idx = 0
        for filepath in filepaths:
            ...
            yield idx, pa_table
            idx += 1

Don’t forget to keep your script memory efficient, in case users run them on machines with a low amount of RAM.