Hugging Face's logo

Datasets documentation

Main classes

You are viewing v2.21.0 version. A newer version v3.2.0 is available.
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Main classes

DatasetInfo

class datasets.DatasetInfo

< >

( description: str = <factory> citation: str = <factory> homepage: str = <factory> license: str = <factory> features: Optional = None post_processed: Optional = None supervised_keys: Optional = None task_templates: Optional = None builder_name: Optional = None dataset_name: Optional = None config_name: Optional = None version: Union = None splits: Optional = None download_checksums: Optional = None download_size: Optional = None post_processing_size: Optional = None dataset_size: Optional = None size_in_bytes: Optional = None )

Parameters

  • description (str) — A description of the dataset.
  • citation (str) — A BibTeX citation of the dataset.
  • homepage (str) — A URL to the official homepage for the dataset.
  • license (str) — The dataset’s license. It can be the name of the license or a paragraph containing the terms of the license.
  • features (Features, optional) — The features used to specify the dataset’s column types.
  • post_processed (PostProcessedInfo, optional) — Information regarding the resources of a possible post-processing of a dataset. For example, it can contain the information of an index.
  • supervised_keys (SupervisedKeysData, optional) — Specifies the input feature and the label for supervised learning if applicable for the dataset (legacy from TFDS).
  • builder_name (str, optional) — The name of the GeneratorBasedBuilder subclass used to create the dataset. Usually matched to the corresponding script name. It is also the snake_case version of the dataset builder class name.
  • config_name (str, optional) — The name of the configuration derived from BuilderConfig.
  • version (str or Version, optional) — The version of the dataset.
  • splits (dict, optional) — The mapping between split name and metadata.
  • download_checksums (dict, optional) — The mapping between the URL to download the dataset’s checksums and corresponding metadata.
  • download_size (int, optional) — The size of the files to download to generate the dataset, in bytes.
  • post_processing_size (int, optional) — Size of the dataset in bytes after post-processing, if any.
  • dataset_size (int, optional) — The combined size in bytes of the Arrow tables for all splits.
  • size_in_bytes (int, optional) — The combined size in bytes of all files associated with the dataset (downloaded files + Arrow files).
  • task_templates (List[TaskTemplate], optional) — The task templates to prepare the dataset for during training and evaluation. Each template casts the dataset’s Features to standardized column names and types as detailed in datasets.tasks.
  • **config_kwargs (additional keyword arguments) — Keyword arguments to be passed to the BuilderConfig and used in the DatasetBuilder.

Information about a dataset.

DatasetInfo documents datasets, including its name, version, and features. See the constructor arguments and properties for a full list.

Not all fields are known on construction and may be updated later.

from_directory

< >

( dataset_info_dir: str fs = 'deprecated' storage_options: Optional = None )

Parameters

  • dataset_info_dir (str) — The directory containing the metadata file. This should be the root directory of a specific dataset version.
  • fs (fsspec.spec.AbstractFileSystem, optional) — Instance of the remote filesystem used to download the files from.

    Deprecated in 2.9.0

    fs was deprecated in version 2.9.0 and will be removed in 3.0.0. Please use storage_options instead, e.g. storage_options=fs.storage_options.
  • storage_options (dict, optional) — Key/value pairs to be passed on to the file-system backend, if any.

    Added in 2.9.0

Create DatasetInfo from the JSON file in dataset_info_dir.

This function updates all the dynamically generated fields (num_examples, hash, time of creation,…) of the DatasetInfo.

This will overwrite all previous metadata.

Example:

>>> from datasets import DatasetInfo
>>> ds_info = DatasetInfo.from_directory("/path/to/directory/")

write_to_directory

< >

( dataset_info_dir pretty_print = False fs = 'deprecated' storage_options: Optional = None )

Parameters

  • dataset_info_dir (str) — Destination directory.
  • pretty_print (bool, defaults to False) — If True, the JSON will be pretty-printed with the indent level of 4.
  • fs (fsspec.spec.AbstractFileSystem, optional) — Instance of the remote filesystem used to download the files from.

    Deprecated in 2.9.0

    fs was deprecated in version 2.9.0 and will be removed in 3.0.0. Please use storage_options instead, e.g. storage_options=fs.storage_options.
  • storage_options (dict, optional) — Key/value pairs to be passed on to the file-system backend, if any.

    Added in 2.9.0

Write DatasetInfo and license (if present) as JSON files to dataset_info_dir.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.info.write_to_directory("/path/to/directory/")

Dataset

The base class Dataset implements a Dataset backed by an Apache Arrow table.

class datasets.Dataset

< >

( arrow_table: Table info: Optional = None split: Optional = None indices_table: Optional = None fingerprint: Optional = None )

A Dataset backed by an Arrow table.

add_column

< >

( name: str column: Union new_fingerprint: str )

Parameters

  • name (str) — Column name.
  • column (list or np.array) — Column data to be added.

Add column to Dataset.

Added in 1.7

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> more_text = ds["text"]
>>> ds.add_column(name="text_2", column=more_text)
Dataset({
    features: ['text', 'label', 'text_2'],
    num_rows: 1066
})

add_item

< >

( item: dict new_fingerprint: str )

Parameters

  • item (dict) — Item data to be added.

Add item to Dataset.

Added in 1.7

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> new_review = {'label': 0, 'text': 'this movie is the absolute worst thing I have ever seen'}
>>> ds = ds.add_item(new_review)
>>> ds[-1]
{'label': 0, 'text': 'this movie is the absolute worst thing I have ever seen'}

from_file

< >

( filename: str info: Optional = None split: Optional = None indices_filename: Optional = None in_memory: bool = False )

Parameters

  • filename (str) — File name of the dataset.
  • info (DatasetInfo, optional) — Dataset information, like description, citation, etc.
  • split (NamedSplit, optional) — Name of the dataset split.
  • indices_filename (str, optional) — File names of the indices.
  • in_memory (bool, defaults to False) — Whether to copy the data in-memory.

Instantiate a Dataset backed by an Arrow table at filename.

from_buffer

< >

( buffer: Buffer info: Optional = None split: Optional = None indices_buffer: Optional = None )

Parameters

  • buffer (pyarrow.Buffer) — Arrow buffer.
  • info (DatasetInfo, optional) — Dataset information, like description, citation, etc.
  • split (NamedSplit, optional) — Name of the dataset split.
  • indices_buffer (pyarrow.Buffer, optional) — Indices Arrow buffer.

Instantiate a Dataset backed by an Arrow buffer.

from_pandas

< >

( df: DataFrame features: Optional = None info: Optional = None split: Optional = None preserve_index: Optional = None )

Parameters

  • df (pandas.DataFrame) — Dataframe that contains the dataset.
  • features (Features, optional) — Dataset features.
  • info (DatasetInfo, optional) — Dataset information, like description, citation, etc.
  • split (NamedSplit, optional) — Name of the dataset split.
  • preserve_index (bool, optional) — Whether to store the index as an additional column in the resulting Dataset. The default of None will store the index as a column, except for RangeIndex which is stored as metadata only. Use preserve_index=True to force it to be stored as a column.

Convert pandas.DataFrame to a pyarrow.Table to create a Dataset.

The column types in the resulting Arrow Table are inferred from the dtypes of the pandas.Series in the DataFrame. In the case of non-object Series, the NumPy dtype is translated to its Arrow equivalent. In the case of object, we need to guess the datatype by looking at the Python objects in this Series.

Be aware that Series of the object dtype don’t carry enough information to always lead to a meaningful Arrow type. In the case that we cannot infer a type, e.g. because the DataFrame is of length 0 or the Series only contains None/nan objects, the type is set to null. This behavior can be avoided by constructing explicit features and passing it to this function.

Example:

>>> ds = Dataset.from_pandas(df)

from_dict

< >

( mapping: dict features: Optional = None info: Optional = None split: Optional = None )

Parameters

  • mapping (Mapping) — Mapping of strings to Arrays or Python lists.
  • features (Features, optional) — Dataset features.
  • info (DatasetInfo, optional) — Dataset information, like description, citation, etc.
  • split (NamedSplit, optional) — Name of the dataset split.

Convert dict to a pyarrow.Table to create a Dataset.

from_generator

< >

( generator: Callable features: Optional = None cache_dir: str = None keep_in_memory: bool = False gen_kwargs: Optional = None num_proc: Optional = None split: NamedSplit = NamedSplit('train') **kwargs )

Parameters

  • generator ( —Callable): A generator function that yields examples.
  • features (Features, optional) — Dataset features.
  • cache_dir (str, optional, defaults to "~/.cache/huggingface/datasets") — Directory to cache data.
  • keep_in_memory (bool, defaults to False) — Whether to copy the data in-memory.
  • gen_kwargs(dict, optional) — Keyword arguments to be passed to the generator callable. You can define a sharded dataset by passing the list of shards in gen_kwargs and setting num_proc greater than 1.
  • num_proc (int, optional, defaults to None) — Number of processes when downloading and generating the dataset locally. This is helpful if the dataset is made of multiple files. Multiprocessing is disabled by default. If num_proc is greater than one, then all list values in gen_kwargs must be the same length. These values will be split between calls to the generator. The number of shards will be the minimum of the shortest list in gen_kwargs and num_proc.

    Added in 2.7.0
  • split (NamedSplit, defaults to Split.TRAIN) — Split name to be assigned to the dataset.

    Added in 2.21.0
  • **kwargs (additional keyword arguments) — Keyword arguments to be passed to :GeneratorConfig.

Create a Dataset from a generator.

Example:

>>> def gen():
...     yield {"text": "Good", "label": 0}
...     yield {"text": "Bad", "label": 1}
...
>>> ds = Dataset.from_generator(gen)
>>> def gen(shards):
...     for shard in shards:
...         with open(shard) as f:
...             for line in f:
...                 yield {"line": line}
...
>>> shards = [f"data{i}.txt" for i in range(32)]
>>> ds = Dataset.from_generator(gen, gen_kwargs={"shards": shards})

data

< >

( )

The Apache Arrow table backing the dataset.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.data
MemoryMappedTable
text: string
label: int64
----
text: [["compassionately explores the seemingly irreconcilable situation between conservative christian parents and their estranged gay and lesbian children .","the soundtrack alone is worth the price of admission .","rodriguez does a splendid job of racial profiling hollywood style--casting excellent latin actors of all ages--a trend long overdue .","beneath the film's obvious determination to shock at any cost lies considerable skill and determination , backed by sheer nerve .","bielinsky is a filmmaker of impressive talent .","so beautifully acted and directed , it's clear that washington most certainly has a new career ahead of him if he so chooses .","a visual spectacle full of stunning images and effects .","a gentle and engrossing character study .","it's enough to watch huppert scheming , with her small , intelligent eyes as steady as any noir villain , and to enjoy the perfectly pitched web of tension that chabrol spins .","an engrossing portrait of uncompromising artists trying to create something original against the backdrop of a corporate music industry that only seems to care about the bottom line .",...,"ultimately , jane learns her place as a girl , softens up and loses some of the intensity that made her an interesting character to begin with .","ah-nuld's action hero days might be over .","it's clear why deuces wild , which was shot two years ago , has been gathering dust on mgm's shelf .","feels like nothing quite so much as a middle-aged moviemaker's attempt to surround himself with beautiful , half-naked women .","when the precise nature of matthew's predicament finally comes into sharp focus , the revelation fails to justify the build-up .","this picture is murder by numbers , and as easy to be bored by as your abc's , despite a few whopping shootouts .","hilarious musical comedy though stymied by accents thick as mud .","if you are into splatter movies , then you will probably have a reasonably good time with the salton sea .","a dull , simple-minded and stereotypical tale of drugs , death and mind-numbing indifference on the inner-city streets .","the feature-length stretch . . . strains the show's concept ."]]
label: [[1,1,1,1,1,1,1,1,1,1,...,0,0,0,0,0,0,0,0,0,0]]

cache_files

< >

( )

The cache files containing the Apache Arrow table backing the dataset.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.cache_files
[{'filename': '/root/.cache/huggingface/datasets/rotten_tomatoes_movie_review/default/1.0.0/40d411e45a6ce3484deed7cc15b82a53dad9a72aafd9f86f8f227134bec5ca46/rotten_tomatoes_movie_review-validation.arrow'}]

num_columns

< >

( )

Number of columns in the dataset.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.num_columns
2

num_rows

< >

( )

Number of rows in the dataset (same as Dataset.len()).

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.num_rows
1066

column_names

< >

( )

Names of the columns in the dataset.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.column_names
['text', 'label']

shape

< >

( )

Shape of the dataset (number of columns, number of rows).

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.shape
(1066, 2)

unique

< >

( column: str ) → list

Parameters

  • column (str) — Column name (list all the column names with column_names).

Returns

list

List of unique elements in the given column.

Return a list of the unique elements in a column.

This is implemented in the low-level backend and as such, very fast.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.unique('label')
[1, 0]

flatten

< >

( new_fingerprint: Optional = None max_depth = 16 ) → Dataset

Parameters

  • new_fingerprint (str, optional) — The new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments.

Returns

Dataset

A copy of the dataset with flattened columns.

Flatten the table. Each column with a struct type is flattened into one column per struct field. Other columns are left unchanged.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("squad", split="train")
>>> ds.features
{'answers': Sequence(feature={'text': Value(dtype='string', id=None), 'answer_start': Value(dtype='int32', id=None)}, length=-1, id=None),
 'context': Value(dtype='string', id=None),
 'id': Value(dtype='string', id=None),
 'question': Value(dtype='string', id=None),
 'title': Value(dtype='string', id=None)}
>>> ds.flatten()
Dataset({
    features: ['id', 'title', 'context', 'question', 'answers.text', 'answers.answer_start'],
    num_rows: 87599
})

cast

< >

( features: Features batch_size: Optional = 1000 keep_in_memory: bool = False load_from_cache_file: Optional = None cache_file_name: Optional = None writer_batch_size: Optional = 1000 num_proc: Optional = None ) → Dataset

Parameters

  • features (Features) — New features to cast the dataset to. The name of the fields in the features must match the current column names. The type of the data must also be convertible from one type to the other. For non-trivial conversion, e.g. str <-> ClassLabel you should use map() to update the Dataset.
  • batch_size (int, defaults to 1000) — Number of examples per batch provided to cast. If batch_size <= 0 or batch_size == None then provide the full dataset as a single batch to cast.
  • keep_in_memory (bool, defaults to False) — Whether to copy the data in-memory.
  • load_from_cache_file (bool, defaults to True if caching is enabled) — If a cache file storing the current computation from function can be identified, use it instead of recomputing.
  • cache_file_name (str, optional, defaults to None) — Provide the name of a path for the cache file. It is used to store the results of the computation instead of the automatically generated cache file name.
  • writer_batch_size (int, defaults to 1000) — Number of rows per write operation for the cache file writer. This value is a good trade-off between memory usage during the processing, and processing speed. Higher value makes the processing do fewer lookups, lower value consume less temporary memory while running map().
  • num_proc (int, optional, defaults to None) — Number of processes for multiprocessing. By default it doesn’t use multiprocessing.

Returns

Dataset

A copy of the dataset with casted features.

Cast the dataset to a new set of features.

Example:

>>> from datasets import load_dataset, ClassLabel, Value
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.features
{'label': ClassLabel(num_classes=2, names=['neg', 'pos'], id=None),
 'text': Value(dtype='string', id=None)}
>>> new_features = ds.features.copy()
>>> new_features['label'] = ClassLabel(names=['bad', 'good'])
>>> new_features['text'] = Value('large_string')
>>> ds = ds.cast(new_features)
>>> ds.features
{'label': ClassLabel(num_classes=2, names=['bad', 'good'], id=None),
 'text': Value(dtype='large_string', id=None)}

cast_column

< >

( column: str feature: Union new_fingerprint: Optional = None )

Parameters

  • column (str) — Column name.
  • feature (FeatureType) — Target feature.
  • new_fingerprint (str, optional) — The new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments.

Cast column to feature for decoding.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.features
{'label': ClassLabel(num_classes=2, names=['neg', 'pos'], id=None),
 'text': Value(dtype='string', id=None)}
>>> ds = ds.cast_column('label', ClassLabel(names=['bad', 'good']))
>>> ds.features
{'label': ClassLabel(num_classes=2, names=['bad', 'good'], id=None),
 'text': Value(dtype='string', id=None)}

remove_columns

< >

( column_names: Union new_fingerprint: Optional = None ) → Dataset

Parameters

  • column_names (Union[str, List[str]]) — Name of the column(s) to remove.
  • new_fingerprint (str, optional) — The new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments.

Returns

Dataset

A copy of the dataset object without the columns to remove.

Remove one or several column(s) in the dataset and the features associated to them.

You can also remove a column using map() with remove_columns but the present method doesn’t copy the data of the remaining columns and is thus faster.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds = ds.remove_columns('label')
Dataset({
    features: ['text'],
    num_rows: 1066
})
>>> ds = ds.remove_columns(column_names=ds.column_names) # Removing all the columns returns an empty dataset with the `num_rows` property set to 0
Dataset({
    features: [],
    num_rows: 0
})

rename_column

< >

( original_column_name: str new_column_name: str new_fingerprint: Optional = None ) → Dataset

Parameters

  • original_column_name (str) — Name of the column to rename.
  • new_column_name (str) — New name for the column.
  • new_fingerprint (str, optional) — The new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments.

Returns

Dataset

A copy of the dataset with a renamed column.

Rename a column in the dataset, and move the features associated to the original column under the new column name.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds = ds.rename_column('label', 'label_new')
Dataset({
    features: ['text', 'label_new'],
    num_rows: 1066
})

rename_columns

< >

( column_mapping: Dict new_fingerprint: Optional = None ) → Dataset

Parameters

  • column_mapping (Dict[str, str]) — A mapping of columns to rename to their new names
  • new_fingerprint (str, optional) — The new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments.

Returns

Dataset

A copy of the dataset with renamed columns

Rename several columns in the dataset, and move the features associated to the original columns under the new column names.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds = ds.rename_columns({'text': 'text_new', 'label': 'label_new'})
Dataset({
    features: ['text_new', 'label_new'],
    num_rows: 1066
})

select_columns

< >

( column_names: Union new_fingerprint: Optional = None ) → Dataset

Parameters

  • column_names (Union[str, List[str]]) — Name of the column(s) to keep.
  • new_fingerprint (str, optional) — The new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments.

Returns

Dataset

A copy of the dataset object which only consists of selected columns.

Select one or several column(s) in the dataset and the features associated to them.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.select_columns(['text'])
Dataset({
    features: ['text'],
    num_rows: 1066
})

class_encode_column

< >

( column: str include_nulls: bool = False )

Parameters

  • column (str) — The name of the column to cast (list all the column names with column_names)
  • include_nulls (bool, defaults to False) — Whether to include null values in the class labels. If True, the null values will be encoded as the "None" class label.

    Added in 1.14.2

Casts the given column as ClassLabel and updates the table.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("boolq", split="validation")
>>> ds.features
{'answer': Value(dtype='bool', id=None),
 'passage': Value(dtype='string', id=None),
 'question': Value(dtype='string', id=None)}
>>> ds = ds.class_encode_column('answer')
>>> ds.features
{'answer': ClassLabel(num_classes=2, names=['False', 'True'], id=None),
 'passage': Value(dtype='string', id=None),
 'question': Value(dtype='string', id=None)}

__len__

< >

( )

Number of rows in the dataset.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.__len__
<bound method Dataset.__len__ of Dataset({
    features: ['text', 'label'],
    num_rows: 1066
})>

__iter__

< >

( )

Iterate through the examples.

If a formatting is set with Dataset.set_format() rows will be returned with the selected format.

iter

< >

( batch_size: int drop_last_batch: bool = False )

Parameters

  • batch_size (int) — size of each batch to yield.
  • drop_last_batch (bool, default False) — Whether a last batch smaller than the batch_size should be dropped

Iterate through the batches of size batch_size.

If a formatting is set with [~datasets.Dataset.set_format] rows will be returned with the selected format.

formatted_as

< >

( type: Optional = None columns: Optional = None output_all_columns: bool = False **format_kwargs )

Parameters

  • type (str, optional) — Output type selected in [None, 'numpy', 'torch', 'tensorflow', 'pandas', 'arrow', 'jax']. None means `getitem“ returns python objects (default).
  • columns (List[str], optional) — Columns to format in the output. None means __getitem__ returns all columns (default).
  • output_all_columns (bool, defaults to False) — Keep un-formatted columns as well in the output (as python objects).
  • **format_kwargs (additional keyword arguments) — Keywords arguments passed to the convert function like np.array, torch.tensor or tensorflow.ragged.constant.

To be used in a with statement. Set __getitem__ return format (type and columns).

set_format

< >

( type: Optional = None columns: Optional = None output_all_columns: bool = False **format_kwargs )

Parameters

  • type (str, optional) — Either output type selected in [None, 'numpy', 'torch', 'tensorflow', 'pandas', 'arrow', 'jax']. None means __getitem__ returns python objects (default).
  • columns (List[str], optional) — Columns to format in the output. None means __getitem__ returns all columns (default).
  • output_all_columns (bool, defaults to False) — Keep un-formatted columns as well in the output (as python objects).
  • **format_kwargs (additional keyword arguments) — Keywords arguments passed to the convert function like np.array, torch.tensor or tensorflow.ragged.constant.

Set __getitem__ return format (type and columns). The data formatting is applied on-the-fly. The format type (for example “numpy”) is used to format batches when using __getitem__. It’s also possible to use custom transforms for formatting using set_transform().

It is possible to call map() after calling set_format. Since map may add new columns, then the list of formatted columns

gets updated. In this case, if you apply map on a dataset to add a new column, then this column will be formatted as:

new formatted columns = (all columns - previously unformatted columns)

Example:

>>> from datasets import load_dataset
>>> from transformers import AutoTokenizer
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
>>> ds = ds.map(lambda x: tokenizer(x['text'], truncation=True, padding=True), batched=True)
>>> ds.set_format(type='numpy', columns=['text', 'label'])
>>> ds.format
{'type': 'numpy',
'format_kwargs': {},
'columns': ['text', 'label'],
'output_all_columns': False}

set_transform

< >

( transform: Optional columns: Optional = None output_all_columns: bool = False )

Parameters

  • transform (Callable, optional) — User-defined formatting transform, replaces the format defined by set_format(). A formatting function is a callable that takes a batch (as a dict) as input and returns a batch. This function is applied right before returning the objects in __getitem__.
  • columns (List[str], optional) — Columns to format in the output. If specified, then the input batch of the transform only contains those columns.
  • output_all_columns (bool, defaults to False) — Keep un-formatted columns as well in the output (as python objects). If set to True, then the other un-formatted columns are kept with the output of the transform.

Set __getitem__ return format using this transform. The transform is applied on-the-fly on batches when __getitem__ is called. As set_format(), this can be reset using reset_format().

Example:

>>> from datasets import load_dataset
>>> from transformers import AutoTokenizer
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')
>>> def encode(batch):
...     return tokenizer(batch['text'], padding=True, truncation=True, return_tensors='pt')
>>> ds.set_transform(encode)
>>> ds[0]
{'attention_mask': tensor([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
 1, 1]),
 'input_ids': tensor([  101, 29353,  2135, 15102,  1996,  9428, 20868,  2890,  8663,  6895,
         20470,  2571,  3663,  2090,  4603,  3017,  3008,  1998,  2037, 24211,
         5637,  1998, 11690,  2336,  1012,   102]),
 'token_type_ids': tensor([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0])}

reset_format

< >

( )

Reset __getitem__ return format to python objects and all columns.

Same as self.set_format()

Example:

>>> from datasets import load_dataset
>>> from transformers import AutoTokenizer
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
>>> ds = ds.map(lambda x: tokenizer(x['text'], truncation=True, padding=True), batched=True)
>>> ds.set_format(type='numpy', columns=['input_ids', 'token_type_ids', 'attention_mask', 'label'])
>>> ds.format
{'columns': ['input_ids', 'token_type_ids', 'attention_mask', 'label'],
 'format_kwargs': {},
 'output_all_columns': False,
 'type': 'numpy'}
>>> ds.reset_format()
>>> ds.format
{'columns': ['text', 'label', 'input_ids', 'token_type_ids', 'attention_mask'],
 'format_kwargs': {},
 'output_all_columns': False,
 'type': None}

with_format

< >

( type: Optional = None columns: Optional = None output_all_columns: bool = False **format_kwargs )

Parameters

  • type (str, optional) — Either output type selected in [None, 'numpy', 'torch', 'tensorflow', 'pandas', 'arrow', 'jax']. None means __getitem__ returns python objects (default).
  • columns (List[str], optional) — Columns to format in the output. None means __getitem__ returns all columns (default).
  • output_all_columns (bool, defaults to False) — Keep un-formatted columns as well in the output (as python objects).
  • **format_kwargs (additional keyword arguments) — Keywords arguments passed to the convert function like np.array, torch.tensor or tensorflow.ragged.constant.

Set __getitem__ return format (type and columns). The data formatting is applied on-the-fly. The format type (for example “numpy”) is used to format batches when using __getitem__.

It’s also possible to use custom transforms for formatting using with_transform().

Contrary to set_format(), with_format returns a new Dataset object.

Example:

>>> from datasets import load_dataset
>>> from transformers import AutoTokenizer
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
>>> ds = ds.map(lambda x: tokenizer(x['text'], truncation=True, padding=True), batched=True)
>>> ds.format
{'columns': ['text', 'label', 'input_ids', 'token_type_ids', 'attention_mask'],
 'format_kwargs': {},
 'output_all_columns': False,
 'type': None}
>>> ds = ds.with_format(type='tensorflow', columns=['input_ids', 'token_type_ids', 'attention_mask', 'label'])
>>> ds.format
{'columns': ['input_ids', 'token_type_ids', 'attention_mask', 'label'],
 'format_kwargs': {},
 'output_all_columns': False,
 'type': 'tensorflow'}

with_transform

< >

( transform: Optional columns: Optional = None output_all_columns: bool = False )

Parameters

  • transform (Callable, optional) — User-defined formatting transform, replaces the format defined by set_format(). A formatting function is a callable that takes a batch (as a dict) as input and returns a batch. This function is applied right before returning the objects in __getitem__.
  • columns (List[str], optional) — Columns to format in the output. If specified, then the input batch of the transform only contains those columns.
  • output_all_columns (bool, defaults to False) — Keep un-formatted columns as well in the output (as python objects). If set to True, then the other un-formatted columns are kept with the output of the transform.

Set __getitem__ return format using this transform. The transform is applied on-the-fly on batches when __getitem__ is called.

As set_format(), this can be reset using reset_format().

Contrary to set_transform(), with_transform returns a new Dataset object.

Example:

>>> from datasets import load_dataset
>>> from transformers import AutoTokenizer
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
>>> def encode(example):
...     return tokenizer(example["text"], padding=True, truncation=True, return_tensors='pt')
>>> ds = ds.with_transform(encode)
>>> ds[0]
{'attention_mask': tensor([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
 1, 1, 1, 1, 1]),
 'input_ids': tensor([  101, 18027, 16310, 16001,  1103,  9321,   178, 11604,  7235,  6617,
         1742,  2165,  2820,  1206,  6588, 22572, 12937,  1811,  2153,  1105,
         1147, 12890, 19587,  6463,  1105, 15026,  1482,   119,   102]),
 'token_type_ids': tensor([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
         0, 0, 0, 0, 0])}

__getitem__

< >

( key )

Can be used to index columns (by string names) or rows (by integer index or iterable of indices or bools).

cleanup_cache_files

< >

( ) → int

Returns

int

Number of removed files.

Clean up all cache files in the dataset cache directory, excepted the currently used cache file if there is one.

Be careful when running this command that no other process is currently using other cache files.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.cleanup_cache_files()
10

map

< >

( function: Optional = None with_indices: bool = False with_rank: bool = False input_columns: Union = None batched: bool = False batch_size: Optional = 1000 drop_last_batch: bool = False remove_columns: Union = None keep_in_memory: bool = False load_from_cache_file: Optional = None cache_file_name: Optional = None writer_batch_size: Optional = 1000 features: Optional = None disable_nullable: bool = False fn_kwargs: Optional = None num_proc: Optional = None suffix_template: str = '_{rank:05d}_of_{num_proc:05d}' new_fingerprint: Optional = None desc: Optional = None )

Parameters

  • function (Callable) — Function with one of the following signatures:

    • function(example: Dict[str, Any]) -> Dict[str, Any] if batched=False and with_indices=False and with_rank=False
    • function(example: Dict[str, Any], *extra_args) -> Dict[str, Any] if batched=False and with_indices=True and/or with_rank=True (one extra arg for each)
    • function(batch: Dict[str, List]) -> Dict[str, List] if batched=True and with_indices=False and with_rank=False
    • function(batch: Dict[str, List], *extra_args) -> Dict[str, List] if batched=True and with_indices=True and/or with_rank=True (one extra arg for each)

    For advanced usage, the function can also return a pyarrow.Table. Moreover if your function returns nothing (None), then map will run your function and return the dataset unchanged. If no function is provided, default to identity function: lambda x: x.

  • with_indices (bool, defaults to False) — Provide example indices to function. Note that in this case the signature of function should be def function(example, idx[, rank]): ....
  • with_rank (bool, defaults to False) — Provide process rank to function. Note that in this case the signature of function should be def function(example[, idx], rank): ....
  • input_columns (Optional[Union[str, List[str]]], defaults to None) — The columns to be passed into function as positional arguments. If None, a dict mapping to all formatted columns is passed as one argument.
  • batched (bool, defaults to False) — Provide batch of examples to function.
  • batch_size (int, optional, defaults to 1000) — Number of examples per batch provided to function if batched=True. If batch_size <= 0 or batch_size == None, provide the full dataset as a single batch to function.
  • drop_last_batch (bool, defaults to False) — Whether a last batch smaller than the batch_size should be dropped instead of being processed by the function.
  • remove_columns (Optional[Union[str, List[str]]], defaults to None) — Remove a selection of columns while doing the mapping. Columns will be removed before updating the examples with the output of function, i.e. if function is adding columns with names in remove_columns, these columns will be kept.
  • keep_in_memory (bool, defaults to False) — Keep the dataset in memory instead of writing it to a cache file.
  • load_from_cache_file (Optional[bool], defaults to True if caching is enabled) — If a cache file storing the current computation from function can be identified, use it instead of recomputing.
  • cache_file_name (str, optional, defaults to None) — Provide the name of a path for the cache file. It is used to store the results of the computation instead of the automatically generated cache file name.
  • writer_batch_size (int, defaults to 1000) — Number of rows per write operation for the cache file writer. This value is a good trade-off between memory usage during the processing, and processing speed. Higher value makes the processing do fewer lookups, lower value consume less temporary memory while running map.
  • features (Optional[datasets.Features], defaults to None) — Use a specific Features to store the cache file instead of the automatically generated one.
  • disable_nullable (bool, defaults to False) — Disallow null values in the table.
  • fn_kwargs (Dict, optional, defaults to None) — Keyword arguments to be passed to function.
  • num_proc (int, optional, defaults to None) — Max number of processes when generating cache. Already cached shards are loaded sequentially.
  • suffix_template (str) — If cache_file_name is specified, then this suffix will be added at the end of the base name of each. Defaults to "_{rank:05d}_of_{num_proc:05d}". For example, if cache_file_name is “processed.arrow”, then for rank=1 and num_proc=4, the resulting file would be "processed_00001_of_00004.arrow" for the default suffix.
  • new_fingerprint (str, optional, defaults to None) — The new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments.
  • desc (str, optional, defaults to None) — Meaningful description to be displayed alongside with the progress bar while mapping examples.

Apply a function to all the examples in the table (individually or in batches) and update the table. If your function returns a column that already exists, then it overwrites it.

You can specify whether the function should be batched or not with the batched parameter:

  • If batched is False, then the function takes 1 example in and should return 1 example. An example is a dictionary, e.g. {"text": "Hello there !"}.
  • If batched is True and batch_size is 1, then the function takes a batch of 1 example as input and can return a batch with 1 or more examples. A batch is a dictionary, e.g. a batch of 1 example is {"text": ["Hello there !"]}.
  • If batched is True and batch_size is n > 1, then the function takes a batch of n examples as input and can return a batch with n examples, or with an arbitrary number of examples. Note that the last batch may have less than n examples. A batch is a dictionary, e.g. a batch of n examples is {"text": ["Hello there !"] * n}.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> def add_prefix(example):
...     example["text"] = "Review: " + example["text"]
...     return example
>>> ds = ds.map(add_prefix)
>>> ds[0:3]["text"]
['Review: compassionately explores the seemingly irreconcilable situation between conservative christian parents and their estranged gay and lesbian children .',
 'Review: the soundtrack alone is worth the price of admission .',
 'Review: rodriguez does a splendid job of racial profiling hollywood style--casting excellent latin actors of all ages--a trend long overdue .']

# process a batch of examples
>>> ds = ds.map(lambda example: tokenizer(example["text"]), batched=True)
# set number of processors
>>> ds = ds.map(add_prefix, num_proc=4)

filter

< >

( function: Optional = None with_indices: bool = False with_rank: bool = False input_columns: Union = None batched: bool = False batch_size: Optional = 1000 keep_in_memory: bool = False load_from_cache_file: Optional = None cache_file_name: Optional = None writer_batch_size: Optional = 1000 fn_kwargs: Optional = None num_proc: Optional = None suffix_template: str = '_{rank:05d}_of_{num_proc:05d}' new_fingerprint: Optional = None desc: Optional = None )

Parameters

  • function (Callable) — Callable with one of the following signatures:

    • function(example: Dict[str, Any]) -> bool if batched=False and with_indices=False and with_rank=False
    • function(example: Dict[str, Any], *extra_args) -> bool if batched=False and with_indices=True and/or with_rank=True (one extra arg for each)
    • function(batch: Dict[str, List]) -> List[bool] if batched=True and with_indices=False and with_rank=False
    • function(batch: Dict[str, List], *extra_args) -> List[bool] if batched=True and with_indices=True and/or with_rank=True (one extra arg for each)

    If no function is provided, defaults to an always True function: lambda x: True.

  • with_indices (bool, defaults to False) — Provide example indices to function. Note that in this case the signature of function should be def function(example, idx[, rank]): ....
  • with_rank (bool, defaults to False) — Provide process rank to function. Note that in this case the signature of function should be def function(example[, idx], rank): ....
  • input_columns (str or List[str], optional) — The columns to be passed into function as positional arguments. If None, a dict mapping to all formatted columns is passed as one argument.
  • batched (bool, defaults to False) — Provide batch of examples to function.
  • batch_size (int, optional, defaults to 1000) — Number of examples per batch provided to function if batched = True. If batched = False, one example per batch is passed to function. If batch_size <= 0 or batch_size == None, provide the full dataset as a single batch to function.
  • keep_in_memory (bool, defaults to False) — Keep the dataset in memory instead of writing it to a cache file.
  • load_from_cache_file (Optional[bool], defaults to True if caching is enabled) — If a cache file storing the current computation from function can be identified, use it instead of recomputing.
  • cache_file_name (str, optional) — Provide the name of a path for the cache file. It is used to store the results of the computation instead of the automatically generated cache file name.
  • writer_batch_size (int, defaults to 1000) — Number of rows per write operation for the cache file writer. This value is a good trade-off between memory usage during the processing, and processing speed. Higher value makes the processing do fewer lookups, lower value consume less temporary memory while running map.
  • fn_kwargs (dict, optional) — Keyword arguments to be passed to function.
  • num_proc (int, optional) — Number of processes for multiprocessing. By default it doesn’t use multiprocessing.
  • suffix_template (str) — If cache_file_name is specified, then this suffix will be added at the end of the base name of each. For example, if cache_file_name is "processed.arrow", then for rank = 1 and num_proc = 4, the resulting file would be "processed_00001_of_00004.arrow" for the default suffix (default _{rank:05d}_of_{num_proc:05d}).
  • new_fingerprint (str, optional) — The new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments.
  • desc (str, optional, defaults to None) — Meaningful description to be displayed alongside with the progress bar while filtering examples.

Apply a filter function to all the elements in the table in batches and update the table so that the dataset only includes examples according to the filter function.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.filter(lambda x: x["label"] == 1)
Dataset({
    features: ['text', 'label'],
    num_rows: 533
})

select

< >

( indices: Iterable keep_in_memory: bool = False indices_cache_file_name: Optional = None writer_batch_size: Optional = 1000 new_fingerprint: Optional = None )

Parameters

  • indices (range, list, iterable, ndarray or Series) — Range, list or 1D-array of integer indices for indexing. If the indices correspond to a contiguous range, the Arrow table is simply sliced. However passing a list of indices that are not contiguous creates indices mapping, which is much less efficient, but still faster than recreating an Arrow table made of the requested rows.
  • keep_in_memory (bool, defaults to False) — Keep the indices mapping in memory instead of writing it to a cache file.
  • indices_cache_file_name (str, optional, defaults to None) — Provide the name of a path for the cache file. It is used to store the indices mapping instead of the automatically generated cache file name.
  • writer_batch_size (int, defaults to 1000) — Number of rows per write operation for the cache file writer. This value is a good trade-off between memory usage during the processing, and processing speed. Higher value makes the processing do fewer lookups, lower value consume less temporary memory while running map.
  • new_fingerprint (str, optional, defaults to None) — The new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments.

Create a new dataset with rows selected following the list/array of indices.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds.select(range(4))
Dataset({
    features: ['text', 'label'],
    num_rows: 4
})

sort

< >

( column_names: Union reverse: Union = False kind = 'deprecated' null_placement: str = 'at_end' keep_in_memory: bool = False load_from_cache_file: Optional = None indices_cache_file_name: Optional = None writer_batch_size: Optional = 1000 new_fingerprint: Optional = None )

Parameters

  • column_names (Union[str, Sequence[str]]) — Column name(s) to sort by.
  • reverse (Union[bool, Sequence[bool]], defaults to False) — If True, sort by descending order rather than ascending. If a single bool is provided, the value is applied to the sorting of all column names. Otherwise a list of bools with the same length and order as column_names must be provided.
  • kind (str, optional) — Pandas algorithm for sorting selected in {quicksort, mergesort, heapsort, stable}, The default is quicksort. Note that both stable and mergesort use timsort under the covers and, in general, the actual implementation will vary with data type. The mergesort option is retained for backwards compatibility.

    Deprecated in 2.8.0

    kind was deprecated in version 2.10.0 and will be removed in 3.0.0.
  • null_placement (str, defaults to at_end) — Put None values at the beginning if at_start or first or at the end if at_end or last

    Added in 1.14.2
  • keep_in_memory (bool, defaults to False) — Keep the sorted indices in memory instead of writing it to a cache file.
  • load_from_cache_file (Optional[bool], defaults to True if caching is enabled) — If a cache file storing the sorted indices can be identified, use it instead of recomputing.
  • indices_cache_file_name (str, optional, defaults to None) — Provide the name of a path for the cache file. It is used to store the sorted indices instead of the automatically generated cache file name.
  • writer_batch_size (int, defaults to 1000) — Number of rows per write operation for the cache file writer. Higher value gives smaller cache files, lower value consume less temporary memory.
  • new_fingerprint (str, optional, defaults to None) — The new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments

Create a new dataset sorted according to a single or multiple columns.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset('rotten_tomatoes', split='validation')
>>> ds['label'][:10]
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
>>> sorted_ds = ds.sort('label')
>>> sorted_ds['label'][:10]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
>>> another_sorted_ds = ds.sort(['label', 'text'], reverse=[True, False])
>>> another_sorted_ds['label'][:10]
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]

shuffle

< >

( seed: Optional = None generator: Optional = None keep_in_memory: bool = False load_from_cache_file: Optional = None indices_cache_file_name: Optional = None writer_batch_size: Optional = 1000 new_fingerprint: Optional = None )

Parameters

  • seed (int, optional) — A seed to initialize the default BitGenerator if generator=None. If None, then fresh, unpredictable entropy will be pulled from the OS. If an int or array_like[ints] is passed, then it will be passed to SeedSequence to derive the initial BitGenerator state.
  • generator (numpy.random.Generator, optional) — Numpy random Generator to use to compute the permutation of the dataset rows. If generator=None (default), uses np.random.default_rng (the default BitGenerator (PCG64) of NumPy).
  • keep_in_memory (bool, default False) — Keep the shuffled indices in memory instead of writing it to a cache file.
  • load_from_cache_file (Optional[bool], defaults to True if caching is enabled) — If a cache file storing the shuffled indices can be identified, use it instead of recomputing.
  • indices_cache_file_name (str, optional) — Provide the name of a path for the cache file. It is used to store the shuffled indices instead of the automatically generated cache file name.
  • writer_batch_size (int, defaults to 1000) — Number of rows per write operation for the cache file writer. This value is a good trade-off between memory usage during the processing, and processing speed. Higher value makes the processing do fewer lookups, lower value consume less temporary memory while running map.
  • new_fingerprint (str, optional, defaults to None) — The new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments.

Create a new Dataset where the rows are shuffled.

Currently shuffling uses numpy random generators. You can either supply a NumPy BitGenerator to use, or a seed to initiate NumPy’s default random generator (PCG64).

Shuffling takes the list of indices [0:len(my_dataset)] and shuffles it to create an indices mapping. However as soon as your Dataset has an indices mapping, the speed can become 10x slower. This is because there is an extra step to get the row index to read using the indices mapping, and most importantly, you aren’t reading contiguous chunks of data anymore. To restore the speed, you’d need to rewrite the entire dataset on your disk again using Dataset.flatten_indices(), which removes the indices mapping.

This may take a lot of time depending of the size of your dataset though:

my_dataset[0]  # fast
my_dataset = my_dataset.shuffle(seed=42)
my_dataset[0]  # up to 10x slower
my_dataset = my_dataset.flatten_indices()  # rewrite the shuffled dataset on disk as contiguous chunks of data
my_dataset[0]  # fast again

In this case, we recommend switching to an IterableDataset and leveraging its fast approximate shuffling method IterableDataset.shuffle().

It only shuffles the shards order and adds a shuffle buffer to your dataset, which keeps the speed of your dataset optimal:

my_iterable_dataset = my_dataset.to_iterable_dataset(num_shards=128)
for example in enumerate(my_iterable_dataset):  # fast
    pass

shuffled_iterable_dataset = my_iterable_dataset.shuffle(seed=42, buffer_size=100)

for example in enumerate(shuffled_iterable_dataset):  # as fast as before
    pass

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds['label'][:10]
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]

# set a seed
>>> shuffled_ds = ds.shuffle(seed=42)
>>> shuffled_ds['label'][:10]
[1, 0, 1, 1, 0, 0, 0, 0, 0, 0]

skip

< >

( n: int )

Parameters

  • n (int) — Number of elements to skip.

Create a new Dataset that skips the first n elements.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train")
>>> list(ds.take(3))
[{'label': 1,
 'text': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'},
 {'label': 1,
 'text': 'the gorgeously elaborate continuation of " the lord of the rings " trilogy is so huge that a column of words cannot adequately describe co-writer/director peter jackson's expanded vision of j . r . r . tolkien's middle-earth .'},
 {'label': 1, 'text': 'effective but too-tepid biopic'}]
>>> ds = ds.skip(1)
>>> list(ds.take(3))
[{'label': 1,
 'text': 'the gorgeously elaborate continuation of " the lord of the rings " trilogy is so huge that a column of words cannot adequately describe co-writer/director peter jackson's expanded vision of j . r . r . tolkien's middle-earth .'},
 {'label': 1, 'text': 'effective but too-tepid biopic'},
 {'label': 1,
 'text': 'if you sometimes like to go to the movies to have fun , wasabi is a good place to start .'}]

take

< >

( n: int )

Parameters

  • n (int) — Number of elements to take.

Create a new Dataset with only the first n elements.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train")
>>> small_ds = ds.take(2)
>>> list(small_ds)
[{'label': 1,
 'text': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'},
 {'label': 1,
 'text': 'the gorgeously elaborate continuation of " the lord of the rings " trilogy is so huge that a column of words cannot adequately describe co-writer/director peter jackson's expanded vision of j . r . r . tolkien's middle-earth .'}]

train_test_split

< >

( test_size: Union = None train_size: Union = None shuffle: bool = True stratify_by_column: Optional = None seed: Optional = None generator: Optional = None keep_in_memory: bool = False load_from_cache_file: Optional = None train_indices_cache_file_name: Optional = None test_indices_cache_file_name: Optional = None writer_batch_size: Optional = 1000 train_new_fingerprint: Optional = None test_new_fingerprint: Optional = None )

Parameters

  • test_size (numpy.random.Generator, optional) — Size of the test split If float, should be between 0.0 and 1.0 and represent the proportion of the dataset to include in the test split. If int, represents the absolute number of test samples. If None, the value is set to the complement of the train size. If train_size is also None, it will be set to 0.25.
  • train_size (numpy.random.Generator, optional) — Size of the train split If float, should be between 0.0 and 1.0 and represent the proportion of the dataset to include in the train split. If int, represents the absolute number of train samples. If None, the value is automatically set to the complement of the test size.
  • shuffle (bool, optional, defaults to True) — Whether or not to shuffle the data before splitting.
  • stratify_by_column (str, optional, defaults to None) — The column name of labels to be used to perform stratified split of data.
  • seed (int, optional) — A seed to initialize the default BitGenerator if generator=None. If None, then fresh, unpredictable entropy will be pulled from the OS. If an int or array_like[ints] is passed, then it will be passed to SeedSequence to derive the initial BitGenerator state.
  • generator (numpy.random.Generator, optional) — Numpy random Generator to use to compute the permutation of the dataset rows. If generator=None (default), uses np.random.default_rng (the default BitGenerator (PCG64) of NumPy).
  • keep_in_memory (bool, defaults to False) — Keep the splits indices in memory instead of writing it to a cache file.
  • load_from_cache_file (Optional[bool], defaults to True if caching is enabled) — If a cache file storing the splits indices can be identified, use it instead of recomputing.
  • train_cache_file_name (str, optional) — Provide the name of a path for the cache file. It is used to store the train split indices instead of the automatically generated cache file name.
  • test_cache_file_name (str, optional) — Provide the name of a path for the cache file. It is used to store the test split indices instead of the automatically generated cache file name.
  • writer_batch_size (int, defaults to 1000) — Number of rows per write operation for the cache file writer. This value is a good trade-off between memory usage during the processing, and processing speed. Higher value makes the processing do fewer lookups, lower value consume less temporary memory while running map.
  • train_new_fingerprint (str, optional, defaults to None) — The new fingerprint of the train set after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments
  • test_new_fingerprint (str, optional, defaults to None) — The new fingerprint of the test set after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments

Return a dictionary (datasets.DatasetDict) with two random train and test subsets (train and test Dataset splits). Splits are created from the dataset according to test_size, train_size and shuffle.

This method is similar to scikit-learn train_test_split.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds = ds.train_test_split(test_size=0.2, shuffle=True)
DatasetDict({
    train: Dataset({
        features: ['text', 'label'],
        num_rows: 852
    })
    test: Dataset({
        features: ['text', 'label'],
        num_rows: 214
    })
})

# set a seed
>>> ds = ds.train_test_split(test_size=0.2, seed=42)

# stratified split
>>> ds = load_dataset("imdb",split="train")
Dataset({
    features: ['text', 'label'],
    num_rows: 25000
})
>>> ds = ds.train_test_split(test_size=0.2, stratify_by_column="label")
DatasetDict({
    train: Dataset({
        features: ['text', 'label'],
        num_rows: 20000
    })
    test: Dataset({
        features: ['text', 'label'],
        num_rows: 5000
    })
})

shard

< >

( num_shards: int index: int contiguous: bool = False keep_in_memory: bool = False indices_cache_file_name: Optional = None writer_batch_size: Optional = 1000 )

Parameters

  • num_shards (int) — How many shards to split the dataset into.
  • index (int) — Which shard to select and return. contiguous — (bool, defaults to False): Whether to select contiguous blocks of indices for shards.
  • keep_in_memory (bool, defaults to False) — Keep the dataset in memory instead of writing it to a cache file.
  • indices_cache_file_name (str, optional) — Provide the name of a path for the cache file. It is used to store the indices of each shard instead of the automatically generated cache file name.
  • writer_batch_size (int, defaults to 1000) — Number of rows per write operation for the cache file writer. This value is a good trade-off between memory usage during the processing, and processing speed. Higher value makes the processing do fewer lookups, lower value consume less temporary memory while running map.

Return the index-nth shard from dataset split into num_shards pieces.

This shards deterministically. dset.shard(n, i) will contain all elements of dset whose index mod n = i.

dset.shard(n, i, contiguous=True) will instead split dset into contiguous chunks, so it can be easily concatenated back together after processing. If n % i == l, then the first l shards will have length (n // i) + 1, and the remaining shards will have length (n // i). datasets.concatenate([dset.shard(n, i, contiguous=True) for i in range(n)]) will return a dataset with the same order as the original.

Be sure to shard before using any randomizing operator (such as shuffle). It is best if the shard operator is used early in the dataset pipeline.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="validation")
>>> ds
Dataset({
    features: ['text', 'label'],
    num_rows: 1066
})
>>> ds.shard(num_shards=2, index=0)
Dataset({
    features: ['text', 'label'],
    num_rows: 533
})

to_tf_dataset

< >

( batch_size: Optional = None columns: Union = None shuffle: bool = False collate_fn: Optional = None drop_remainder: bool = False collate_fn_args: Optional = None label_cols: Union = None prefetch: bool = True num_workers: int = 0 num_test_batches: int = 20 )

Parameters

  • batch_size (int, optional) — Size of batches to load from the dataset. Defaults to None, which implies that the dataset won’t be batched, but the returned dataset can be batched later with tf_dataset.batch(batch_size).
  • columns (List[str] or str, optional) — Dataset column(s) to load in the tf.data.Dataset. Column names that are created by the collate_fn and that do not exist in the original dataset can be used.
  • shuffle(bool, defaults to False) — Shuffle the dataset order when loading. Recommended True for training, False for validation/evaluation.
  • drop_remainder(bool, defaults to False) — Drop the last incomplete batch when loading. Ensures that all batches yielded by the dataset will have the same length on the batch dimension.
  • collate_fn(Callable, optional) — A function or callable object (such as a DataCollator) that will collate lists of samples into a batch.
  • collate_fn_args (Dict, optional) — An optional dict of keyword arguments to be passed to the collate_fn.
  • label_cols (List[str] or str, defaults to None) — Dataset column(s) to load as labels. Note that many models compute loss internally rather than letting Keras do it, in which case passing the labels here is optional, as long as they’re in the input columns.
  • prefetch (bool, defaults to True) — Whether to run the dataloader in a separate thread and maintain a small buffer of batches for training. Improves performance by allowing data to be loaded in the background while the model is training.
  • num_workers (int, defaults to 0) — Number of workers to use for loading the dataset. Only supported on Python versions >= 3.8.
  • num_test_batches (int, defaults to 20) — Number of batches to use to infer the output signature of the dataset. The higher this number, the more accurate the signature will be, but the longer it will take to create the dataset.

Create a tf.data.Dataset from the underlying Dataset. This tf.data.Dataset will load and collate batches from the Dataset, and is suitable for passing to methods like model.fit() or model.predict(). The dataset will yield dicts for both inputs and labels unless the dict would contain only a single key, in which case a raw tf.Tensor is yielded instead.

Example:

>>> ds_train = ds["train"].to_tf_dataset(
...    columns=['input_ids', 'token_type_ids', 'attention_mask', 'label'],
...    shuffle=True,
...    batch_size=16,
...    collate_fn=data_collator,
... )

push_to_hub

< >

( repo_id: str config_name: str = 'default' set_default: Optional = None split: Optional = None data_dir: Optional = None commit_message: Optional = None commit_description: Optional = None private: Optional = False token: Optional = None revision: Optional = None branch = 'deprecated' create_pr: Optional = False max_shard_size: Union = None num_shards: Optional = None embed_external_files: bool = True )

Parameters

  • repo_id (str) — The ID of the repository to push to in the following format: <user>/<dataset_name> or <org>/<dataset_name>. Also accepts <dataset_name>, which will default to the namespace of the logged-in user.
  • config_name (str, defaults to “default”) — The configuration name (or subset) of a dataset. Defaults to “default”.
  • set_default (bool, optional) — Whether to set this configuration as the default one. Otherwise, the default configuration is the one named “default”.
  • split (str, optional) — The name of the split that will be given to that dataset. Defaults to self.split.
  • data_dir (str, optional) — Directory name that will contain the uploaded data files. Defaults to the config_name if different from “default”, else “data”.

    Added in 2.17.0
  • commit_message (str, optional) — Message to commit while pushing. Will default to "Upload dataset".
  • commit_description (str, optional) — Description of the commit that will be created. Additionally, description of the PR if a PR is created (create_pr is True).

    Added in 2.16.0
  • private (bool, optional, defaults to False) — Whether the dataset repository should be set to private or not. Only affects repository creation: a repository that already exists will not be affected by that parameter.
  • token (str, optional) — An optional authentication token for the Hugging Face Hub. If no token is passed, will default to the token saved locally when logging in with huggingface-cli login. Will raise an error if no token is passed and the user is not logged-in.
  • revision (str, optional) — Branch to push the uploaded files to. Defaults to the "main" branch.

    Added in 2.15.0
  • branch (str, optional) — The git branch on which to push the dataset. This defaults to the default branch as specified in your repository, which defaults to "main".

    Deprecated in 2.15.0

    branch was deprecated in favor of revision in version 2.15.0 and will be removed in 3.0.0.
  • create_pr (bool, optional, defaults to False) — Whether to create a PR with the uploaded files or directly commit.

    Added in 2.15.0
  • max_shard_size (int or str, optional, defaults to "500MB") — The maximum size of the dataset shards to be uploaded to the hub. If expressed as a string, needs to be digits followed by a unit (like "5MB").
  • num_shards (int, optional) — Number of shards to write. By default, the number of shards depends on max_shard_size.

    Added in 2.8.0
  • embed_external_files (bool, defaults to True) — Whether to embed file bytes in the shards. In particular, this will do the following before the push for the fields of type:

    • Audio and Image: remove local path information and embed file content in the Parquet files.

Pushes the dataset to the hub as a Parquet dataset. The dataset is pushed using HTTP requests and does not need to have neither git or git-lfs installed.

The resulting Parquet files are self-contained by default. If your dataset contains Image or Audio data, the Parquet files will store the bytes of your images or audio files. You can disable this by setting embed_external_files to False.

Example:

>>> dataset.push_to_hub("<organization>/<dataset_id>")
>>> dataset_dict.push_to_hub("<organization>/<dataset_id>", private=True)
>>> dataset.push_to_hub("<organization>/<dataset_id>", max_shard_size="1GB")
>>> dataset.push_to_hub("<organization>/<dataset_id>", num_shards=1024)

If your dataset has multiple splits (e.g. train/validation/test):

>>> train_dataset.push_to_hub("<organization>/<dataset_id>", split="train")
>>> val_dataset.push_to_hub("<organization>/<dataset_id>", split="validation")
>>> # later
>>> dataset = load_dataset("<organization>/<dataset_id>")
>>> train_dataset = dataset["train"]
>>> val_dataset = dataset["validation"]

If you want to add a new configuration (or subset) to a dataset (e.g. if the dataset has multiple tasks/versions/languages):

>>> english_dataset.push_to_hub("<organization>/<dataset_id>", "en")
>>> french_dataset.push_to_hub("<organization>/<dataset_id>", "fr")
>>> # later
>>> english_dataset = load_dataset("<organization>/<dataset_id>", "en")
>>> french_dataset = load_dataset("<organization>/<dataset_id>", "fr")

save_to_disk

< >

( dataset_path: Union fs = 'deprecated' max_shard_size: Union = None num_shards: Optional = None num_proc: Optional = None storage_options: Optional = None )

Parameters

  • dataset_path (path-like) — Path (e.g. dataset/train) or remote URI (e.g. s3://my-bucket/dataset/train) of the dataset directory where the dataset will be saved to.
  • fs (fsspec.spec.AbstractFileSystem, optional) — Instance of the remote filesystem where the dataset will be saved to.

    Deprecated in 2.8.0

    fs was deprecated in version 2.8.0 and will be removed in 3.0.0. Please use storage_options instead, e.g. storage_options=fs.storage_options
  • max_shard_size (int or str, optional, defaults to "500MB") — The maximum size of the dataset shards to be uploaded to the hub. If expressed as a string, needs to be digits followed by a unit (like "50MB").
  • num_shards (int, optional) — Number of shards to write. By default the number of shards depends on max_shard_size and num_proc.

    Added in 2.8.0
  • num_proc (int, optional) — Number of processes when downloading and generating the dataset locally. Multiprocessing is disabled by default.

    Added in 2.8.0
  • storage_options (dict, optional) — Key/value pairs to be passed on to the file-system backend, if any.

    Added in 2.8.0

Saves a dataset to a dataset directory, or in a filesystem using any implementation of fsspec.spec.AbstractFileSystem.

For Image and Audio data:

All the Image() and Audio() data are stored in the arrow files. If you want to store paths or urls, please use the Value(“string”) type.

Example:

>>> ds.save_to_disk("path/to/dataset/directory")
>>> ds.save_to_disk("path/to/dataset/directory", max_shard_size="1GB")
>>> ds.save_to_disk("path/to/dataset/directory", num_shards=1024)

load_from_disk

< >

( dataset_path: Union fs = 'deprecated' keep_in_memory: Optional = None storage_options: Optional = None ) → Dataset or DatasetDict

Parameters

  • dataset_path (path-like) — Path (e.g. "dataset/train") or remote URI (e.g. "s3//my-bucket/dataset/train") of the dataset directory where the dataset will be loaded from.
  • fs (fsspec.spec.AbstractFileSystem, optional) — Instance of the remote filesystem where the dataset will be saved to.

    Deprecated in 2.8.0

    fs was deprecated in version 2.8.0 and will be removed in 3.0.0. Please use storage_options instead, e.g. storage_options=fs.storage_options
  • keep_in_memory (bool, defaults to None) — Whether to copy the dataset in-memory. If None, the dataset will not be copied in-memory unless explicitly enabled by setting datasets.config.IN_MEMORY_MAX_SIZE to nonzero. See more details in the improve performance section.
  • storage_options (dict, optional) — Key/value pairs to be passed on to the file-system backend, if any.

    Added in 2.8.0

Returns

Dataset or DatasetDict

  • If dataset_path is a path of a dataset directory, the dataset requested.
  • If dataset_path is a path of a dataset dict directory, a datasets.DatasetDict with each split.

Loads a dataset that was previously saved using save_to_disk from a dataset directory, or from a filesystem using any implementation of fsspec.spec.AbstractFileSystem.

Example:

>>> ds = load_from_disk("path/to/dataset/directory")

flatten_indices

< >

( keep_in_memory: bool = False cache_file_name: Optional = None writer_batch_size: Optional = 1000 features: Optional = None disable_nullable: bool = False num_proc: Optional = None new_fingerprint: Optional = None )

Parameters

  • keep_in_memory (bool, defaults to False) — Keep the dataset in memory instead of writing it to a cache file.
  • cache_file_name (str, optional, default None) — Provide the name of a path for the cache file. It is used to store the results of the computation instead of the automatically generated cache file name.
  • writer_batch_size (int, defaults to 1000) — Number of rows per write operation for the cache file writer. This value is a good trade-off between memory usage during the processing, and processing speed. Higher value makes the processing do fewer lookups, lower value consume less temporary memory while running map.
  • features (Optional[datasets.Features], defaults to None) — Use a specific Features to store the cache file instead of the automatically generated one.
  • disable_nullable (bool, defaults to False) — Allow null values in the table.
  • num_proc (int, optional, default None) — Max number of processes when generating cache. Already cached shards are loaded sequentially
  • new_fingerprint (str, optional, defaults to None) — The new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments

Create and cache a new Dataset by flattening the indices mapping.

to_csv

< >

( path_or_buf: Union batch_size: Optional = None num_proc: Optional = None storage_options: Optional = None **to_csv_kwargs ) → int

Parameters

  • path_or_buf (PathLike or FileOrBuffer) — Either a path to a file (e.g. file.csv), a remote URI (e.g. hf://datasets/username/my_dataset_name/data.csv), or a BinaryIO, where the dataset will be saved to in the specified format.
  • batch_size (int, optional) — Size of the batch to load in memory and write at once. Defaults to datasets.config.DEFAULT_MAX_BATCH_SIZE.
  • num_proc (int, optional) — Number of processes for multiprocessing. By default it doesn’t use multiprocessing. batch_size in this case defaults to datasets.config.DEFAULT_MAX_BATCH_SIZE but feel free to make it 5x or 10x of the default value if you have sufficient compute power.
  • storage_options (dict, optional) — Key/value pairs to be passed on to the file-system backend, if any.

    Added in 2.19.0
  • **to_csv_kwargs (additional keyword arguments) — Parameters to pass to pandas’s pandas.DataFrame.to_csv.

    Changed in 2.10.0

    Now, index defaults to False if not specified.

    If you would like to write the index, pass index=True and also set a name for the index column by passing index_label.

Returns

int

The number of characters or bytes written.

Exports the dataset to csv

Example:

>>> ds.to_csv("path/to/dataset/directory")

to_pandas

< >

( batch_size: Optional = None batched: bool = False )

Parameters

  • batched (bool) — Set to True to return a generator that yields the dataset as batches of batch_size rows. Defaults to False (returns the whole datasets once).
  • batch_size (int, optional) — The size (number of rows) of the batches if batched is True. Defaults to datasets.config.DEFAULT_MAX_BATCH_SIZE.

Returns the dataset as a pandas.DataFrame. Can also return a generator for large datasets.

Example:

>>> ds.to_pandas()

to_dict

< >

( batch_size: Optional = None batched = 'deprecated' )

Parameters

  • batched (bool) — Set to True to return a generator that yields the dataset as batches of batch_size rows. Defaults to False (returns the whole datasets once).

    Deprecated in 2.11.0

    Use .iter(batch_size=batch_size) followed by .to_dict() on the individual batches instead.
  • batch_size (int, optional) — The size (number of rows) of the batches if batched is True. Defaults to datasets.config.DEFAULT_MAX_BATCH_SIZE.

Returns the dataset as a Python dict. Can also return a generator for large datasets.

Example:

>>> ds.to_dict()

to_json

< >

( path_or_buf: Union batch_size: Optional = None num_proc: Optional = None storage_options: Optional = None **to_json_kwargs ) → int

Parameters

  • path_or_buf (PathLike or FileOrBuffer) — Either a path to a file (e.g. file.json), a remote URI (e.g. hf://datasets/username/my_dataset_name/data.json), or a BinaryIO, where the dataset will be saved to in the specified format.
  • batch_size (int, optional) — Size of the batch to load in memory and write at once. Defaults to datasets.config.DEFAULT_MAX_BATCH_SIZE.
  • num_proc (int, optional) — Number of processes for multiprocessing. By default, it doesn’t use multiprocessing. batch_size in this case defaults to datasets.config.DEFAULT_MAX_BATCH_SIZE but feel free to make it 5x or 10x of the default value if you have sufficient compute power.
  • storage_options (dict, optional) — Key/value pairs to be passed on to the file-system backend, if any.

    Added in 2.19.0
  • **to_json_kwargs (additional keyword arguments) — Parameters to pass to pandas’s pandas.DataFrame.to_json. Default arguments are lines=True and `orient=“records”.

    Changed in 2.11.0

    The parameter index defaults to False if orient is "split" or "table".

    If you would like to write the index, pass index=True.

Returns

int

The number of characters or bytes written.

Export the dataset to JSON Lines or JSON.

The default output format is JSON Lines. To export to JSON, pass lines=False argument and the desired orient.

Example:

>>> ds.to_json("path/to/dataset/directory/filename.jsonl")

to_parquet

< >

( path_or_buf: Union batch_size: Optional = None storage_options: Optional = None **parquet_writer_kwargs ) → int

Parameters

  • path_or_buf (PathLike or FileOrBuffer) — Either a path to a file (e.g. file.parquet), a remote URI (e.g. hf://datasets/username/my_dataset_name/data.parquet), or a BinaryIO, where the dataset will be saved to in the specified format.
  • batch_size (int, optional) — Size of the batch to load in memory and write at once. Defaults to datasets.config.DEFAULT_MAX_BATCH_SIZE.
  • storage_options (dict, optional) — Key/value pairs to be passed on to the file-system backend, if any.

    Added in 2.19.0
  • **parquet_writer_kwargs (additional keyword arguments) — Parameters to pass to PyArrow’s pyarrow.parquet.ParquetWriter.

Returns

int

The number of characters or bytes written.

Exports the dataset to parquet

Example:

>>> ds.to_parquet("path/to/dataset/directory")

to_sql

< >

( name: str con: Union batch_size: Optional = None **sql_writer_kwargs ) → int

Parameters

  • name (str) — Name of SQL table.
  • con (str or sqlite3.Connection or sqlalchemy.engine.Connection or sqlalchemy.engine.Connection) — A URI string or a SQLite3/SQLAlchemy connection object used to write to a database.
  • batch_size (int, optional) — Size of the batch to load in memory and write at once. Defaults to datasets.config.DEFAULT_MAX_BATCH_SIZE.
  • **sql_writer_kwargs (additional keyword arguments) — Parameters to pass to pandas’s pandas.DataFrame.to_sql.

    Changed in 2.11.0

    Now, index defaults to False if not specified.

    If you would like to write the index, pass index=True and also set a name for the index column by passing index_label.

Returns

int

The number of records written.

Exports the dataset to a SQL database.

Example:

>>> # con provided as a connection URI string
>>> ds.to_sql("data", "sqlite:///my_own_db.sql")
>>> # con provided as a sqlite3 connection object
>>> import sqlite3
>>> con = sqlite3.connect("my_own_db.sql")
>>> with con:
...     ds.to_sql("data", con)

to_iterable_dataset

< >

( num_shards: Optional = 1 )

Parameters

  • num_shards (int, default to 1) — Number of shards to define when instantiating the iterable dataset. This is especially useful for big datasets to be able to shuffle properly, and also to enable fast parallel loading using a PyTorch DataLoader or in distributed setups for example. Shards are defined using datasets.Dataset.shard(): it simply slices the data without writing anything on disk.

Get an datasets.IterableDataset from a map-style datasets.Dataset. This is equivalent to loading a dataset in streaming mode with datasets.load_dataset(), but much faster since the data is streamed from local files.

Contrary to map-style datasets, iterable datasets are lazy and can only be iterated over (e.g. using a for loop). Since they are read sequentially in training loops, iterable datasets are much faster than map-style datasets. All the transformations applied to iterable datasets like filtering or processing are done on-the-fly when you start iterating over the dataset.

Still, it is possible to shuffle an iterable dataset using datasets.IterableDataset.shuffle(). This is a fast approximate shuffling that works best if you have multiple shards and if you specify a buffer size that is big enough.

To get the best speed performance, make sure your dataset doesn’t have an indices mapping. If this is the case, the data are not read contiguously, which can be slow sometimes. You can use ds = ds.flatten_indices() to write your dataset in contiguous chunks of data and have optimal speed before switching to an iterable dataset.

Example:

Basic usage:

>>> ids = ds.to_iterable_dataset()
>>> for example in ids:
...     pass

With lazy filtering and processing:

>>> ids = ds.to_iterable_dataset()
>>> ids = ids.filter(filter_fn).map(process_fn)  # will filter and process on-the-fly when you start iterating over the iterable dataset
>>> for example in ids:
...     pass

With sharding to enable efficient shuffling:

>>> ids = ds.to_iterable_dataset(num_shards=64)  # the dataset is split into 64 shards to be iterated over
>>> ids = ids.shuffle(buffer_size=10_000)  # will shuffle the shards order and use a shuffle buffer for fast approximate shuffling when you start iterating
>>> for example in ids:
...     pass

With a PyTorch DataLoader:

>>> import torch
>>> ids = ds.to_iterable_dataset(num_shards=64)
>>> ids = ids.filter(filter_fn).map(process_fn)
>>> dataloader = torch.utils.data.DataLoader(ids, num_workers=4)  # will assign 64 / 4 = 16 shards to each worker to load, filter and process when you start iterating
>>> for example in ids:
...     pass

With a PyTorch DataLoader and shuffling:

>>> import torch
>>> ids = ds.to_iterable_dataset(num_shards=64)
>>> ids = ids.shuffle(buffer_size=10_000)  # will shuffle the shards order and use a shuffle buffer when you start iterating
>>> dataloader = torch.utils.data.DataLoader(ids, num_workers=4)  # will assign 64 / 4 = 16 shards from the shuffled list of shards to each worker when you start iterating
>>> for example in ids:
...     pass

In a distributed setup like PyTorch DDP with a PyTorch DataLoader and shuffling

>>> from datasets.distributed import split_dataset_by_node
>>> ids = ds.to_iterable_dataset(num_shards=512)
>>> ids = ids.shuffle(buffer_size=10_000)  # will shuffle the shards order and use a shuffle buffer when you start iterating
>>> ids = split_dataset_by_node(ds, world_size=8, rank=0)  # will keep only 512 / 8 = 64 shards from the shuffled lists of shards when you start iterating
>>> dataloader = torch.utils.data.DataLoader(ids, num_workers=4)  # will assign 64 / 4 = 16 shards from this node's list of shards to each worker when you start iterating
>>> for example in ids:
...     pass

With shuffling and multiple epochs:

>>> ids = ds.to_iterable_dataset(num_shards=64)
>>> ids = ids.shuffle(buffer_size=10_000, seed=42)  # will shuffle the shards order and use a shuffle buffer when you start iterating
>>> for epoch in range(n_epochs):
...     ids.set_epoch(epoch)  # will use effective_seed = seed + epoch to shuffle the shards and for the shuffle buffer when you start iterating
...     for example in ids:
...         pass
Feel free to also use `IterableDataset.set_epoch()` when using a PyTorch DataLoader or in distributed setups.

add_faiss_index

< >

( column: str index_name: Optional = None device: Optional = None string_factory: Optional = None metric_type: Optional = None custom_index: Optional = None batch_size: int = 1000 train_size: Optional = None faiss_verbose: bool = False dtype = <class 'numpy.float32'> )

Parameters

  • column (str) — The column of the vectors to add to the index.
  • index_name (str, optional) — The index_name/identifier of the index. This is the index_name that is used to call get_nearest_examples() or search(). By default it corresponds to column.
  • device (Union[int, List[int]], optional) — If positive integer, this is the index of the GPU to use. If negative integer, use all GPUs. If a list of positive integers is passed in, run only on those GPUs. By default it uses the CPU.
  • string_factory (str, optional) — This is passed to the index factory of Faiss to create the index. Default index class is IndexFlat.
  • metric_type (int, optional) — Type of metric. Ex: faiss.METRIC_INNER_PRODUCT or faiss.METRIC_L2.
  • custom_index (faiss.Index, optional) — Custom Faiss index that you already have instantiated and configured for your needs.
  • batch_size (int) — Size of the batch to use while adding vectors to the FaissIndex. Default value is 1000.

    Added in 2.4.0
  • train_size (int, optional) — If the index needs a training step, specifies how many vectors will be used to train the index.
  • faiss_verbose (bool, defaults to False) — Enable the verbosity of the Faiss index.
  • dtype (data-type) — The dtype of the numpy arrays that are indexed. Default is np.float32.

Add a dense index using Faiss for fast retrieval. By default the index is done over the vectors of the specified column. You can specify device if you want to run it on GPU (device must be the GPU index). You can find more information about Faiss here:

Example:

>>> ds = datasets.load_dataset('crime_and_punish', split='train')
>>> ds_with_embeddings = ds.map(lambda example: {'embeddings': embed(example['line']}))
>>> ds_with_embeddings.add_faiss_index(column='embeddings')
>>> # query
>>> scores, retrieved_examples = ds_with_embeddings.get_nearest_examples('embeddings', embed('my new query'), k=10)
>>> # save index
>>> ds_with_embeddings.save_faiss_index('embeddings', 'my_index.faiss')

>>> ds = datasets.load_dataset('crime_and_punish', split='train')
>>> # load index
>>> ds.load_faiss_index('embeddings', 'my_index.faiss')
>>> # query
>>> scores, retrieved_examples = ds.get_nearest_examples('embeddings', embed('my new query'), k=10)

add_faiss_index_from_external_arrays

< >

( external_arrays: array index_name: str device: Optional = None string_factory: Optional = None metric_type: Optional = None custom_index: Optional = None batch_size: int = 1000 train_size: Optional = None faiss_verbose: bool = False dtype = <class 'numpy.float32'> )

Parameters

  • external_arrays (np.array) — If you want to use arrays from outside the lib for the index, you can set external_arrays. It will use external_arrays to create the Faiss index instead of the arrays in the given column.
  • index_name (str) — The index_name/identifier of the index. This is the index_name that is used to call get_nearest_examples() or search().
  • device (Optional Union[int, List[int]], optional) — If positive integer, this is the index of the GPU to use. If negative integer, use all GPUs. If a list of positive integers is passed in, run only on those GPUs. By default it uses the CPU.
  • string_factory (str, optional) — This is passed to the index factory of Faiss to create the index. Default index class is IndexFlat.
  • metric_type (int, optional) — Type of metric. Ex: faiss.faiss.METRIC_INNER_PRODUCT or faiss.METRIC_L2.
  • custom_index (faiss.Index, optional) — Custom Faiss index that you already have instantiated and configured for your needs.
  • batch_size (int, optional) — Size of the batch to use while adding vectors to the FaissIndex. Default value is 1000.

    Added in 2.4.0
  • train_size (int, optional) — If the index needs a training step, specifies how many vectors will be used to train the index.
  • faiss_verbose (bool, defaults to False) — Enable the verbosity of the Faiss index.
  • dtype (numpy.dtype) — The dtype of the numpy arrays that are indexed. Default is np.float32.

Add a dense index using Faiss for fast retrieval. The index is created using the vectors of external_arrays. You can specify device if you want to run it on GPU (device must be the GPU index). You can find more information about Faiss here:

save_faiss_index

< >

( index_name: str file: Union storage_options: Optional = None )

Parameters

  • index_name (str) — The index_name/identifier of the index. This is the index_name that is used to call .get_nearest or .search.
  • file (str) — The path to the serialized faiss index on disk or remote URI (e.g. "s3://my-bucket/index.faiss").
  • storage_options (dict, optional) — Key/value pairs to be passed on to the file-system backend, if any.

    Added in 2.11.0

Save a FaissIndex on disk.

load_faiss_index

< >

( index_name: str file: Union device: Union = None storage_options: Optional = None )

Parameters

  • index_name (str) — The index_name/identifier of the index. This is the index_name that is used to call .get_nearest or .search.
  • file (str) — The path to the serialized faiss index on disk or remote URI (e.g. "s3://my-bucket/index.faiss").
  • device (Optional Union[int, List[int]]) — If positive integer, this is the index of the GPU to use. If negative integer, use all GPUs. If a list of positive integers is passed in, run only on those GPUs. By default it uses the CPU.
  • storage_options (dict, optional) — Key/value pairs to be passed on to the file-system backend, if any.

    Added in 2.11.0

Load a FaissIndex from disk.

If you want to do additional configurations, you can have access to the faiss index object by doing .get_index(index_name).faiss_index to make it fit your needs.

add_elasticsearch_index

< >

( column: str index_name: Optional = None host: Optional = None port: Optional = None es_client: Optional = None es_index_name: Optional = None es_index_config: Optional = None )

Parameters

  • column (str) — The column of the documents to add to the index.
  • index_name (str, optional) — The index_name/identifier of the index. This is the index name that is used to call get_nearest_examples() or search(). By default it corresponds to column.
  • host (str, optional, defaults to localhost) — Host of where ElasticSearch is running.
  • port (str, optional, defaults to 9200) — Port of where ElasticSearch is running.
  • es_client (elasticsearch.Elasticsearch, optional) — The elasticsearch client used to create the index if host and port are None.
  • es_index_name (str, optional) — The elasticsearch index name used to create the index.
  • es_index_config (dict, optional) — The configuration of the elasticsearch index. Default config is:

Add a text index using ElasticSearch for fast retrieval. This is done in-place.

Example:

>>> es_client = elasticsearch.Elasticsearch()
>>> ds = datasets.load_dataset('crime_and_punish', split='train')
>>> ds.add_elasticsearch_index(column='line', es_client=es_client, es_index_name="my_es_index")
>>> scores, retrieved_examples = ds.get_nearest_examples('line', 'my new query', k=10)

load_elasticsearch_index

< >

( index_name: str es_index_name: str host: Optional = None port: Optional = None es_client: Optional = None es_index_config: Optional = None )

Parameters

  • index_name (str) — The index_name/identifier of the index. This is the index name that is used to call get_nearest or search.
  • es_index_name (str) — The name of elasticsearch index to load.
  • host (str, optional, defaults to localhost) — Host of where ElasticSearch is running.
  • port (str, optional, defaults to 9200) — Port of where ElasticSearch is running.
  • es_client (elasticsearch.Elasticsearch, optional) — The elasticsearch client used to create the index if host and port are None.
  • es_index_config (dict, optional) — The configuration of the elasticsearch index. Default config is:

Load an existing text index using ElasticSearch for fast retrieval.

list_indexes

< >

( )

List the colindex_nameumns/identifiers of all the attached indexes.

get_index

< >

( index_name: str )

Parameters

  • index_name (str) — Index name.

List the index_name/identifiers of all the attached indexes.

drop_index

< >

( index_name: str )

Parameters

  • index_name (str) — The index_name/identifier of the index.

Drop the index with the specified column.

search

< >

( index_name: str query: Union k: int = 10 **kwargs ) → (scores, indices)

Parameters

  • index_name (str) — The name/identifier of the index.
  • query (Union[str, np.ndarray]) — The query as a string if index_name is a text index or as a numpy array if index_name is a vector index.
  • k (int) — The number of examples to retrieve.

Returns

(scores, indices)

A tuple of (scores, indices) where:

  • scores (List[List[float]): the retrieval scores from either FAISS (IndexFlatL2 by default) or ElasticSearch of the retrieved examples
  • indices (List[List[int]]): the indices of the retrieved examples

Find the nearest examples indices in the dataset to the query.

search_batch

< >

( index_name: str queries: Union k: int = 10 **kwargs ) → (total_scores, total_indices)

Parameters

  • index_name (str) — The index_name/identifier of the index.
  • queries (Union[List[str], np.ndarray]) — The queries as a list of strings if index_name is a text index or as a numpy array if index_name is a vector index.
  • k (int) — The number of examples to retrieve per query.

Returns

(total_scores, total_indices)

A tuple of (total_scores, total_indices) where:

  • total_scores (List[List[float]): the retrieval scores from either FAISS (IndexFlatL2 by default) or ElasticSearch of the retrieved examples per query
  • total_indices (List[List[int]]): the indices of the retrieved examples per query

Find the nearest examples indices in the dataset to the query.

get_nearest_examples

< >

( index_name: str query: Union k: int = 10 **kwargs ) → (scores, examples)

Parameters

  • index_name (str) — The index_name/identifier of the index.
  • query (Union[str, np.ndarray]) — The query as a string if index_name is a text index or as a numpy array if index_name is a vector index.
  • k (int) — The number of examples to retrieve.

Returns

(scores, examples)

A tuple of (scores, examples) where:

  • scores (List[float]): the retrieval scores from either FAISS (IndexFlatL2 by default) or ElasticSearch of the retrieved examples
  • examples (dict): the retrieved examples

Find the nearest examples in the dataset to the query.

get_nearest_examples_batch

< >

( index_name: str queries: Union k: int = 10 **kwargs ) → (total_scores, total_examples)

Parameters

  • index_name (str) — The index_name/identifier of the index.
  • queries (Union[List[str], np.ndarray]) — The queries as a list of strings if index_name is a text index or as a numpy array if index_name is a vector index.
  • k (int) — The number of examples to retrieve per query.

Returns

(total_scores, total_examples)

A tuple of (total_scores, total_examples) where:

  • total_scores (List[List[float]): the retrieval scores from either FAISS (IndexFlatL2 by default) or ElasticSearch of the retrieved examples per query
  • total_examples (List[dict]): the retrieved examples per query

Find the nearest examples in the dataset to the query.

info

< >

( )

DatasetInfo object containing all the metadata in the dataset.

split

< >

( )

NamedSplit object corresponding to a named dataset split.

builder_name

< >

( )

citation

< >

( )

config_name

< >

( )

dataset_size

< >

( )

description

< >

( )

download_checksums

< >

( )

download_size

< >

( )

features

< >

( )

homepage

< >

( )

license

< >

( )

size_in_bytes

< >

( )

supervised_keys

< >

( )

version

< >

( )

from_csv

< >

( path_or_paths: Union split: Optional = None features: Optional = None cache_dir: str = None keep_in_memory: bool = False num_proc: Optional = None **kwargs )

Parameters

  • path_or_paths (path-like or list of path-like) — Path(s) of the CSV file(s).
  • split (NamedSplit, optional) — Split name to be assigned to the dataset.
  • features (Features, optional) — Dataset features.
  • cache_dir (str, optional, defaults to "~/.cache/huggingface/datasets") — Directory to cache data.
  • keep_in_memory (bool, defaults to False) — Whether to copy the data in-memory.
  • num_proc (int, optional, defaults to None) — Number of processes when downloading and generating the dataset locally. This is helpful if the dataset is made of multiple files. Multiprocessing is disabled by default.

    Added in 2.8.0
  • **kwargs (additional keyword arguments) — Keyword arguments to be passed to pandas.read_csv.

Create Dataset from CSV file(s).

Example:

>>> ds = Dataset.from_csv('path/to/dataset.csv')

from_json

< >

( path_or_paths: Union split: Optional = None features: Optional = None cache_dir: str = None keep_in_memory: bool = False field: Optional = None num_proc: Optional = None **kwargs )

Parameters

  • path_or_paths (path-like or list of path-like) — Path(s) of the JSON or JSON Lines file(s).
  • split (NamedSplit, optional) — Split name to be assigned to the dataset.
  • features (Features, optional) — Dataset features.
  • cache_dir (str, optional, defaults to "~/.cache/huggingface/datasets") — Directory to cache data.
  • keep_in_memory (bool, defaults to False) — Whether to copy the data in-memory.
  • field (str, optional) — Field name of the JSON file where the dataset is contained in.
  • num_proc (int, optional defaults to None) — Number of processes when downloading and generating the dataset locally. This is helpful if the dataset is made of multiple files. Multiprocessing is disabled by default.

    Added in 2.8.0
  • **kwargs (additional keyword arguments) — Keyword arguments to be passed to JsonConfig.

Create Dataset from JSON or JSON Lines file(s).

Example:

>>> ds = Dataset.from_json('path/to/dataset.json')

from_parquet

< >

( path_or_paths: Union split: Optional = None features: Optional = None cache_dir: str = None keep_in_memory: bool = False columns: Optional = None num_proc: Optional = None **kwargs )

Parameters

  • path_or_paths (path-like or list of path-like) — Path(s) of the Parquet file(s).
  • split (NamedSplit, optional) — Split name to be assigned to the dataset.
  • features (Features, optional) — Dataset features.
  • cache_dir (str, optional, defaults to "~/.cache/huggingface/datasets") — Directory to cache data.
  • keep_in_memory (bool, defaults to False) — Whether to copy the data in-memory.
  • columns (List[str], optional) — If not None, only these columns will be read from the file. A column name may be a prefix of a nested field, e.g. ‘a’ will select ‘a.b’, ‘a.c’, and ‘a.d.e’.
  • num_proc (int, optional, defaults to None) — Number of processes when downloading and generating the dataset locally. This is helpful if the dataset is made of multiple files. Multiprocessing is disabled by default.

    Added in 2.8.0
  • **kwargs (additional keyword arguments) — Keyword arguments to be passed to ParquetConfig.

Create Dataset from Parquet file(s).

Example:

>>> ds = Dataset.from_parquet('path/to/dataset.parquet')

from_text

< >

( path_or_paths: Union split: Optional = None features: Optional = None cache_dir: str = None keep_in_memory: bool = False num_proc: Optional = None **kwargs )

Parameters

  • path_or_paths (path-like or list of path-like) — Path(s) of the text file(s).
  • split (NamedSplit, optional) — Split name to be assigned to the dataset.
  • features (Features, optional) — Dataset features.
  • cache_dir (str, optional, defaults to "~/.cache/huggingface/datasets") — Directory to cache data.
  • keep_in_memory (bool, defaults to False) — Whether to copy the data in-memory.
  • num_proc (int, optional, defaults to None) — Number of processes when downloading and generating the dataset locally. This is helpful if the dataset is made of multiple files. Multiprocessing is disabled by default.

    Added in 2.8.0
  • **kwargs (additional keyword arguments) — Keyword arguments to be passed to TextConfig.

Create Dataset from text file(s).

Example:

>>> ds = Dataset.from_text('path/to/dataset.txt')

from_sql

< >

( sql: Union con: Union features: Optional = None cache_dir: str = None keep_in_memory: bool = False **kwargs )

Parameters

  • sql (str or sqlalchemy.sql.Selectable) — SQL query to be executed or a table name.
  • con (str or sqlite3.Connection or sqlalchemy.engine.Connection or sqlalchemy.engine.Connection) — A URI string used to instantiate a database connection or a SQLite3/SQLAlchemy connection object.
  • features (Features, optional) — Dataset features.
  • cache_dir (str, optional, defaults to "~/.cache/huggingface/datasets") — Directory to cache data.
  • keep_in_memory (bool, defaults to False) — Whether to copy the data in-memory.
  • **kwargs (additional keyword arguments) — Keyword arguments to be passed to SqlConfig.

Create Dataset from SQL query or database table.

Example:

>>> # Fetch a database table
>>> ds = Dataset.from_sql("test_data", "postgres:///db_name")
>>> # Execute a SQL query on the table
>>> ds = Dataset.from_sql("SELECT sentence FROM test_data", "postgres:///db_name")
>>> # Use a Selectable object to specify the query
>>> from sqlalchemy import select, text
>>> stmt = select([text("sentence")]).select_from(text("test_data"))
>>> ds = Dataset.from_sql(stmt, "postgres:///db_name")

The returned dataset can only be cached if con is specified as URI string.

prepare_for_task

< >

( task: Union id: int = 0 )

Parameters

  • task (Union[str, TaskTemplate]) — The task to prepare the dataset for during training and evaluation. If str, supported tasks include:

    • "text-classification"
    • "question-answering"

    If TaskTemplate, must be one of the task templates in datasets.tasks.

  • id (int, defaults to 0) — The id required to unambiguously identify the task template when multiple task templates of the same type are supported.

Prepare a dataset for the given task by casting the dataset’s Features to standardized column names and types as detailed in datasets.tasks.

Casts datasets.DatasetInfo.features according to a task-specific schema. Intended for single-use only, so all task templates are removed from datasets.DatasetInfo.task_templates after casting.

align_labels_with_mapping

< >

( label2id: Dict label_column: str )

Parameters

  • label2id (dict) — The label name to ID mapping to align the dataset with.
  • label_column (str) — The column name of labels to align on.

Align the dataset’s label ID and label name mapping to match an input label2id mapping. This is useful when you want to ensure that a model’s predicted labels are aligned with the dataset. The alignment in done using the lowercase label names.

Example:

>>> # dataset with mapping {'entailment': 0, 'neutral': 1, 'contradiction': 2}
>>> ds = load_dataset("glue", "mnli", split="train")
>>> # mapping to align with
>>> label2id = {'CONTRADICTION': 0, 'NEUTRAL': 1, 'ENTAILMENT': 2}
>>> ds_aligned = ds.align_labels_with_mapping(label2id, "label")

datasets.concatenate_datasets

< >

( dsets: List info: Optional = None split: Optional = None axis: int = 0 )

Parameters

  • dsets (List[datasets.Dataset]) — List of Datasets to concatenate.
  • info (DatasetInfo, optional) — Dataset information, like description, citation, etc.
  • split (NamedSplit, optional) — Name of the dataset split.
  • axis ({0, 1}, defaults to 0) — Axis to concatenate over, where 0 means over rows (vertically) and 1 means over columns (horizontally).

    Added in 1.6.0

Converts a list of Dataset with the same schema into a single Dataset.

Example:

>>> ds3 = concatenate_datasets([ds1, ds2])

datasets.interleave_datasets

< >

( datasets: List probabilities: Optional = None seed: Optional = None info: Optional = None split: Optional = None stopping_strategy: Literal = 'first_exhausted' ) → Dataset or IterableDataset

Parameters

  • datasets (List[Dataset] or List[IterableDataset]) — List of datasets to interleave.
  • probabilities (List[float], optional, defaults to None) — If specified, the new dataset is constructed by sampling examples from one source at a time according to these probabilities.
  • seed (int, optional, defaults to None) — The random seed used to choose a source for each example.
  • info (DatasetInfo, optional) — Dataset information, like description, citation, etc.

    Added in 2.4.0
  • split (NamedSplit, optional) — Name of the dataset split.

    Added in 2.4.0
  • stopping_strategy (str, defaults to first_exhausted) — Two strategies are proposed right now, first_exhausted and all_exhausted. By default, first_exhausted is an undersampling strategy, i.e the dataset construction is stopped as soon as one dataset has ran out of samples. If the strategy is all_exhausted, we use an oversampling strategy, i.e the dataset construction is stopped as soon as every samples of every dataset has been added at least once. Note that if the strategy is all_exhausted, the interleaved dataset size can get enormous:
    • with no probabilities, the resulting dataset will have max_length_datasets*nb_dataset samples.
    • with given probabilities, the resulting dataset will have more samples if some datasets have really low probability of visiting.

Returns

Dataset or IterableDataset

Return type depends on the input datasets parameter. Dataset if the input is a list of Dataset, IterableDataset if the input is a list of IterableDataset.

Interleave several datasets (sources) into a single dataset. The new dataset is constructed by alternating between the sources to get the examples.

You can use this function on a list of Dataset objects, or on a list of IterableDataset objects.

  • If probabilities is None (default) the new dataset is constructed by cycling between each source to get the examples.
  • If probabilities is not None, the new dataset is constructed by getting examples from a random source at a time according to the provided probabilities.

The resulting dataset ends when one of the source datasets runs out of examples except when oversampling is True, in which case, the resulting dataset ends when all datasets have ran out of examples at least one time.

Note for iterable datasets:

In a distributed setup or in PyTorch DataLoader workers, the stopping strategy is applied per process. Therefore the “first_exhausted” strategy on an sharded iterable dataset can generate less samples in total (up to 1 missing sample per subdataset per worker).

Example:

For regular datasets (map-style):

>>> from datasets import Dataset, interleave_datasets
>>> d1 = Dataset.from_dict({"a": [0, 1, 2]})
>>> d2 = Dataset.from_dict({"a": [10, 11, 12]})
>>> d3 = Dataset.from_dict({"a": [20, 21, 22]})
>>> dataset = interleave_datasets([d1, d2, d3], probabilities=[0.7, 0.2, 0.1], seed=42, stopping_strategy="all_exhausted")
>>> dataset["a"]
[10, 0, 11, 1, 2, 20, 12, 10, 0, 1, 2, 21, 0, 11, 1, 2, 0, 1, 12, 2, 10, 0, 22]
>>> dataset = interleave_datasets([d1, d2, d3], probabilities=[0.7, 0.2, 0.1], seed=42)
>>> dataset["a"]
[10, 0, 11, 1, 2]
>>> dataset = interleave_datasets([d1, d2, d3])
>>> dataset["a"]
[0, 10, 20, 1, 11, 21, 2, 12, 22]
>>> dataset = interleave_datasets([d1, d2, d3], stopping_strategy="all_exhausted")
>>> dataset["a"]
[0, 10, 20, 1, 11, 21, 2, 12, 22]
>>> d1 = Dataset.from_dict({"a": [0, 1, 2]})
>>> d2 = Dataset.from_dict({"a": [10, 11, 12, 13]})
>>> d3 = Dataset.from_dict({"a": [20, 21, 22, 23, 24]})
>>> dataset = interleave_datasets([d1, d2, d3])
>>> dataset["a"]
[0, 10, 20, 1, 11, 21, 2, 12, 22]
>>> dataset = interleave_datasets([d1, d2, d3], stopping_strategy="all_exhausted")
>>> dataset["a"]
[0, 10, 20, 1, 11, 21, 2, 12, 22, 0, 13, 23, 1, 10, 24]
>>> dataset = interleave_datasets([d1, d2, d3], probabilities=[0.7, 0.2, 0.1], seed=42)
>>> dataset["a"]
[10, 0, 11, 1, 2]
>>> dataset = interleave_datasets([d1, d2, d3], probabilities=[0.7, 0.2, 0.1], seed=42, stopping_strategy="all_exhausted")
>>> dataset["a"]
[10, 0, 11, 1, 2, 20, 12, 13, ..., 0, 1, 2, 0, 24]
For datasets in streaming mode (iterable):

>>> from datasets import load_dataset, interleave_datasets
>>> d1 = load_dataset("oscar", "unshuffled_deduplicated_en", split="train", streaming=True)
>>> d2 = load_dataset("oscar", "unshuffled_deduplicated_fr", split="train", streaming=True)
>>> dataset = interleave_datasets([d1, d2])
>>> iterator = iter(dataset)
>>> next(iterator)
{'text': 'Mtendere Village was inspired by the vision...}
>>> next(iterator)
{'text': "Média de débat d'idées, de culture...}

datasets.distributed.split_dataset_by_node

< >

( dataset: DatasetType rank: int world_size: int ) → Dataset or IterableDataset

Parameters

  • dataset (Dataset or IterableDataset) — The dataset to split by node.
  • rank (int) — Rank of the current node.
  • world_size (int) — Total number of nodes.

Returns

Dataset or IterableDataset

The dataset to be used on the node at rank rank.

Split a dataset for the node at rank rank in a pool of nodes of size world_size.

For map-style datasets:

Each node is assigned a chunk of data, e.g. rank 0 is given the first chunk of the dataset. To maximize data loading throughput, chunks are made of contiguous data on disk if possible.

For iterable datasets:

If the dataset has a number of shards that is a factor of world_size (i.e. if dataset.n_shards % world_size == 0), then the shards are evenly assigned across the nodes, which is the most optimized. Otherwise, each node keeps 1 example out of world_size, skipping the other examples.

datasets.enable_caching

< >

( )

When applying transforms on a dataset, the data are stored in cache files. The caching mechanism allows to reload an existing cache file if it’s already been computed.

Reloading a dataset is possible since the cache files are named using the dataset fingerprint, which is updated after each transform.

If disabled, the library will no longer reload cached datasets files when applying transforms to the datasets. More precisely, if the caching is disabled:

  • cache files are always recreated
  • cache files are written to a temporary directory that is deleted when session closes
  • cache files are named using a random hash instead of the dataset fingerprint
  • use save_to_disk() to save a transformed dataset or it will be deleted when session closes
  • caching doesn’t affect load_dataset(). If you want to regenerate a dataset from scratch you should use the download_mode parameter in load_dataset().

datasets.disable_caching

< >

( )

When applying transforms on a dataset, the data are stored in cache files. The caching mechanism allows to reload an existing cache file if it’s already been computed.

Reloading a dataset is possible since the cache files are named using the dataset fingerprint, which is updated after each transform.

If disabled, the library will no longer reload cached datasets files when applying transforms to the datasets. More precisely, if the caching is disabled:

  • cache files are always recreated
  • cache files are written to a temporary directory that is deleted when session closes
  • cache files are named using a random hash instead of the dataset fingerprint
  • use save_to_disk() to save a transformed dataset or it will be deleted when session closes
  • caching doesn’t affect load_dataset(). If you want to regenerate a dataset from scratch you should use the download_mode parameter in load_dataset().

datasets.is_caching_enabled

< >

( )

When applying transforms on a dataset, the data are stored in cache files. The caching mechanism allows to reload an existing cache file if it’s already been computed.

Reloading a dataset is possible since the cache files are named using the dataset fingerprint, which is updated after each transform.

If disabled, the library will no longer reload cached datasets files when applying transforms to the datasets. More precisely, if the caching is disabled:

  • cache files are always recreated
  • cache files are written to a temporary directory that is deleted when session closes
  • cache files are named using a random hash instead of the dataset fingerprint
  • use save_to_disk()] to save a transformed dataset or it will be deleted when session closes
  • caching doesn’t affect load_dataset(). If you want to regenerate a dataset from scratch you should use the download_mode parameter in load_dataset().

DatasetDict

Dictionary with split names as keys (‘train’, ‘test’ for example), and Dataset objects as values. It also has dataset transform methods like map or filter, to process all the splits at once.

class datasets.DatasetDict

< >

( )

A dictionary (dict of str: datasets.Dataset) with dataset transforms methods (map, filter, etc.)

data

< >

( )

The Apache Arrow tables backing each split.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds.data

cache_files

< >

( )

The cache files containing the Apache Arrow table backing each split.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds.cache_files
{'test': [{'filename': '/root/.cache/huggingface/datasets/rotten_tomatoes_movie_review/default/1.0.0/40d411e45a6ce3484deed7cc15b82a53dad9a72aafd9f86f8f227134bec5ca46/rotten_tomatoes_movie_review-test.arrow'}],
 'train': [{'filename': '/root/.cache/huggingface/datasets/rotten_tomatoes_movie_review/default/1.0.0/40d411e45a6ce3484deed7cc15b82a53dad9a72aafd9f86f8f227134bec5ca46/rotten_tomatoes_movie_review-train.arrow'}],
 'validation': [{'filename': '/root/.cache/huggingface/datasets/rotten_tomatoes_movie_review/default/1.0.0/40d411e45a6ce3484deed7cc15b82a53dad9a72aafd9f86f8f227134bec5ca46/rotten_tomatoes_movie_review-validation.arrow'}]}

num_columns

< >

( )

Number of columns in each split of the dataset.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds.num_columns
{'test': 2, 'train': 2, 'validation': 2}

num_rows

< >

( )

Number of rows in each split of the dataset.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds.num_rows
{'test': 1066, 'train': 8530, 'validation': 1066}

column_names

< >

( )

Names of the columns in each split of the dataset.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds.column_names
{'test': ['text', 'label'],
 'train': ['text', 'label'],
 'validation': ['text', 'label']}

shape

< >

( )

Shape of each split of the dataset (number of rows, number of columns).

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds.shape
{'test': (1066, 2), 'train': (8530, 2), 'validation': (1066, 2)}

unique

< >

( column: str ) → Dict[str, list]

Parameters

  • column (str) — column name (list all the column names with column_names)

Returns

Dict[str, list]

Dictionary of unique elements in the given column.

Return a list of the unique elements in a column for each split.

This is implemented in the low-level backend and as such, very fast.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds.unique("label")
{'test': [1, 0], 'train': [1, 0], 'validation': [1, 0]}

cleanup_cache_files

< >

( )

Clean up all cache files in the dataset cache directory, excepted the currently used cache file if there is one. Be careful when running this command that no other process is currently using other cache files.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds.cleanup_cache_files()
{'test': 0, 'train': 0, 'validation': 0}

map

< >

( function: Optional = None with_indices: bool = False with_rank: bool = False input_columns: Union = None batched: bool = False batch_size: Optional = 1000 drop_last_batch: bool = False remove_columns: Union = None keep_in_memory: bool = False load_from_cache_file: Optional = None cache_file_names: Optional = None writer_batch_size: Optional = 1000 features: Optional = None disable_nullable: bool = False fn_kwargs: Optional = None num_proc: Optional = None desc: Optional = None )

Parameters

  • function (callable) — with one of the following signature:

    • function(example: Dict[str, Any]) -> Dict[str, Any] if batched=False and with_indices=False
    • function(example: Dict[str, Any], indices: int) -> Dict[str, Any] if batched=False and with_indices=True
    • function(batch: Dict[str, List]) -> Dict[str, List] if batched=True and with_indices=False
    • function(batch: Dict[str, List], indices: List[int]) -> Dict[str, List] if batched=True and with_indices=True

    For advanced usage, the function can also return a pyarrow.Table. Moreover if your function returns nothing (None), then map will run your function and return the dataset unchanged.

  • with_indices (bool, defaults to False) — Provide example indices to function. Note that in this case the signature of function should be def function(example, idx): ....
  • with_rank (bool, defaults to False) — Provide process rank to function. Note that in this case the signature of function should be def function(example[, idx], rank): ....
  • input_columns ([Union[str, List[str]]], optional, defaults to None) — The columns to be passed into function as positional arguments. If None, a dict mapping to all formatted columns is passed as one argument.
  • batched (bool, defaults to False) — Provide batch of examples to function.
  • batch_size (int, optional, defaults to 1000) — Number of examples per batch provided to function if batched=True, batch_size <= 0 or batch_size == None then provide the full dataset as a single batch to function.
  • drop_last_batch (bool, defaults to False) — Whether a last batch smaller than the batch_size should be dropped instead of being processed by the function.
  • remove_columns ([Union[str, List[str]]], optional, defaults to None) — Remove a selection of columns while doing the mapping. Columns will be removed before updating the examples with the output of function, i.e. if function is adding columns with names in remove_columns, these columns will be kept.
  • keep_in_memory (bool, defaults to False) — Keep the dataset in memory instead of writing it to a cache file.
  • load_from_cache_file (Optional[bool], defaults to True if caching is enabled) — If a cache file storing the current computation from function can be identified, use it instead of recomputing.
  • cache_file_names ([Dict[str, str]], optional, defaults to None) — Provide the name of a path for the cache file. It is used to store the results of the computation instead of the automatically generated cache file name. You have to provide one cache_file_name per dataset in the dataset dictionary.
  • writer_batch_size (int, default 1000) — Number of rows per write operation for the cache file writer. This value is a good trade-off between memory usage during the processing, and processing speed. Higher value makes the processing do fewer lookups, lower value consume less temporary memory while running map.
  • features ([datasets.Features], optional, defaults to None) — Use a specific Features to store the cache file instead of the automatically generated one.
  • disable_nullable (bool, defaults to False) — Disallow null values in the table.
  • fn_kwargs (Dict, optional, defaults to None) — Keyword arguments to be passed to function
  • num_proc (int, optional, defaults to None) — Number of processes for multiprocessing. By default it doesn’t use multiprocessing.
  • desc (str, optional, defaults to None) — Meaningful description to be displayed alongside with the progress bar while mapping examples.

Apply a function to all the elements in the table (individually or in batches) and update the table (if function does updated examples). The transformation is applied to all the datasets of the dataset dictionary.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> def add_prefix(example):
...     example["text"] = "Review: " + example["text"]
...     return example
>>> ds = ds.map(add_prefix)
>>> ds["train"][0:3]["text"]
['Review: the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .',
 'Review: the gorgeously elaborate continuation of " the lord of the rings " trilogy is so huge that a column of words cannot adequately describe co-writer/director peter jackson's expanded vision of j . r . r . tolkien's middle-earth .',
 'Review: effective but too-tepid biopic']

# process a batch of examples
>>> ds = ds.map(lambda example: tokenizer(example["text"]), batched=True)
# set number of processors
>>> ds = ds.map(add_prefix, num_proc=4)

filter

< >

( function: Optional = None with_indices: bool = False with_rank: bool = False input_columns: Union = None batched: bool = False batch_size: Optional = 1000 keep_in_memory: bool = False load_from_cache_file: Optional = None cache_file_names: Optional = None writer_batch_size: Optional = 1000 fn_kwargs: Optional = None num_proc: Optional = None desc: Optional = None )

Parameters

  • function (Callable) — Callable with one of the following signatures:

    • function(example: Dict[str, Any]) -> bool if batched=False and with_indices=False and with_rank=False
    • function(example: Dict[str, Any], *extra_args) -> bool if batched=False and with_indices=True and/or with_rank=True (one extra arg for each)
    • function(batch: Dict[str, List]) -> List[bool] if batched=True and with_indices=False and with_rank=False
    • function(batch: Dict[str, List], *extra_args) -> List[bool] if batched=True and with_indices=True and/or with_rank=True (one extra arg for each)

    If no function is provided, defaults to an always True function: lambda x: True.

  • with_indices (bool, defaults to False) — Provide example indices to function. Note that in this case the signature of function should be def function(example, idx[, rank]): ....
  • with_rank (bool, defaults to False) — Provide process rank to function. Note that in this case the signature of function should be def function(example[, idx], rank): ....
  • input_columns ([Union[str, List[str]]], optional, defaults to None) — The columns to be passed into function as positional arguments. If None, a dict mapping to all formatted columns is passed as one argument.
  • batched (bool, defaults to False) — Provide batch of examples to function.
  • batch_size (int, optional, defaults to 1000) — Number of examples per batch provided to function if batched=True batch_size <= 0 or batch_size == None then provide the full dataset as a single batch to function.
  • keep_in_memory (bool, defaults to False) — Keep the dataset in memory instead of writing it to a cache file.
  • load_from_cache_file (Optional[bool], defaults to True if caching is enabled) — If a cache file storing the current computation from function can be identified, use it instead of recomputing.
  • cache_file_names ([Dict[str, str]], optional, defaults to None) — Provide the name of a path for the cache file. It is used to store the results of the computation instead of the automatically generated cache file name. You have to provide one cache_file_name per dataset in the dataset dictionary.
  • writer_batch_size (int, defaults to 1000) — Number of rows per write operation for the cache file writer. This value is a good trade-off between memory usage during the processing, and processing speed. Higher value makes the processing do fewer lookups, lower value consume less temporary memory while running map.
  • fn_kwargs (Dict, optional, defaults to None) — Keyword arguments to be passed to function
  • num_proc (int, optional, defaults to None) — Number of processes for multiprocessing. By default it doesn’t use multiprocessing.
  • desc (str, optional, defaults to None) — Meaningful description to be displayed alongside with the progress bar while filtering examples.

Apply a filter function to all the elements in the table in batches and update the table so that the dataset only includes examples according to the filter function. The transformation is applied to all the datasets of the dataset dictionary.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds.filter(lambda x: x["label"] == 1)
DatasetDict({
    train: Dataset({
        features: ['text', 'label'],
        num_rows: 4265
    })
    validation: Dataset({
        features: ['text', 'label'],
        num_rows: 533
    })
    test: Dataset({
        features: ['text', 'label'],
        num_rows: 533
    })
})

sort

< >

( column_names: Union reverse: Union = False kind = 'deprecated' null_placement: str = 'at_end' keep_in_memory: bool = False load_from_cache_file: Optional = None indices_cache_file_names: Optional = None writer_batch_size: Optional = 1000 )

Parameters

  • column_names (Union[str, Sequence[str]]) — Column name(s) to sort by.
  • reverse (Union[bool, Sequence[bool]], defaults to False) — If True, sort by descending order rather than ascending. If a single bool is provided, the value is applied to the sorting of all column names. Otherwise a list of bools with the same length and order as column_names must be provided.
  • kind (str, optional) — Pandas algorithm for sorting selected in {quicksort, mergesort, heapsort, stable}, The default is quicksort. Note that both stable and mergesort use timsort under the covers and, in general, the actual implementation will vary with data type. The mergesort option is retained for backwards compatibility.

    Deprecated in 2.8.0

    kind was deprecated in version 2.10.0 and will be removed in 3.0.0.
  • null_placement (str, defaults to at_end) — Put None values at the beginning if at_start or first or at the end if at_end or last
  • keep_in_memory (bool, defaults to False) — Keep the sorted indices in memory instead of writing it to a cache file.
  • load_from_cache_file (Optional[bool], defaults to True if caching is enabled) — If a cache file storing the sorted indices can be identified, use it instead of recomputing.
  • indices_cache_file_names ([Dict[str, str]], optional, defaults to None) — Provide the name of a path for the cache file. It is used to store the indices mapping instead of the automatically generated cache file name. You have to provide one cache_file_name per dataset in the dataset dictionary.
  • writer_batch_size (int, defaults to 1000) — Number of rows per write operation for the cache file writer. Higher value gives smaller cache files, lower value consume less temporary memory.

Create a new dataset sorted according to a single or multiple columns.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset('rotten_tomatoes')
>>> ds['train']['label'][:10]
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
>>> sorted_ds = ds.sort('label')
>>> sorted_ds['train']['label'][:10]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
>>> another_sorted_ds = ds.sort(['label', 'text'], reverse=[True, False])
>>> another_sorted_ds['train']['label'][:10]
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]

shuffle

< >

( seeds: Union = None seed: Optional = None generators: Optional = None keep_in_memory: bool = False load_from_cache_file: Optional = None indices_cache_file_names: Optional = None writer_batch_size: Optional = 1000 )

Parameters

  • seeds (Dict[str, int] or int, optional) — A seed to initialize the default BitGenerator if generator=None. If None, then fresh, unpredictable entropy will be pulled from the OS. If an int or array_like[ints] is passed, then it will be passed to SeedSequence to derive the initial BitGenerator state. You can provide one seed per dataset in the dataset dictionary.
  • seed (int, optional) — A seed to initialize the default BitGenerator if generator=None. Alias for seeds (a ValueError is raised if both are provided).
  • generators (Dict[str, *optional*, np.random.Generator]) — Numpy random Generator to use to compute the permutation of the dataset rows. If generator=None (default), uses np.random.default_rng (the default BitGenerator (PCG64) of NumPy). You have to provide one generator per dataset in the dataset dictionary.
  • keep_in_memory (bool, defaults to False) — Keep the dataset in memory instead of writing it to a cache file.
  • load_from_cache_file (Optional[bool], defaults to True if caching is enabled) — If a cache file storing the current computation from function can be identified, use it instead of recomputing.
  • indices_cache_file_names (Dict[str, str], optional) — Provide the name of a path for the cache file. It is used to store the indices mappings instead of the automatically generated cache file name. You have to provide one cache_file_name per dataset in the dataset dictionary.
  • writer_batch_size (int, defaults to 1000) — Number of rows per write operation for the cache file writer. This value is a good trade-off between memory usage during the processing, and processing speed. Higher value makes the processing do fewer lookups, lower value consume less temporary memory while running map.

Create a new Dataset where the rows are shuffled.

The transformation is applied to all the datasets of the dataset dictionary.

Currently shuffling uses numpy random generators. You can either supply a NumPy BitGenerator to use, or a seed to initiate NumPy’s default random generator (PCG64).

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds["train"]["label"][:10]
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1]

# set a seed
>>> shuffled_ds = ds.shuffle(seed=42)
>>> shuffled_ds["train"]["label"][:10]
[0, 1, 0, 1, 0, 0, 0, 0, 0, 0]

set_format

< >

( type: Optional = None columns: Optional = None output_all_columns: bool = False **format_kwargs )

Parameters

  • type (str, optional) — Output type selected in [None, 'numpy', 'torch', 'tensorflow', 'pandas', 'arrow', 'jax']. None means __getitem__ returns python objects (default).
  • columns (List[str], optional) — Columns to format in the output. None means __getitem__ returns all columns (default).
  • output_all_columns (bool, defaults to False) — Keep un-formatted columns as well in the output (as python objects),
  • **format_kwargs (additional keyword arguments) — Keywords arguments passed to the convert function like np.array, torch.tensor or tensorflow.ragged.constant.

Set __getitem__ return format (type and columns). The format is set for every dataset in the dataset dictionary.

It is possible to call map after calling set_format. Since map may add new columns, then the list of formatted columns gets updated. In this case, if you apply map on a dataset to add a new column, then this column will be formatted:

new formatted columns = (all columns - previously unformatted columns)

Example:

>>> from datasets import load_dataset
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
>>> ds = ds.map(lambda x: tokenizer(x["text"], truncation=True, padding=True), batched=True)
>>> ds.set_format(type="numpy", columns=['input_ids', 'token_type_ids', 'attention_mask', 'label'])
>>> ds["train"].format
{'columns': ['input_ids', 'token_type_ids', 'attention_mask', 'label'],
 'format_kwargs': {},
 'output_all_columns': False,
 'type': 'numpy'}

reset_format

< >

( )

Reset __getitem__ return format to python objects and all columns. The transformation is applied to all the datasets of the dataset dictionary.

Same as self.set_format()

Example:

>>> from datasets import load_dataset
>>> from transformers import AutoTokenizer
>>> ds = load_dataset("rotten_tomatoes")
>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
>>> ds = ds.map(lambda x: tokenizer(x["text"], truncation=True, padding=True), batched=True)
>>> ds.set_format(type="numpy", columns=['input_ids', 'token_type_ids', 'attention_mask', 'label'])
>>> ds["train"].format
{'columns': ['input_ids', 'token_type_ids', 'attention_mask', 'label'],
 'format_kwargs': {},
 'output_all_columns': False,
 'type': 'numpy'}
>>> ds.reset_format()
>>> ds["train"].format
{'columns': ['text', 'label', 'input_ids', 'token_type_ids', 'attention_mask'],
 'format_kwargs': {},
 'output_all_columns': False,
 'type': None}

formatted_as

< >

( type: Optional = None columns: Optional = None output_all_columns: bool = False **format_kwargs )

Parameters

  • type (str, optional) — Output type selected in [None, 'numpy', 'torch', 'tensorflow', 'pandas', 'arrow', 'jax']. None means __getitem__ returns python objects (default).
  • columns (List[str], optional) — Columns to format in the output. None means __getitem__ returns all columns (default).
  • output_all_columns (bool, defaults to False) — Keep un-formatted columns as well in the output (as python objects).
  • **format_kwargs (additional keyword arguments) — Keywords arguments passed to the convert function like np.array, torch.tensor or tensorflow.ragged.constant.

To be used in a with statement. Set __getitem__ return format (type and columns). The transformation is applied to all the datasets of the dataset dictionary.

with_format

< >

( type: Optional = None columns: Optional = None output_all_columns: bool = False **format_kwargs )

Parameters

  • type (str, optional) — Output type selected in [None, 'numpy', 'torch', 'tensorflow', 'pandas', 'arrow', 'jax']. None means __getitem__ returns python objects (default).
  • columns (List[str], optional) — Columns to format in the output. None means __getitem__ returns all columns (default).
  • output_all_columns (bool, defaults to False) — Keep un-formatted columns as well in the output (as python objects).
  • **format_kwargs (additional keyword arguments) — Keywords arguments passed to the convert function like np.array, torch.tensor or tensorflow.ragged.constant.

Set __getitem__ return format (type and columns). The data formatting is applied on-the-fly. The format type (for example “numpy”) is used to format batches when using __getitem__. The format is set for every dataset in the dataset dictionary.

It’s also possible to use custom transforms for formatting using with_transform().

Contrary to set_format(), with_format returns a new DatasetDict object with new Dataset objects.

Example:

>>> from datasets import load_dataset
>>> from transformers import AutoTokenizer
>>> ds = load_dataset("rotten_tomatoes")
>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
>>> ds = ds.map(lambda x: tokenizer(x['text'], truncation=True, padding=True), batched=True)
>>> ds["train"].format
{'columns': ['text', 'label', 'input_ids', 'token_type_ids', 'attention_mask'],
 'format_kwargs': {},
 'output_all_columns': False,
 'type': None}
>>> ds = ds.with_format(type='tensorflow', columns=['input_ids', 'token_type_ids', 'attention_mask', 'label'])
>>> ds["train"].format
{'columns': ['input_ids', 'token_type_ids', 'attention_mask', 'label'],
 'format_kwargs': {},
 'output_all_columns': False,
 'type': 'tensorflow'}

with_transform

< >

( transform: Optional columns: Optional = None output_all_columns: bool = False )

Parameters

  • transform (Callable, optional) — User-defined formatting transform, replaces the format defined by set_format(). A formatting function is a callable that takes a batch (as a dict) as input and returns a batch. This function is applied right before returning the objects in __getitem__.
  • columns (List[str], optional) — Columns to format in the output. If specified, then the input batch of the transform only contains those columns.
  • output_all_columns (bool, defaults to False) — Keep un-formatted columns as well in the output (as python objects). If set to True, then the other un-formatted columns are kept with the output of the transform.

Set __getitem__ return format using this transform. The transform is applied on-the-fly on batches when __getitem__ is called. The transform is set for every dataset in the dataset dictionary

As set_format(), this can be reset using reset_format().

Contrary to set_transform(), with_transform returns a new DatasetDict object with new Dataset objects.

Example:

>>> from datasets import load_dataset
>>> from transformers import AutoTokenizer
>>> ds = load_dataset("rotten_tomatoes")
>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
>>> def encode(example):
...     return tokenizer(example['text'], truncation=True, padding=True, return_tensors="pt")
>>> ds = ds.with_transform(encode)
>>> ds["train"][0]
{'attention_mask': tensor([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
 1, 1, 1, 1, 1, 1, 1, 1, 1]),
 'input_ids': tensor([  101,  1103,  2067,  1110, 17348,  1106,  1129,  1103,  6880,  1432,
        112,   188,  1207,   107, 14255,  1389,   107,  1105,  1115,  1119,
        112,   188,  1280,  1106,  1294,   170, 24194,  1256,  3407,  1190,
        170, 11791,  5253,   188,  1732,  7200, 10947, 12606,  2895,   117,
        179,  7766,   118,   172, 15554,  1181,  3498,  6961,  3263,  1137,
        188,  1566,  7912, 14516,  6997,   119,   102]),
 'token_type_ids': tensor([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0])}

flatten

< >

( max_depth = 16 )

Flatten the Apache Arrow Table of each split (nested features are flatten). Each column with a struct type is flattened into one column per struct field. Other columns are left unchanged.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("squad")
>>> ds["train"].features
{'answers': Sequence(feature={'text': Value(dtype='string', id=None), 'answer_start': Value(dtype='int32', id=None)}, length=-1, id=None),
 'context': Value(dtype='string', id=None),
 'id': Value(dtype='string', id=None),
 'question': Value(dtype='string', id=None),
 'title': Value(dtype='string', id=None)}
>>> ds.flatten()
DatasetDict({
    train: Dataset({
        features: ['id', 'title', 'context', 'question', 'answers.text', 'answers.answer_start'],
        num_rows: 87599
    })
    validation: Dataset({
        features: ['id', 'title', 'context', 'question', 'answers.text', 'answers.answer_start'],
        num_rows: 10570
    })
})

cast

< >

( features: Features )

Parameters

  • features (Features) — New features to cast the dataset to. The name and order of the fields in the features must match the current column names. The type of the data must also be convertible from one type to the other. For non-trivial conversion, e.g. string <-> ClassLabel you should use map() to update the dataset.

Cast the dataset to a new set of features. The transformation is applied to all the datasets of the dataset dictionary.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds["train"].features
{'label': ClassLabel(num_classes=2, names=['neg', 'pos'], id=None),
 'text': Value(dtype='string', id=None)}
>>> new_features = ds["train"].features.copy()
>>> new_features['label'] = ClassLabel(names=['bad', 'good'])
>>> new_features['text'] = Value('large_string')
>>> ds = ds.cast(new_features)
>>> ds["train"].features
{'label': ClassLabel(num_classes=2, names=['bad', 'good'], id=None),
 'text': Value(dtype='large_string', id=None)}

cast_column

< >

( column: str feature )

Parameters

  • column (str) — Column name.
  • feature (Feature) — Target feature.

Cast column to feature for decoding.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds["train"].features
{'label': ClassLabel(num_classes=2, names=['neg', 'pos'], id=None),
 'text': Value(dtype='string', id=None)}
>>> ds = ds.cast_column('label', ClassLabel(names=['bad', 'good']))
>>> ds["train"].features
{'label': ClassLabel(num_classes=2, names=['bad', 'good'], id=None),
 'text': Value(dtype='string', id=None)}

remove_columns

< >

( column_names: Union ) → DatasetDict

Parameters

  • column_names (Union[str, List[str]]) — Name of the column(s) to remove.

Returns

DatasetDict

A copy of the dataset object without the columns to remove.

Remove one or several column(s) from each split in the dataset and the features associated to the column(s).

The transformation is applied to all the splits of the dataset dictionary.

You can also remove a column using map() with remove_columns but the present method doesn’t copy the data of the remaining columns and is thus faster.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds = ds.remove_columns("label")
DatasetDict({
    train: Dataset({
        features: ['text'],
        num_rows: 8530
    })
    validation: Dataset({
        features: ['text'],
        num_rows: 1066
    })
    test: Dataset({
        features: ['text'],
        num_rows: 1066
    })
})

rename_column

< >

( original_column_name: str new_column_name: str )

Parameters

  • original_column_name (str) — Name of the column to rename.
  • new_column_name (str) — New name for the column.

Rename a column in the dataset and move the features associated to the original column under the new column name. The transformation is applied to all the datasets of the dataset dictionary.

You can also rename a column using map() with remove_columns but the present method:

  • takes care of moving the original features under the new column name.
  • doesn’t copy the data to a new dataset and is thus much faster.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds = ds.rename_column("label", "label_new")
DatasetDict({
    train: Dataset({
        features: ['text', 'label_new'],
        num_rows: 8530
    })
    validation: Dataset({
        features: ['text', 'label_new'],
        num_rows: 1066
    })
    test: Dataset({
        features: ['text', 'label_new'],
        num_rows: 1066
    })
})

rename_columns

< >

( column_mapping: Dict ) → DatasetDict

Parameters

  • column_mapping (Dict[str, str]) — A mapping of columns to rename to their new names.

Returns

DatasetDict

A copy of the dataset with renamed columns.

Rename several columns in the dataset, and move the features associated to the original columns under the new column names. The transformation is applied to all the datasets of the dataset dictionary.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds.rename_columns({'text': 'text_new', 'label': 'label_new'})
DatasetDict({
    train: Dataset({
        features: ['text_new', 'label_new'],
        num_rows: 8530
    })
    validation: Dataset({
        features: ['text_new', 'label_new'],
        num_rows: 1066
    })
    test: Dataset({
        features: ['text_new', 'label_new'],
        num_rows: 1066
    })
})

select_columns

< >

( column_names: Union )

Parameters

  • column_names (Union[str, List[str]]) — Name of the column(s) to keep.

Select one or several column(s) from each split in the dataset and the features associated to the column(s).

The transformation is applied to all the splits of the dataset dictionary.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes")
>>> ds.select_columns("text")
DatasetDict({
    train: Dataset({
        features: ['text'],
        num_rows: 8530
    })
    validation: Dataset({
        features: ['text'],
        num_rows: 1066
    })
    test: Dataset({
        features: ['text'],
        num_rows: 1066
    })
})

class_encode_column

< >

( column: str include_nulls: bool = False )

Parameters

  • column (str) — The name of the column to cast.
  • include_nulls (bool, defaults to False) — Whether to include null values in the class labels. If True, the null values will be encoded as the "None" class label.

    Added in 1.14.2

Casts the given column as ClassLabel and updates the tables.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("boolq")
>>> ds["train"].features
{'answer': Value(dtype='bool', id=None),
 'passage': Value(dtype='string', id=None),
 'question': Value(dtype='string', id=None)}
>>> ds = ds.class_encode_column("answer")
>>> ds["train"].features
{'answer': ClassLabel(num_classes=2, names=['False', 'True'], id=None),
 'passage': Value(dtype='string', id=None),
 'question': Value(dtype='string', id=None)}

push_to_hub

< >

( repo_id config_name: str = 'default' set_default: Optional = None data_dir: Optional = None commit_message: Optional = None commit_description: Optional = None private: Optional = False token: Optional = None revision: Optional = None branch = 'deprecated' create_pr: Optional = False max_shard_size: Union = None num_shards: Optional = None embed_external_files: bool = True )

Parameters

  • repo_id (str) — The ID of the repository to push to in the following format: <user>/<dataset_name> or <org>/<dataset_name>. Also accepts <dataset_name>, which will default to the namespace of the logged-in user.
  • config_name (str) — Configuration name of a dataset. Defaults to “default”.
  • set_default (bool, optional) — Whether to set this configuration as the default one. Otherwise, the default configuration is the one named “default”.
  • data_dir (str, optional) — Directory name that will contain the uploaded data files. Defaults to the config_name if different from “default”, else “data”.

    Added in 2.17.0
  • commit_message (str, optional) — Message to commit while pushing. Will default to "Upload dataset".
  • commit_description (str, optional) — Description of the commit that will be created. Additionally, description of the PR if a PR is created (create_pr is True).

    Added in 2.16.0
  • private (bool, optional) — Whether the dataset repository should be set to private or not. Only affects repository creation: a repository that already exists will not be affected by that parameter.
  • token (str, optional) — An optional authentication token for the Hugging Face Hub. If no token is passed, will default to the token saved locally when logging in with huggingface-cli login. Will raise an error if no token is passed and the user is not logged-in.
  • revision (str, optional) — Branch to push the uploaded files to. Defaults to the "main" branch.

    Added in 2.15.0
  • branch (str, optional) — The git branch on which to push the dataset. This defaults to the default branch as specified in your repository, which defaults to "main".

    Deprecated in 2.15.0

    branch was deprecated in favor of revision in version 2.15.0 and will be removed in 3.0.0.
  • create_pr (bool, optional, defaults to False) — Whether to create a PR with the uploaded files or directly commit.

    Added in 2.15.0
  • max_shard_size (int or str, optional, defaults to "500MB") — The maximum size of the dataset shards to be uploaded to the hub. If expressed as a string, needs to be digits followed by a unit (like "500MB" or "1GB").
  • num_shards (Dict[str, int], optional) — Number of shards to write. By default, the number of shards depends on max_shard_size. Use a dictionary to define a different num_shards for each split.

    Added in 2.8.0
  • embed_external_files (bool, defaults to True) — Whether to embed file bytes in the shards. In particular, this will do the following before the push for the fields of type:

    • Audio and Image removes local path information and embed file content in the Parquet files.

Pushes the DatasetDict to the hub as a Parquet dataset. The DatasetDict is pushed using HTTP requests and does not need to have neither git or git-lfs installed.

Each dataset split will be pushed independently. The pushed dataset will keep the original split names.

The resulting Parquet files are self-contained by default: if your dataset contains Image or Audio data, the Parquet files will store the bytes of your images or audio files. You can disable this by setting embed_external_files to False.

Example:

>>> dataset_dict.push_to_hub("<organization>/<dataset_id>")
>>> dataset_dict.push_to_hub("<organization>/<dataset_id>", private=True)
>>> dataset_dict.push_to_hub("<organization>/<dataset_id>", max_shard_size="1GB")
>>> dataset_dict.push_to_hub("<organization>/<dataset_id>", num_shards={"train": 1024, "test": 8})

If you want to add a new configuration (or subset) to a dataset (e.g. if the dataset has multiple tasks/versions/languages):

>>> english_dataset.push_to_hub("<organization>/<dataset_id>", "en")
>>> french_dataset.push_to_hub("<organization>/<dataset_id>", "fr")
>>> # later
>>> english_dataset = load_dataset("<organization>/<dataset_id>", "en")
>>> french_dataset = load_dataset("<organization>/<dataset_id>", "fr")

save_to_disk

< >

( dataset_dict_path: Union fs = 'deprecated' max_shard_size: Union = None num_shards: Optional = None num_proc: Optional = None storage_options: Optional = None )

Parameters

  • dataset_dict_path (path-like) — Path (e.g. dataset/train) or remote URI (e.g. s3://my-bucket/dataset/train) of the dataset dict directory where the dataset dict will be saved to.
  • fs (fsspec.spec.AbstractFileSystem, optional) — Instance of the remote filesystem where the dataset will be saved to.

    Deprecated in 2.8.0

    fs was deprecated in version 2.8.0 and will be removed in 3.0.0. Please use storage_options instead, e.g. storage_options=fs.storage_options
  • max_shard_size (int or str, optional, defaults to "500MB") — The maximum size of the dataset shards to be uploaded to the hub. If expressed as a string, needs to be digits followed by a unit (like "50MB").
  • num_shards (Dict[str, int], optional) — Number of shards to write. By default the number of shards depends on max_shard_size and num_proc. You need to provide the number of shards for each dataset in the dataset dictionary. Use a dictionary to define a different num_shards for each split.

    Added in 2.8.0
  • num_proc (int, optional, default None) — Number of processes when downloading and generating the dataset locally. Multiprocessing is disabled by default.

    Added in 2.8.0
  • storage_options (dict, optional) — Key/value pairs to be passed on to the file-system backend, if any.

    Added in 2.8.0

Saves a dataset dict to a filesystem using fsspec.spec.AbstractFileSystem.

For Image and Audio data:

All the Image() and Audio() data are stored in the arrow files. If you want to store paths or urls, please use the Value(“string”) type.

Example:

>>> dataset_dict.save_to_disk("path/to/dataset/directory")
>>> dataset_dict.save_to_disk("path/to/dataset/directory", max_shard_size="1GB")
>>> dataset_dict.save_to_disk("path/to/dataset/directory", num_shards={"train": 1024, "test": 8})

load_from_disk

< >

( dataset_dict_path: Union fs = 'deprecated' keep_in_memory: Optional = None storage_options: Optional = None )

Parameters

  • dataset_dict_path (path-like) — Path (e.g. "dataset/train") or remote URI (e.g. "s3//my-bucket/dataset/train") of the dataset dict directory where the dataset dict will be loaded from.
  • fs (fsspec.spec.AbstractFileSystem, optional) — Instance of the remote filesystem where the dataset will be saved to.

    Deprecated in 2.8.0

    fs was deprecated in version 2.8.0 and will be removed in 3.0.0. Please use storage_options instead, e.g. storage_options=fs.storage_options
  • keep_in_memory (bool, defaults to None) — Whether to copy the dataset in-memory. If None, the dataset will not be copied in-memory unless explicitly enabled by setting datasets.config.IN_MEMORY_MAX_SIZE to nonzero. See more details in the improve performance section.
  • storage_options (dict, optional) — Key/value pairs to be passed on to the file-system backend, if any.

    Added in 2.8.0

Load a dataset that was previously saved using save_to_disk from a filesystem using fsspec.spec.AbstractFileSystem.

Example:

>>> ds = load_from_disk('path/to/dataset/directory')

from_csv

< >

( path_or_paths: Dict features: Optional = None cache_dir: str = None keep_in_memory: bool = False **kwargs )

Parameters

  • path_or_paths (dict of path-like) — Path(s) of the CSV file(s).
  • features (Features, optional) — Dataset features.
  • cache_dir (str, optional, defaults to "~/.cache/huggingface/datasets") — Directory to cache data.
  • keep_in_memory (bool, defaults to False) — Whether to copy the data in-memory.
  • **kwargs (additional keyword arguments) — Keyword arguments to be passed to pandas.read_csv.

Create DatasetDict from CSV file(s).

Example:

>>> from datasets import DatasetDict
>>> ds = DatasetDict.from_csv({'train': 'path/to/dataset.csv'})

from_json

< >

( path_or_paths: Dict features: Optional = None cache_dir: str = None keep_in_memory: bool = False **kwargs )

Parameters

  • path_or_paths (path-like or list of path-like) — Path(s) of the JSON Lines file(s).
  • features (Features, optional) — Dataset features.
  • cache_dir (str, optional, defaults to "~/.cache/huggingface/datasets") — Directory to cache data.
  • keep_in_memory (bool, defaults to False) — Whether to copy the data in-memory.
  • **kwargs (additional keyword arguments) — Keyword arguments to be passed to JsonConfig.

Create DatasetDict from JSON Lines file(s).

Example:

>>> from datasets import DatasetDict
>>> ds = DatasetDict.from_json({'train': 'path/to/dataset.json'})

from_parquet

< >

( path_or_paths: Dict features: Optional = None cache_dir: str = None keep_in_memory: bool = False columns: Optional = None **kwargs )

Parameters

  • path_or_paths (dict of path-like) — Path(s) of the CSV file(s).
  • features (Features, optional) — Dataset features.
  • cache_dir (str, optional, defaults to "~/.cache/huggingface/datasets") — Directory to cache data.
  • keep_in_memory (bool, defaults to False) — Whether to copy the data in-memory.
  • columns (List[str], optional) — If not None, only these columns will be read from the file. A column name may be a prefix of a nested field, e.g. ‘a’ will select ‘a.b’, ‘a.c’, and ‘a.d.e’.
  • **kwargs (additional keyword arguments) — Keyword arguments to be passed to ParquetConfig.

Create DatasetDict from Parquet file(s).

Example:

>>> from datasets import DatasetDict
>>> ds = DatasetDict.from_parquet({'train': 'path/to/dataset/parquet'})

from_text

< >

( path_or_paths: Dict features: Optional = None cache_dir: str = None keep_in_memory: bool = False **kwargs )

Parameters

  • path_or_paths (dict of path-like) — Path(s) of the text file(s).
  • features (Features, optional) — Dataset features.
  • cache_dir (str, optional, defaults to "~/.cache/huggingface/datasets") — Directory to cache data.
  • keep_in_memory (bool, defaults to False) — Whether to copy the data in-memory.
  • **kwargs (additional keyword arguments) — Keyword arguments to be passed to TextConfig.

Create DatasetDict from text file(s).

Example:

>>> from datasets import DatasetDict
>>> ds = DatasetDict.from_text({'train': 'path/to/dataset.txt'})

prepare_for_task

< >

( task: Union id: int = 0 )

Parameters

  • task (Union[str, TaskTemplate]) — The task to prepare the dataset for during training and evaluation. If str, supported tasks include:

    • "text-classification"
    • "question-answering"

    If TaskTemplate, must be one of the task templates in datasets.tasks.

  • id (int, defaults to 0) — The id required to unambiguously identify the task template when multiple task templates of the same type are supported.

Prepare a dataset for the given task by casting the dataset’s Features to standardized column names and types as detailed in datasets.tasks.

Casts datasets.DatasetInfo.features according to a task-specific schema. Intended for single-use only, so all task templates are removed from datasets.DatasetInfo.task_templates after casting.

IterableDataset

The base class IterableDataset implements an iterable Dataset backed by python generators.

class datasets.IterableDataset

< >

( ex_iterable: _BaseExamplesIterable info: Optional = None split: Optional = None formatting: Optional = None shuffling: Optional = None distributed: Optional = None token_per_repo_id: Optional = None format_type = 'deprecated' )

A Dataset backed by an iterable.

from_generator

< >

( generator: Callable features: Optional = None gen_kwargs: Optional = None split: NamedSplit = NamedSplit('train') ) → IterableDataset

Parameters

  • generator (Callable) — A generator function that yields examples.
  • features (Features, optional) — Dataset features.
  • gen_kwargs(dict, optional) — Keyword arguments to be passed to the generator callable. You can define a sharded iterable dataset by passing the list of shards in gen_kwargs. This can be used to improve shuffling and when iterating over the dataset with multiple workers.
  • split (NamedSplit, defaults to Split.TRAIN) — Split name to be assigned to the dataset.

    Added in 2.21.0

Returns

IterableDataset

Create an Iterable Dataset from a generator.

Example:

>>> def gen():
...     yield {"text": "Good", "label": 0}
...     yield {"text": "Bad", "label": 1}
...
>>> ds = IterableDataset.from_generator(gen)
>>> def gen(shards):
...     for shard in shards:
...         with open(shard) as f:
...             for line in f:
...                 yield {"line": line}
...
>>> shards = [f"data{i}.txt" for i in range(32)]
>>> ds = IterableDataset.from_generator(gen, gen_kwargs={"shards": shards})
>>> ds = ds.shuffle(seed=42, buffer_size=10_000)  # shuffles the shards order + uses a shuffle buffer
>>> from torch.utils.data import DataLoader
>>> dataloader = DataLoader(ds.with_format("torch"), num_workers=4)  # give each worker a subset of 32/4=8 shards

remove_columns

< >

( column_names: Union ) → IterableDataset

Parameters

  • column_names (Union[str, List[str]]) — Name of the column(s) to remove.

Returns

IterableDataset

A copy of the dataset object without the columns to remove.

Remove one or several column(s) in the dataset and the features associated to them. The removal is done on-the-fly on the examples when iterating over the dataset.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train", streaming=True)
>>> next(iter(ds))
{'text': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .', 'label': 1}
>>> ds = ds.remove_columns("label")
>>> next(iter(ds))
{'text': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'}

select_columns

< >

( column_names: Union ) → IterableDataset

Parameters

  • column_names (Union[str, List[str]]) — Name of the column(s) to select.

Returns

IterableDataset

A copy of the dataset object with selected columns.

Select one or several column(s) in the dataset and the features associated to them. The selection is done on-the-fly on the examples when iterating over the dataset.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train", streaming=True)
>>> next(iter(ds))
{'text': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .', 'label': 1}
>>> ds = ds.select_columns("text")
>>> next(iter(ds))
{'text': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'}

cast_column

< >

( column: str feature: Union ) → IterableDataset

Parameters

  • column (str) — Column name.
  • feature (Feature) — Target feature.

Returns

IterableDataset

Cast column to feature for decoding.

Example:

>>> from datasets import load_dataset, Audio
>>> ds = load_dataset("PolyAI/minds14", name="en-US", split="train", streaming=True)
>>> ds.features
{'audio': Audio(sampling_rate=8000, mono=True, decode=True, id=None),
 'english_transcription': Value(dtype='string', id=None),
 'intent_class': ClassLabel(num_classes=14, names=['abroad', 'address', 'app_error', 'atm_limit', 'balance', 'business_loan',  'card_issues', 'cash_deposit', 'direct_debit', 'freeze', 'high_value_payment', 'joint_account', 'latest_transactions', 'pay_bill'], id=None),
 'lang_id': ClassLabel(num_classes=14, names=['cs-CZ', 'de-DE', 'en-AU', 'en-GB', 'en-US', 'es-ES', 'fr-FR', 'it-IT', 'ko-KR',  'nl-NL', 'pl-PL', 'pt-PT', 'ru-RU', 'zh-CN'], id=None),
 'path': Value(dtype='string', id=None),
 'transcription': Value(dtype='string', id=None)}
>>> ds = ds.cast_column("audio", Audio(sampling_rate=16000))
>>> ds.features
{'audio': Audio(sampling_rate=16000, mono=True, decode=True, id=None),
 'english_transcription': Value(dtype='string', id=None),
 'intent_class': ClassLabel(num_classes=14, names=['abroad', 'address', 'app_error', 'atm_limit', 'balance', 'business_loan',  'card_issues', 'cash_deposit', 'direct_debit', 'freeze', 'high_value_payment', 'joint_account', 'latest_transactions', 'pay_bill'], id=None),
 'lang_id': ClassLabel(num_classes=14, names=['cs-CZ', 'de-DE', 'en-AU', 'en-GB', 'en-US', 'es-ES', 'fr-FR', 'it-IT', 'ko-KR',  'nl-NL', 'pl-PL', 'pt-PT', 'ru-RU', 'zh-CN'], id=None),
 'path': Value(dtype='string', id=None),
 'transcription': Value(dtype='string', id=None)}

cast

< >

( features: Features ) → IterableDataset

Parameters

  • features (Features) — New features to cast the dataset to. The name of the fields in the features must match the current column names. The type of the data must also be convertible from one type to the other. For non-trivial conversion, e.g. string <-> ClassLabel you should use map() to update the Dataset.

Returns

IterableDataset

A copy of the dataset with casted features.

Cast the dataset to a new set of features.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train", streaming=True)
>>> ds.features
{'label': ClassLabel(num_classes=2, names=['neg', 'pos'], id=None),
 'text': Value(dtype='string', id=None)}
>>> new_features = ds.features.copy()
>>> new_features["label"] = ClassLabel(names=["bad", "good"])
>>> new_features["text"] = Value("large_string")
>>> ds = ds.cast(new_features)
>>> ds.features
{'label': ClassLabel(num_classes=2, names=['bad', 'good'], id=None),
 'text': Value(dtype='large_string', id=None)}

__iter__

< >

( )

iter

< >

( batch_size: int drop_last_batch: bool = False )

Parameters

  • batch_size (int) — size of each batch to yield.
  • drop_last_batch (bool, default False) — Whether a last batch smaller than the batch_size should be dropped

Iterate through the batches of size batch_size.

map

< >

( function: Optional = None with_indices: bool = False input_columns: Union = None batched: bool = False batch_size: Optional = 1000 drop_last_batch: bool = False remove_columns: Union = None features: Optional = None fn_kwargs: Optional = None )

Parameters

  • function (Callable, optional, defaults to None) — Function applied on-the-fly on the examples when you iterate on the dataset. It must have one of the following signatures:

    • function(example: Dict[str, Any]) -> Dict[str, Any] if batched=False and with_indices=False
    • function(example: Dict[str, Any], idx: int) -> Dict[str, Any] if batched=False and with_indices=True
    • function(batch: Dict[str, List]) -> Dict[str, List] if batched=True and with_indices=False
    • function(batch: Dict[str, List], indices: List[int]) -> Dict[str, List] if batched=True and with_indices=True

    For advanced usage, the function can also return a pyarrow.Table. Moreover if your function returns nothing (None), then map will run your function and return the dataset unchanged. If no function is provided, default to identity function: lambda x: x.

  • with_indices (bool, defaults to False) — Provide example indices to function. Note that in this case the signature of function should be def function(example, idx[, rank]): ....
  • input_columns (Optional[Union[str, List[str]]], defaults to None) — The columns to be passed into function as positional arguments. If None, a dict mapping to all formatted columns is passed as one argument.
  • batched (bool, defaults to False) — Provide batch of examples to function.
  • batch_size (int, optional, defaults to 1000) — Number of examples per batch provided to function if batched=True. batch_size <= 0 or batch_size == None then provide the full dataset as a single batch to function.
  • drop_last_batch (bool, defaults to False) — Whether a last batch smaller than the batch_size should be dropped instead of being processed by the function.
  • remove_columns ([List[str]], optional, defaults to None) — Remove a selection of columns while doing the mapping. Columns will be removed before updating the examples with the output of function, i.e. if function is adding columns with names in remove_columns, these columns will be kept.
  • features ([Features], optional, defaults to None) — Feature types of the resulting dataset.
  • fn_kwargs (Dict, optional, default None) — Keyword arguments to be passed to function.

Apply a function to all the examples in the iterable dataset (individually or in batches) and update them. If your function returns a column that already exists, then it overwrites it. The function is applied on-the-fly on the examples when iterating over the dataset.

You can specify whether the function should be batched or not with the batched parameter:

  • If batched is False, then the function takes 1 example in and should return 1 example. An example is a dictionary, e.g. {"text": "Hello there !"}.
  • If batched is True and batch_size is 1, then the function takes a batch of 1 example as input and can return a batch with 1 or more examples. A batch is a dictionary, e.g. a batch of 1 example is {“text”: [“Hello there !”]}.
  • If batched is True and batch_size is n > 1, then the function takes a batch of n examples as input and can return a batch with n examples, or with an arbitrary number of examples. Note that the last batch may have less than n examples. A batch is a dictionary, e.g. a batch of n examples is {"text": ["Hello there !"] * n}.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train", streaming=True)
>>> def add_prefix(example):
...     example["text"] = "Review: " + example["text"]
...     return example
>>> ds = ds.map(add_prefix)
>>> list(ds.take(3))
[{'label': 1,
 'text': 'Review: the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'},
 {'label': 1,
 'text': 'Review: the gorgeously elaborate continuation of " the lord of the rings " trilogy is so huge that a column of words cannot adequately describe co-writer/director peter jackson's expanded vision of j . r . r . tolkien's middle-earth .'},
 {'label': 1, 'text': 'Review: effective but too-tepid biopic'}]

rename_column

< >

( original_column_name: str new_column_name: str ) → IterableDataset

Parameters

  • original_column_name (str) — Name of the column to rename.
  • new_column_name (str) — New name for the column.

Returns

IterableDataset

A copy of the dataset with a renamed column.

Rename a column in the dataset, and move the features associated to the original column under the new column name.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train", streaming=True)
>>> next(iter(ds))
{'label': 1,
 'text': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'}
>>> ds = ds.rename_column("text", "movie_review")
>>> next(iter(ds))
{'label': 1,
 'movie_review': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'}

filter

< >

( function: Optional = None with_indices = False input_columns: Union = None batched: bool = False batch_size: Optional = 1000 fn_kwargs: Optional = None )

Parameters

  • function (Callable) — Callable with one of the following signatures:

    • function(example: Dict[str, Any]) -> bool if with_indices=False, batched=False
    • function(example: Dict[str, Any], indices: int) -> bool if with_indices=True, batched=False
    • function(example: Dict[str, List]) -> List[bool] if with_indices=False, batched=True
    • function(example: Dict[str, List], indices: List[int]) -> List[bool] if with_indices=True, batched=True

    If no function is provided, defaults to an always True function: lambda x: True.

  • with_indices (bool, defaults to False) — Provide example indices to function. Note that in this case the signature of function should be def function(example, idx): ....
  • input_columns (str or List[str], optional) — The columns to be passed into function as positional arguments. If None, a dict mapping to all formatted columns is passed as one argument.
  • batched (bool, defaults to False) — Provide batch of examples to function.
  • batch_size (int, optional, default 1000) — Number of examples per batch provided to function if batched=True.
  • fn_kwargs (Dict, optional, default None) — Keyword arguments to be passed to function.

Apply a filter function to all the elements so that the dataset only includes examples according to the filter function. The filtering is done on-the-fly when iterating over the dataset.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train", streaming=True)
>>> ds = ds.filter(lambda x: x["label"] == 0)
>>> list(ds.take(3))
[{'label': 0, 'movie_review': 'simplistic , silly and tedious .'},
 {'label': 0,
 'movie_review': "it's so laddish and juvenile , only teenage boys could possibly find it funny ."},
 {'label': 0,
 'movie_review': 'exploitative and largely devoid of the depth or sophistication that would make watching such a graphic treatment of the crimes bearable .'}]

shuffle

< >

( seed = None generator: Optional = None buffer_size: int = 1000 )

Parameters

  • seed (int, optional, defaults to None) — Random seed that will be used to shuffle the dataset. It is used to sample from the shuffle buffer and also to shuffle the data shards.
  • generator (numpy.random.Generator, optional) — Numpy random Generator to use to compute the permutation of the dataset rows. If generator=None (default), uses np.random.default_rng (the default BitGenerator (PCG64) of NumPy).
  • buffer_size (int, defaults to 1000) — Size of the buffer.

Randomly shuffles the elements of this dataset.

This dataset fills a buffer with buffer_size elements, then randomly samples elements from this buffer, replacing the selected elements with new elements. For perfect shuffling, a buffer size greater than or equal to the full size of the dataset is required.

For instance, if your dataset contains 10,000 elements but buffer_size is set to 1000, then shuffle will initially select a random element from only the first 1000 elements in the buffer. Once an element is selected, its space in the buffer is replaced by the next (i.e. 1,001-st) element, maintaining the 1000 element buffer.

If the dataset is made of several shards, it also does shuffle the order of the shards. However if the order has been fixed by using skip() or take() then the order of the shards is kept unchanged.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train", streaming=True)
>>> list(ds.take(3))
[{'label': 1,
 'text': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'},
 {'label': 1,
 'text': 'the gorgeously elaborate continuation of " the lord of the rings " trilogy is so huge that a column of words cannot adequately describe co-writer/director peter jackson's expanded vision of j . r . r . tolkien's middle-earth .'},
 {'label': 1, 'text': 'effective but too-tepid biopic'}]
>>> shuffled_ds = ds.shuffle(seed=42)
>>> list(shuffled_ds.take(3))
[{'label': 1,
 'text': "a sports movie with action that's exciting on the field and a story you care about off it ."},
 {'label': 1,
 'text': 'at its best , the good girl is a refreshingly adult take on adultery . . .'},
 {'label': 1,
 'text': "sam jones became a very lucky filmmaker the day wilco got dropped from their record label , proving that one man's ruin may be another's fortune ."}]

batch

< >

( batch_size: int drop_last_batch: bool = False )

Parameters

  • batch_size (int) — The number of samples in each batch.
  • drop_last_batch (bool, defaults to False) — Whether to drop the last incomplete batch.

Group samples from the dataset into batches.

Example:

>>> ds = load_dataset("some_dataset", streaming=True)
>>> batched_ds = ds.batch(batch_size=32)

skip

< >

( n: int )

Parameters

  • n (int) — Number of elements to skip.

Create a new IterableDataset that skips the first n elements.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train", streaming=True)
>>> list(ds.take(3))
[{'label': 1,
 'text': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'},
 {'label': 1,
 'text': 'the gorgeously elaborate continuation of " the lord of the rings " trilogy is so huge that a column of words cannot adequately describe co-writer/director peter jackson's expanded vision of j . r . r . tolkien's middle-earth .'},
 {'label': 1, 'text': 'effective but too-tepid biopic'}]
>>> ds = ds.skip(1)
>>> list(ds.take(3))
[{'label': 1,
 'text': 'the gorgeously elaborate continuation of " the lord of the rings " trilogy is so huge that a column of words cannot adequately describe co-writer/director peter jackson's expanded vision of j . r . r . tolkien's middle-earth .'},
 {'label': 1, 'text': 'effective but too-tepid biopic'},
 {'label': 1,
 'text': 'if you sometimes like to go to the movies to have fun , wasabi is a good place to start .'}]

take

< >

( n: int )

Parameters

  • n (int) — Number of elements to take.

Create a new IterableDataset with only the first n elements.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train", streaming=True)
>>> small_ds = ds.take(2)
>>> list(small_ds)
[{'label': 1,
 'text': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'},
 {'label': 1,
 'text': 'the gorgeously elaborate continuation of " the lord of the rings " trilogy is so huge that a column of words cannot adequately describe co-writer/director peter jackson's expanded vision of j . r . r . tolkien's middle-earth .'}]

load_state_dict

< >

( state_dict: dict )

Load the state_dict of the dataset. The iteration will restart at the next example from when the state was saved.

Resuming returns exactly where the checkpoint was saved except in two cases:

  1. examples from shuffle buffers are lost when resuming and the buffers are refilled with new data
  2. combinations of .with_format(arrow) and batched .map() may skip one batch.

Example:

>>> from datasets import Dataset, concatenate_datasets
>>> ds = Dataset.from_dict({"a": range(6)}).to_iterable_dataset(num_shards=3)
>>> for idx, example in enumerate(ds):
...     print(example)
...     if idx == 2:
...         state_dict = ds.state_dict()
...         print("checkpoint")
...         break
>>> ds.load_state_dict(state_dict)
>>> print(f"restart from checkpoint")
>>> for example in ds:
...     print(example)

which returns:

{'a': 0}
{'a': 1}
{'a': 2}
checkpoint
restart from checkpoint
{'a': 3}
{'a': 4}
{'a': 5}
>>> from torchdata.stateful_dataloader import StatefulDataLoader
>>> ds = load_dataset("deepmind/code_contests", streaming=True, split="train")
>>> dataloader = StatefulDataLoader(ds, batch_size=32, num_workers=4)
>>> # checkpoint
>>> state_dict = dataloader.state_dict()  # uses ds.state_dict() under the hood
>>> # resume from checkpoint
>>> dataloader.load_state_dict(state_dict)  # uses ds.load_state_dict() under the hood

state_dict

< >

( ) → dict

Returns

dict

Get the current state_dict of the dataset. It corresponds to the state at the latest example it yielded.

Resuming returns exactly where the checkpoint was saved except in two cases:

  1. examples from shuffle buffers are lost when resuming and the buffers are refilled with new data
  2. combinations of .with_format(arrow) and batched .map() may skip one batch.

Example:

>>> from datasets import Dataset, concatenate_datasets
>>> ds = Dataset.from_dict({"a": range(6)}).to_iterable_dataset(num_shards=3)
>>> for idx, example in enumerate(ds):
...     print(example)
...     if idx == 2:
...         state_dict = ds.state_dict()
...         print("checkpoint")
...         break
>>> ds.load_state_dict(state_dict)
>>> print(f"restart from checkpoint")
>>> for example in ds:
...     print(example)

which returns:

{'a': 0}
{'a': 1}
{'a': 2}
checkpoint
restart from checkpoint
{'a': 3}
{'a': 4}
{'a': 5}
>>> from torchdata.stateful_dataloader import StatefulDataLoader
>>> ds = load_dataset("deepmind/code_contests", streaming=True, split="train")
>>> dataloader = StatefulDataLoader(ds, batch_size=32, num_workers=4)
>>> # checkpoint
>>> state_dict = dataloader.state_dict()  # uses ds.state_dict() under the hood
>>> # resume from checkpoint
>>> dataloader.load_state_dict(state_dict)  # uses ds.load_state_dict() under the hood

info

< >

( )

DatasetInfo object containing all the metadata in the dataset.

split

< >

( )

NamedSplit object corresponding to a named dataset split.

builder_name

< >

( )

citation

< >

( )

config_name

< >

( )

dataset_size

< >

( )

description

< >

( )

download_checksums

< >

( )

download_size

< >

( )

features

< >

( )

homepage

< >

( )

license

< >

( )

size_in_bytes

< >

( )

supervised_keys

< >

( )

version

< >

( )

IterableDatasetDict

Dictionary with split names as keys (‘train’, ‘test’ for example), and IterableDataset objects as values.

class datasets.IterableDatasetDict

< >

( )

map

< >

( function: Optional = None with_indices: bool = False input_columns: Union = None batched: bool = False batch_size: int = 1000 drop_last_batch: bool = False remove_columns: Union = None fn_kwargs: Optional = None )

Parameters

  • function (Callable, optional, defaults to None) — Function applied on-the-fly on the examples when you iterate on the dataset. It must have one of the following signatures:

    • function(example: Dict[str, Any]) -> Dict[str, Any] if batched=False and with_indices=False
    • function(example: Dict[str, Any], idx: int) -> Dict[str, Any] if batched=False and with_indices=True
    • function(batch: Dict[str, List]) -> Dict[str, List] if batched=True and with_indices=False
    • function(batch: Dict[str, List], indices: List[int]) -> Dict[str, List] if batched=True and with_indices=True

    For advanced usage, the function can also return a pyarrow.Table. Moreover if your function returns nothing (None), then map will run your function and return the dataset unchanged. If no function is provided, default to identity function: lambda x: x.

  • with_indices (bool, defaults to False) — Provide example indices to function. Note that in this case the signature of function should be def function(example, idx[, rank]): ....
  • input_columns ([Union[str, List[str]]], optional, defaults to None) — The columns to be passed into function as positional arguments. If None, a dict mapping to all formatted columns is passed as one argument.
  • batched (bool, defaults to False) — Provide batch of examples to function.
  • batch_size (int, optional, defaults to 1000) — Number of examples per batch provided to function if batched=True.
  • drop_last_batch (bool, defaults to False) — Whether a last batch smaller than the batch_size should be dropped instead of being processed by the function.
  • remove_columns ([List[str]], optional, defaults to None) — Remove a selection of columns while doing the mapping. Columns will be removed before updating the examples with the output of function, i.e. if function is adding columns with names in remove_columns, these columns will be kept.
  • fn_kwargs (Dict, optional, defaults to None) — Keyword arguments to be passed to function

Apply a function to all the examples in the iterable dataset (individually or in batches) and update them. If your function returns a column that already exists, then it overwrites it. The function is applied on-the-fly on the examples when iterating over the dataset. The transformation is applied to all the datasets of the dataset dictionary.

You can specify whether the function should be batched or not with the batched parameter:

  • If batched is False, then the function takes 1 example in and should return 1 example. An example is a dictionary, e.g. {"text": "Hello there !"}.
  • If batched is True and batch_size is 1, then the function takes a batch of 1 example as input and can return a batch with 1 or more examples. A batch is a dictionary, e.g. a batch of 1 example is {"text": ["Hello there !"]}.
  • If batched is True and batch_size is n > 1, then the function takes a batch of n examples as input and can return a batch with n examples, or with an arbitrary number of examples. Note that the last batch may have less than n examples. A batch is a dictionary, e.g. a batch of n examples is {"text": ["Hello there !"] * n}.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", streaming=True)
>>> def add_prefix(example):
...     example["text"] = "Review: " + example["text"]
...     return example
>>> ds = ds.map(add_prefix)
>>> next(iter(ds["train"]))
{'label': 1,
 'text': 'Review: the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'}

filter

< >

( function: Optional = None with_indices = False input_columns: Union = None batched: bool = False batch_size: Optional = 1000 fn_kwargs: Optional = None )

Parameters

  • function (Callable) — Callable with one of the following signatures:

    • function(example: Dict[str, Any]) -> bool if with_indices=False, batched=False
    • function(example: Dict[str, Any], indices: int) -> bool if with_indices=True, batched=False
    • function(example: Dict[str, List]) -> List[bool] if with_indices=False, batched=True
    • function(example: Dict[str, List], indices: List[int]) -> List[bool] if with_indices=True, batched=True

    If no function is provided, defaults to an always True function: lambda x: True.

  • with_indices (bool, defaults to False) — Provide example indices to function. Note that in this case the signature of function should be def function(example, idx): ....
  • input_columns (str or List[str], optional) — The columns to be passed into function as positional arguments. If None, a dict mapping to all formatted columns is passed as one argument.
  • batched (bool, defaults to False) — Provide batch of examples to function
  • batch_size (int, optional, defaults to 1000) — Number of examples per batch provided to function if batched=True.
  • fn_kwargs (Dict, optional, defaults to None) — Keyword arguments to be passed to function

Apply a filter function to all the elements so that the dataset only includes examples according to the filter function. The filtering is done on-the-fly when iterating over the dataset. The filtering is applied to all the datasets of the dataset dictionary.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", streaming=True)
>>> ds = ds.filter(lambda x: x["label"] == 0)
>>> list(ds["train"].take(3))
[{'label': 0, 'text': 'Review: simplistic , silly and tedious .'},
 {'label': 0,
 'text': "Review: it's so laddish and juvenile , only teenage boys could possibly find it funny ."},
 {'label': 0,
 'text': 'Review: exploitative and largely devoid of the depth or sophistication that would make watching such a graphic treatment of the crimes bearable .'}]

shuffle

< >

( seed = None generator: Optional = None buffer_size: int = 1000 )

Parameters

  • seed (int, optional, defaults to None) — Random seed that will be used to shuffle the dataset. It is used to sample from the shuffle buffer and also to shuffle the data shards.
  • generator (numpy.random.Generator, optional) — Numpy random Generator to use to compute the permutation of the dataset rows. If generator=None (default), uses np.random.default_rng (the default BitGenerator (PCG64) of NumPy).
  • buffer_size (int, defaults to 1000) — Size of the buffer.

Randomly shuffles the elements of this dataset. The shuffling is applied to all the datasets of the dataset dictionary.

This dataset fills a buffer with buffer_size elements, then randomly samples elements from this buffer, replacing the selected elements with new elements. For perfect shuffling, a buffer size greater than or equal to the full size of the dataset is required.

For instance, if your dataset contains 10,000 elements but buffer_size is set to 1000, then shuffle will initially select a random element from only the first 1000 elements in the buffer. Once an element is selected, its space in the buffer is replaced by the next (i.e. 1,001-st) element, maintaining the 1000 element buffer.

If the dataset is made of several shards, it also does shuffle the order of the shards. However if the order has been fixed by using skip() or take() then the order of the shards is kept unchanged.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", streaming=True)
>>> list(ds["train"].take(3))
[{'label': 1,
 'text': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'},
 {'label': 1,
 'text': 'the gorgeously elaborate continuation of " the lord of the rings " trilogy is so huge that a column of words cannot adequately describe co-writer/director peter jackson's expanded vision of j . r . r . tolkien's middle-earth .'},
 {'label': 1, 'text': 'effective but too-tepid biopic'}]
>>> ds = ds.shuffle(seed=42)
>>> list(ds["train"].take(3))
[{'label': 1,
 'text': "a sports movie with action that's exciting on the field and a story you care about off it ."},
 {'label': 1,
 'text': 'at its best , the good girl is a refreshingly adult take on adultery . . .'},
 {'label': 1,
 'text': "sam jones became a very lucky filmmaker the day wilco got dropped from their record label , proving that one man's ruin may be another's fortune ."}]

with_format

< >

( type: Optional = None )

Parameters

  • type (str, optional, defaults to None) — If set to “torch”, the returned dataset will be a subclass of torch.utils.data.IterableDataset to be used in a DataLoader.

Return a dataset with the specified format. This method only supports the “torch” format for now. The format is set to all the datasets of the dataset dictionary.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", streaming=True)
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
>>> def encode(example):
...     return tokenizer(examples["text"], truncation=True, padding="max_length")
>>> ds = ds.map(encode, batched=True, remove_columns=["text"])
>>> ds = ds.with_format("torch")

cast

< >

( features: Features ) → IterableDatasetDict

Parameters

  • features (Features) — New features to cast the dataset to. The name of the fields in the features must match the current column names. The type of the data must also be convertible from one type to the other. For non-trivial conversion, e.g. string <-> ClassLabel you should use map to update the Dataset.

Returns

IterableDatasetDict

A copy of the dataset with casted features.

Cast the dataset to a new set of features. The type casting is applied to all the datasets of the dataset dictionary.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", streaming=True)
>>> ds["train"].features
{'label': ClassLabel(num_classes=2, names=['neg', 'pos'], id=None),
 'text': Value(dtype='string', id=None)}
>>> new_features = ds["train"].features.copy()
>>> new_features['label'] = ClassLabel(names=['bad', 'good'])
>>> new_features['text'] = Value('large_string')
>>> ds = ds.cast(new_features)
>>> ds["train"].features
{'label': ClassLabel(num_classes=2, names=['bad', 'good'], id=None),
 'text': Value(dtype='large_string', id=None)}

cast_column

< >

( column: str feature: Union )

Parameters

  • column (str) — Column name.
  • feature (Feature) — Target feature.

Cast column to feature for decoding. The type casting is applied to all the datasets of the dataset dictionary.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", streaming=True)
>>> ds["train"].features
{'label': ClassLabel(num_classes=2, names=['neg', 'pos'], id=None),
 'text': Value(dtype='string', id=None)}
>>> ds = ds.cast_column('label', ClassLabel(names=['bad', 'good']))
>>> ds["train"].features
{'label': ClassLabel(num_classes=2, names=['bad', 'good'], id=None),
 'text': Value(dtype='string', id=None)}

remove_columns

< >

( column_names: Union ) → IterableDatasetDict

Parameters

  • column_names (Union[str, List[str]]) — Name of the column(s) to remove.

Returns

IterableDatasetDict

A copy of the dataset object without the columns to remove.

Remove one or several column(s) in the dataset and the features associated to them. The removal is done on-the-fly on the examples when iterating over the dataset. The removal is applied to all the datasets of the dataset dictionary.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", streaming=True)
>>> ds = ds.remove_columns("label")
>>> next(iter(ds["train"]))
{'text': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'}

rename_column

< >

( original_column_name: str new_column_name: str ) → IterableDatasetDict

Parameters

  • original_column_name (str) — Name of the column to rename.
  • new_column_name (str) — New name for the column.

Returns

IterableDatasetDict

A copy of the dataset with a renamed column.

Rename a column in the dataset, and move the features associated to the original column under the new column name. The renaming is applied to all the datasets of the dataset dictionary.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", streaming=True)
>>> ds = ds.rename_column("text", "movie_review")
>>> next(iter(ds["train"]))
{'label': 1,
 'movie_review': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'}

rename_columns

< >

( column_mapping: Dict ) → IterableDatasetDict

Parameters

  • column_mapping (Dict[str, str]) — A mapping of columns to rename to their new names.

Returns

IterableDatasetDict

A copy of the dataset with renamed columns

Rename several columns in the dataset, and move the features associated to the original columns under the new column names. The renaming is applied to all the datasets of the dataset dictionary.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", streaming=True)
>>> ds = ds.rename_columns({"text": "movie_review", "label": "rating"})
>>> next(iter(ds["train"]))
{'movie_review': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .',
 'rating': 1}

select_columns

< >

( column_names: Union ) → IterableDatasetDict

Parameters

  • column_names (Union[str, List[str]]) — Name of the column(s) to keep.

Returns

IterableDatasetDict

A copy of the dataset object with only selected columns.

Select one or several column(s) in the dataset and the features associated to them. The selection is done on-the-fly on the examples when iterating over the dataset. The selection is applied to all the datasets of the dataset dictionary.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", streaming=True)
>>> ds = ds.select("text")
>>> next(iter(ds["train"]))
{'text': 'the rock is destined to be the 21st century's new " conan " and that he's going to make a splash even greater than arnold schwarzenegger , jean-claud van damme or steven segal .'}

Features

class datasets.Features

< >

( *args **kwargs )

A special dictionary that defines the internal structure of a dataset.

Instantiated with a dictionary of type dict[str, FieldType], where keys are the desired column names, and values are the type of that column.

FieldType can be one of the following:

  • a Value feature specifies a single typed value, e.g. int64 or string.

  • a ClassLabel feature specifies a field with a predefined set of classes which can have labels associated to them and will be stored as integers in the dataset.

  • a python dict which specifies that the field is a nested field containing a mapping of sub-fields to sub-fields features. It’s possible to have nested fields of nested fields in an arbitrary manner.

  • a python list or a Sequence specifies that the field contains a list of objects. The python list or Sequence should be provided with a single sub-feature as an example of the feature type hosted in this list.

    A Sequence with a internal dictionary feature will be automatically converted into a dictionary of lists. This behavior is implemented to have a compatibility layer with the TensorFlow Datasets library but may be un-wanted in some cases. If you don’t want this behavior, you can use a python list instead of the Sequence.

  • a Array2D, Array3D, Array4D or Array5D feature for multidimensional arrays.

  • an Audio feature to store the absolute path to an audio file or a dictionary with the relative path to an audio file (“path” key) and its bytes content (“bytes” key). This feature extracts the audio data.

  • an Image feature to store the absolute path to an image file, an np.ndarray object, a PIL.Image.Image object or a dictionary with the relative path to an image file (“path” key) and its bytes content (“bytes” key). This feature extracts the image data.

  • Translation and TranslationVariableLanguages, the two features specific to Machine Translation.

copy

< >

( )

Make a deep copy of Features.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train")
>>> copy_of_features = ds.features.copy()
>>> copy_of_features
{'label': ClassLabel(num_classes=2, names=['neg', 'pos'], id=None),
 'text': Value(dtype='string', id=None)}

decode_batch

< >

( batch: dict token_per_repo_id: Optional = None )

Parameters

  • batch (dict[str, list[Any]]) — Dataset batch data.
  • token_per_repo_id (dict, optional) — To access and decode audio or image files from private repositories on the Hub, you can pass a dictionary repo_id (str) -> token (bool or str)

Decode batch with custom feature decoding.

decode_column

< >

( column: list column_name: str )

Parameters

  • column (list[Any]) — Dataset column data.
  • column_name (str) — Dataset column name.

Decode column with custom feature decoding.

decode_example

< >

( example: dict token_per_repo_id: Optional = None )

Parameters

  • example (dict[str, Any]) — Dataset row data.
  • token_per_repo_id (dict, optional) — To access and decode audio or image files from private repositories on the Hub, you can pass a dictionary repo_id (str) -> token (bool or str).

Decode example with custom feature decoding.

encode_batch

< >

( batch )

Parameters

  • batch (dict[str, list[Any]]) — Data in a Dataset batch.

Encode batch into a format for Arrow.

encode_column

< >

( column column_name: str )

Parameters

  • column (list[Any]) — Data in a Dataset column.
  • column_name (str) — Dataset column name.

Encode column into a format for Arrow.

encode_example

< >

( example )

Parameters

  • example (dict[str, Any]) — Data in a Dataset row.

Encode example into a format for Arrow.

flatten

< >

( max_depth = 16 ) → Features

Returns

Features

The flattened features.

Flatten the features. Every dictionary column is removed and is replaced by all the subfields it contains. The new fields are named by concatenating the name of the original column and the subfield name like this: <original>.<subfield>.

If a column contains nested dictionaries, then all the lower-level subfields names are also concatenated to form new columns: <original>.<subfield>.<subsubfield>, etc.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("squad", split="train")
>>> ds.features.flatten()
{'answers.answer_start': Sequence(feature=Value(dtype='int32', id=None), length=-1, id=None),
 'answers.text': Sequence(feature=Value(dtype='string', id=None), length=-1, id=None),
 'context': Value(dtype='string', id=None),
 'id': Value(dtype='string', id=None),
 'question': Value(dtype='string', id=None),
 'title': Value(dtype='string', id=None)}

from_arrow_schema

< >

( pa_schema: Schema )

Parameters

  • pa_schema (pyarrow.Schema) — Arrow Schema.

Construct Features from Arrow Schema. It also checks the schema metadata for Hugging Face Datasets features. Non-nullable fields are not supported and set to nullable.

Also, pa.dictionary is not supported and it uses its underlying type instead. Therefore datasets convert DictionaryArray objects to their actual values.

from_dict

< >

( dic ) → Features

Parameters

  • dic (dict[str, Any]) — Python dictionary.

Returns

Features

Construct [Features] from dict.

Regenerate the nested feature object from a deserialized dict. We use the _type key to infer the dataclass name of the feature FieldType.

It allows for a convenient constructor syntax to define features from deserialized JSON dictionaries. This function is used in particular when deserializing a [DatasetInfo] that was dumped to a JSON object. This acts as an analogue to [Features.from_arrow_schema] and handles the recursive field-by-field instantiation, but doesn’t require any mapping to/from pyarrow, except for the fact that it takes advantage of the mapping of pyarrow primitive dtypes that [Value] automatically performs.

Example:

>>> Features.from_dict({'_type': {'dtype': 'string', 'id': None, '_type': 'Value'}})
{'_type': Value(dtype='string', id=None)}

reorder_fields_as

< >

( other: Features )

Parameters

  • other ([Features]) — The other [Features] to align with.

Reorder Features fields to match the field order of other [Features].

The order of the fields is important since it matters for the underlying arrow data. Re-ordering the fields allows to make the underlying arrow data type match.

Example:

>>> from datasets import Features, Sequence, Value
>>> # let's say we have two features with a different order of nested fields (for a and b for example)
>>> f1 = Features({"root": Sequence({"a": Value("string"), "b": Value("string")})})
>>> f2 = Features({"root": {"b": Sequence(Value("string")), "a": Sequence(Value("string"))}})
>>> assert f1.type != f2.type
>>> # re-ordering keeps the base structure (here Sequence is defined at the root level), but makes the fields order match
>>> f1.reorder_fields_as(f2)
{'root': Sequence(feature={'b': Value(dtype='string', id=None), 'a': Value(dtype='string', id=None)}, length=-1, id=None)}
>>> assert f1.reorder_fields_as(f2).type == f2.type

class datasets.Value

< >

( dtype: str id: Optional = None )

The Value dtypes are as follows:

  • null
  • bool
  • int8
  • int16
  • int32
  • int64
  • uint8
  • uint16
  • uint32
  • uint64
  • float16
  • float32 (alias float)
  • float64 (alias double)
  • time32[(s|ms)]
  • time64[(us|ns)]
  • timestamp[(s|ms|us|ns)]
  • timestamp[(s|ms|us|ns), tz=(tzstring)]
  • date32
  • date64
  • duration[(s|ms|us|ns)]
  • decimal128(precision, scale)
  • decimal256(precision, scale)
  • binary
  • large_binary
  • string
  • large_string

Example:

>>> from datasets import Features
>>> features = Features({'stars': Value(dtype='int32')})
>>> features
{'stars': Value(dtype='int32', id=None)}

class datasets.ClassLabel

< >

( num_classes: dataclasses.InitVar[typing.Optional[int]] = None names: List = None names_file: dataclasses.InitVar[typing.Optional[str]] = None id: Optional = None )

Parameters

  • num_classes (int, optional) — Number of classes. All labels must be < num_classes.
  • names (list of str, optional) — String names for the integer classes. The order in which the names are provided is kept.
  • names_file (str, optional) — Path to a file with names for the integer classes, one per line.

Feature type for integer class labels.

There are 3 ways to define a ClassLabel, which correspond to the 3 arguments:

  • num_classes: Create 0 to (num_classes-1) labels.
  • names: List of label strings.
  • names_file: File containing the list of labels.

Under the hood the labels are stored as integers. You can use negative integers to represent unknown/missing labels.

Example:

>>> from datasets import Features
>>> features = Features({'label': ClassLabel(num_classes=3, names=['bad', 'ok', 'good'])})
>>> features
{'label': ClassLabel(num_classes=3, names=['bad', 'ok', 'good'], id=None)}

cast_storage

< >

( storage: Union ) → pa.Int64Array

Parameters

  • storage (Union[pa.StringArray, pa.IntegerArray]) — PyArrow array to cast.

Returns

pa.Int64Array

Array in the ClassLabel arrow storage type.

Cast an Arrow array to the ClassLabel arrow storage type. The Arrow types that can be converted to the ClassLabel pyarrow storage type are:

  • pa.string()
  • pa.int()

int2str

< >

( values: Union )

Conversion integer => class name string.

Regarding unknown/missing labels: passing negative integers raises ValueError.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train")
>>> ds.features["label"].int2str(0)
'neg'

str2int

< >

( values: Union )

Conversion class name string => integer.

Example:

>>> from datasets import load_dataset
>>> ds = load_dataset("rotten_tomatoes", split="train")
>>> ds.features["label"].str2int('neg')
0

class datasets.LargeList

< >

( dtype: Any id: Optional = None )

Feature type for large list data composed of child feature data type.

It is backed by pyarrow.LargeListType, which is like pyarrow.ListType but with 64-bit rather than 32-bit offsets.

class datasets.Sequence

< >

( feature: Any length: int = -1 id: Optional = None )

Parameters

  • length (int) — Length of the sequence.

Construct a list of feature from a single type or a dict of types. Mostly here for compatiblity with tfds.

Example:

>>> from datasets import Features, Sequence, Value, ClassLabel
>>> features = Features({'post': Sequence(feature={'text': Value(dtype='string'), 'upvotes': Value(dtype='int32'), 'label': ClassLabel(num_classes=2, names=['hot', 'cold'])})})
>>> features
{'post': Sequence(feature={'text': Value(dtype='string', id=None), 'upvotes': Value(dtype='int32', id=None), 'label': ClassLabel(num_classes=2, names=['hot', 'cold'], id=None)}, length=-1, id=None)}

class datasets.Translation

< >

( languages: List id: Optional = None )

Parameters

  • languages (dict) — A dictionary for each example mapping string language codes to string translations.

FeatureConnector for translations with fixed languages per example. Here for compatiblity with tfds.

Example:

>>> # At construction time:
>>> datasets.features.Translation(languages=['en', 'fr', 'de'])
>>> # During data generation:
>>> yield {
...         'en': 'the cat',
...         'fr': 'le chat',
...         'de': 'die katze'
... }

flatten

< >

( )

Flatten the Translation feature into a dictionary.

class datasets.TranslationVariableLanguages

< >

( languages: Optional = None num_languages: Optional = None id: Optional = None ) →

  • language or translation (variable-length 1D tf.Tensor of tf.string)

Parameters

  • languages (dict) — A dictionary for each example mapping string language codes to one or more string translations. The languages present may vary from example to example.

Returns

  • language or translation (variable-length 1D tf.Tensor of tf.string)

Language codes sorted in ascending order or plain text translations, sorted to align with language codes.

FeatureConnector for translations with variable languages per example. Here for compatiblity with tfds.

Example:

>>> # At construction time:
>>> datasets.features.TranslationVariableLanguages(languages=['en', 'fr', 'de'])
>>> # During data generation:
>>> yield {
...         'en': 'the cat',
...         'fr': ['le chat', 'la chatte,']
...         'de': 'die katze'
... }
>>> # Tensor returned :
>>> {
...         'language': ['en', 'de', 'fr', 'fr'],
...         'translation': ['the cat', 'die katze', 'la chatte', 'le chat'],
... }

flatten

< >

( )

Flatten the TranslationVariableLanguages feature into a dictionary.

class datasets.Array2D

< >

( shape: tuple dtype: str id: Optional = None )

Parameters

  • shape (tuple) — The size of each dimension.
  • dtype (str) — The value of the data type.

Create a two-dimensional array.

Example:

>>> from datasets import Features
>>> features = Features({'x': Array2D(shape=(1, 3), dtype='int32')})

class datasets.Array3D

< >

( shape: tuple dtype: str id: Optional = None )

Parameters

  • shape (tuple) — The size of each dimension.
  • dtype (str) — The value of the data type.

Create a three-dimensional array.

Example:

>>> from datasets import Features
>>> features = Features({'x': Array3D(shape=(1, 2, 3), dtype='int32')})

class datasets.Array4D

< >

( shape: tuple dtype: str id: Optional = None )

Parameters

  • shape (tuple) — The size of each dimension.
  • dtype (str) — The value of the data type.

Create a four-dimensional array.

Example:

>>> from datasets import Features
>>> features = Features({'x': Array4D(shape=(1, 2, 2, 3), dtype='int32')})

class datasets.Array5D

< >

( shape: tuple dtype: str id: Optional = None )

Parameters

  • shape (tuple) — The size of each dimension.
  • dtype (str) — The value of the data type.

Create a five-dimensional array.

Example:

>>> from datasets import Features
>>> features = Features({'x': Array5D(shape=(1, 2, 2, 3, 3), dtype='int32')})

class datasets.Audio

< >

( sampling_rate: Optional = None mono: bool = True decode: bool = True id: Optional = None )

Parameters

  • sampling_rate (int, optional) — Target sampling rate. If None, the native sampling rate is used.
  • mono (bool, defaults to True) — Whether to convert the audio signal to mono by averaging samples across channels.
  • decode (bool, defaults to True) — Whether to decode the audio data. If False, returns the underlying dictionary in the format {"path": audio_path, "bytes": audio_bytes}.

Audio Feature to extract audio data from an audio file.

Input: The Audio feature accepts as input:

  • A str: Absolute path to the audio file (i.e. random access is allowed).

  • A dict with the keys:

    • path: String with relative path of the audio file to the archive file.
    • bytes: Bytes content of the audio file.

    This is useful for archived files with sequential access.

  • A dict with the keys:

    • path: String with relative path of the audio file to the archive file.
    • array: Array containing the audio sample
    • sampling_rate: Integer corresponding to the sampling rate of the audio sample.

    This is useful for archived files with sequential access.

Example:

>>> from datasets import load_dataset, Audio
>>> ds = load_dataset("PolyAI/minds14", name="en-US", split="train")
>>> ds = ds.cast_column("audio", Audio(sampling_rate=16000))
>>> ds[0]["audio"]
{'array': array([ 2.3443763e-05,  2.1729663e-04,  2.2145823e-04, ...,
     3.8356509e-05, -7.3497440e-06, -2.1754686e-05], dtype=float32),
 'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~JOINT_ACCOUNT/602ba55abb1e6d0fbce92065.wav',
 'sampling_rate': 16000}

cast_storage

< >

( storage: Union ) → pa.StructArray

Parameters

  • storage (Union[pa.StringArray, pa.StructArray]) — PyArrow array to cast.

Returns

pa.StructArray

Array in the Audio arrow storage type, that is pa.struct({"bytes": pa.binary(), "path": pa.string()})

Cast an Arrow array to the Audio arrow storage type. The Arrow types that can be converted to the Audio pyarrow storage type are:

  • pa.string() - it must contain the “path” data
  • pa.binary() - it must contain the audio bytes
  • pa.struct({"bytes": pa.binary()})
  • pa.struct({"path": pa.string()})
  • pa.struct({"bytes": pa.binary(), "path": pa.string()}) - order doesn’t matter

decode_example

< >

( value: dict token_per_repo_id: Optional = None ) → dict

Parameters

  • value (dict) — A dictionary with keys:

    • path: String with relative audio file path.
    • bytes: Bytes of the audio file.
  • token_per_repo_id (dict, optional) — To access and decode audio files from private repositories on the Hub, you can pass a dictionary repo_id (str) -> token (bool or str)

Returns

dict

Decode example audio file into audio data.

embed_storage

< >

( storage: StructArray ) → pa.StructArray

Parameters

  • storage (pa.StructArray) — PyArrow array to embed.

Returns

pa.StructArray

Array in the Audio arrow storage type, that is pa.struct({"bytes": pa.binary(), "path": pa.string()}).

Embed audio files into the Arrow array.

encode_example

< >

( value: Union ) → dict

Parameters

  • value (str or dict) — Data passed as input to Audio feature.

Returns

dict

Encode example into a format for Arrow.

flatten

< >

( )

If in the decodable state, raise an error, otherwise flatten the feature into a dictionary.

class datasets.Image

< >

( mode: Optional = None decode: bool = True id: Optional = None )

Parameters

  • mode (str, optional) — The mode to convert the image to. If None, the native mode of the image is used.
  • decode (bool, defaults to True) — Whether to decode the image data. If False, returns the underlying dictionary in the format {"path": image_path, "bytes": image_bytes}.

Image Feature to read image data from an image file.

Input: The Image feature accepts as input:

  • A str: Absolute path to the image file (i.e. random access is allowed).

  • A dict with the keys:

    • path: String with relative path of the image file to the archive file.
    • bytes: Bytes of the image file.

    This is useful for archived files with sequential access.

  • An np.ndarray: NumPy array representing an image.

  • A PIL.Image.Image: PIL image object.

Examples:

>>> from datasets import load_dataset, Image
>>> ds = load_dataset("beans", split="train")
>>> ds.features["image"]
Image(decode=True, id=None)
>>> ds[0]["image"]
<PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=500x500 at 0x15E52E7F0>
>>> ds = ds.cast_column('image', Image(decode=False))
{'bytes': None,
 'path': '/root/.cache/huggingface/datasets/downloads/extracted/b0a21163f78769a2cf11f58dfc767fb458fc7cea5c05dccc0144a2c0f0bc1292/train/healthy/healthy_train.85.jpg'}

cast_storage

< >

( storage: Union ) → pa.StructArray

Parameters

  • storage (Union[pa.StringArray, pa.StructArray, pa.ListArray]) — PyArrow array to cast.

Returns

pa.StructArray

Array in the Image arrow storage type, that is pa.struct({"bytes": pa.binary(), "path": pa.string()}).

Cast an Arrow array to the Image arrow storage type. The Arrow types that can be converted to the Image pyarrow storage type are:

  • pa.string() - it must contain the “path” data
  • pa.binary() - it must contain the image bytes
  • pa.struct({"bytes": pa.binary()})
  • pa.struct({"path": pa.string()})
  • pa.struct({"bytes": pa.binary(), "path": pa.string()}) - order doesn’t matter
  • pa.list(*) - it must contain the image array data

decode_example

< >

( value: dict token_per_repo_id = None )

Parameters

  • value (str or dict) — A string with the absolute image file path, a dictionary with keys:

    • path: String with absolute or relative image file path.
    • bytes: The bytes of the image file.
  • token_per_repo_id (dict, optional) — To access and decode image files from private repositories on the Hub, you can pass a dictionary repo_id (str) -> token (bool or str).

Decode example image file into image data.

embed_storage

< >

( storage: StructArray ) → pa.StructArray

Parameters

  • storage (pa.StructArray) — PyArrow array to embed.

Returns

pa.StructArray

Array in the Image arrow storage type, that is pa.struct({"bytes": pa.binary(), "path": pa.string()}).

Embed image files into the Arrow array.

encode_example

< >

( value: Union )

Parameters

  • value (str, np.ndarray, PIL.Image.Image or dict) — Data passed as input to Image feature.

Encode example into a format for Arrow.

flatten

< >

( )

If in the decodable state, return the feature itself, otherwise flatten the feature into a dictionary.

MetricInfo

class datasets.MetricInfo

< >

( description: str citation: str features: Features inputs_description: str = <factory> homepage: str = <factory> license: str = <factory> codebase_urls: List = <factory> reference_urls: List = <factory> streamable: bool = False format: Optional = None metric_name: Optional = None config_name: Optional = None experiment_id: Optional = None )

Information about a metric.

MetricInfo documents a metric, including its name, version, and features. See the constructor arguments and properties for a full list.

Note: Not all fields are known on construction and may be updated later.

from_directory

< >

( metric_info_dir )

Create MetricInfo from the JSON file in metric_info_dir.

Example:

>>> from datasets import MetricInfo
>>> metric_info = MetricInfo.from_directory("/path/to/directory/")

write_to_directory

< >

( metric_info_dir pretty_print = False )

Write MetricInfo as JSON to metric_info_dir. Also save the license separately in LICENCE. If pretty_print is True, the JSON will be pretty-printed with the indent level of 4.

Example:

>>> from datasets import load_metric
>>> metric = load_metric("accuracy")
>>> metric.info.write_to_directory("/path/to/directory/")

Metric

The base class Metric implements a Metric backed by one or several Dataset.

class datasets.Metric

< >

( config_name: Optional = None keep_in_memory: bool = False cache_dir: Optional = None num_process: int = 1 process_id: int = 0 seed: Optional = None experiment_id: Optional = None max_concurrent_cache_files: int = 10000 timeout: Union = 100 **kwargs )

Parameters

  • config_name (str) — This is used to define a hash specific to a metrics computation script and prevents the metric’s data to be overridden when the metric loading script is modified.
  • keep_in_memory (bool) — keep all predictions and references in memory. Not possible in distributed settings.
  • cache_dir (str) — Path to a directory in which temporary prediction/references data will be stored. The data directory should be located on a shared file-system in distributed setups.
  • num_process (int) — specify the total number of nodes in a distributed settings. This is useful to compute metrics in distributed setups (in particular non-additive metrics like F1).
  • process_id (int) — specify the id of the current process in a distributed setup (between 0 and num_process-1) This is useful to compute metrics in distributed setups (in particular non-additive metrics like F1).
  • seed (int, optional) — If specified, this will temporarily set numpy’s random seed when datasets.Metric.compute() is run.
  • experiment_id (str) — A specific experiment id. This is used if several distributed evaluations share the same file system. This is useful to compute metrics in distributed setups (in particular non-additive metrics like F1).
  • max_concurrent_cache_files (int) — Max number of concurrent metrics cache files (default 10000).
  • timeout (Union[int, float]) — Timeout in second for distributed setting synchronization.

A Metric is the base class and common API for all metrics.

Deprecated in 2.5.0

Use the new library 🤗 Evaluate instead: https://huggingface.co/docs/evaluate

add

< >

( prediction = None reference = None **kwargs )

Parameters

  • prediction (list/array/tensor, optional) — Predictions.
  • reference (list/array/tensor, optional) — References.

Add one prediction and reference for the metric’s stack.

Example:

>>> from datasets import load_metric
>>> metric = load_metric("accuracy")
>>> metric.add(predictions=model_predictions, references=labels)

add_batch

< >

( predictions = None references = None **kwargs )

Parameters

  • predictions (list/array/tensor, optional) — Predictions.
  • references (list/array/tensor, optional) — References.

Add a batch of predictions and references for the metric’s stack.

Example:

>>> from datasets import load_metric
>>> metric = load_metric("accuracy")
>>> metric.add_batch(predictions=model_prediction, references=labels)

compute

< >

( predictions = None references = None **kwargs )

Parameters

  • predictions (list/array/tensor, optional) — Predictions.
  • references (list/array/tensor, optional) — References.
  • **kwargs (optional) — Keyword arguments that will be forwarded to the metrics _compute method (see details in the docstring).

Compute the metrics.

Usage of positional arguments is not allowed to prevent mistakes.

Example:

>>> from datasets import load_metric
>>> metric = load_metric("accuracy")
>>> accuracy = metric.compute(predictions=model_prediction, references=labels)

download_and_prepare

< >

( download_config: Optional = None dl_manager: Optional = None )

Parameters

  • download_config (DownloadConfig, optional) — Specific download configuration parameters.
  • dl_manager (DownloadManager, optional) — Specific download manager to use.

Downloads and prepares dataset for reading.

Filesystems

class datasets.filesystems.S3FileSystem

< >

( *args **kwargs )

Parameters

  • anon (bool, default to False) — Whether to use anonymous connection (public buckets only). If False, uses the key/secret given, or boto’s credential resolver (client_kwargs, environment, variables, config files, EC2 IAM server, in that order).
  • key (str) — If not anonymous, use this access key ID, if specified.
  • secret (str) — If not anonymous, use this secret access key, if specified.
  • token (str) — If not anonymous, use this security token, if specified.
  • use_ssl (bool, defaults to True) — Whether to use SSL in connections to S3; may be faster without, but insecure. If use_ssl is also set in client_kwargs, the value set in client_kwargs will take priority.
  • s3_additional_kwargs (dict) — Parameters that are used when calling S3 API methods. Typically used for things like ServerSideEncryption.
  • client_kwargs (dict) — Parameters for the botocore client.
  • requester_pays (bool, defaults to False) — Whether RequesterPays buckets are supported.
  • default_block_size (int) — If given, the default block size value used for open(), if no specific value is given at all time. The built-in default is 5MB.
  • default_fill_cache (bool, defaults to True) — Whether to use cache filling with open by default. Refer to S3File.open.
  • default_cache_type (str, defaults to bytes) — If given, the default cache_type value used for open(). Set to none if no caching is desired. See fsspec’s documentation for other available cache_type values.
  • version_aware (bool, defaults to False) — Whether to support bucket versioning. If enable this will require the user to have the necessary IAM permissions for dealing with versioned objects.
  • cache_regions (bool, defaults to False) — Whether to cache bucket regions. Whenever a new bucket is used, it will first find out which region it belongs to and then use the client for that region.
  • asynchronous (bool, defaults to False) — Whether this instance is to be used from inside coroutines.
  • config_kwargs (dict) — Parameters passed to botocore.client.Config. **kwargs — Other parameters for core session.
  • session (aiobotocore.session.AioSession) — Session to be used for all connections. This session will be used inplace of creating a new session inside S3FileSystem. For example: aiobotocore.session.AioSession(profile='test_user').
  • skip_instance_cache (bool) — Control reuse of instances. Passed on to fsspec.
  • use_listings_cache (bool) — Control reuse of directory listings. Passed on to fsspec.
  • listings_expiry_time (int or float) — Control reuse of directory listings. Passed on to fsspec.
  • max_paths (int) — Control reuse of directory listings. Passed on to fsspec.

datasets.filesystems.S3FileSystem is a subclass of s3fs.S3FileSystem.

Users can use this class to access S3 as if it were a file system. It exposes a filesystem-like API (ls, cp, open, etc.) on top of S3 storage. Provide credentials either explicitly (key=, secret=) or with boto’s credential methods. See botocore documentation for more information. If no credentials are available, use anon=True.

Examples:

Listing files from public S3 bucket.

>>> import datasets
>>> s3 = datasets.filesystems.S3FileSystem(anon=True)
>>> s3.ls('public-datasets/imdb/train')
['dataset_info.json.json','dataset.arrow','state.json']

Listing files from private S3 bucket using aws_access_key_id and aws_secret_access_key.

>>> import datasets
>>> s3 = datasets.filesystems.S3FileSystem(key=aws_access_key_id, secret=aws_secret_access_key)
>>> s3.ls('my-private-datasets/imdb/train')
['dataset_info.json.json','dataset.arrow','state.json']

Using S3Filesystem with botocore.session.Session and custom aws_profile.

>>> import botocore
>>> from datasets.filesystems import S3Filesystem

>>> s3_session = botocore.session.Session(profile_name='my_profile_name')
>>> s3 = S3FileSystem(session=s3_session)

Loading dataset from S3 using S3Filesystem and load_from_disk().

>>> from datasets import load_from_disk
>>> from datasets.filesystems import S3Filesystem

>>> s3 = S3FileSystem(key=aws_access_key_id, secret=aws_secret_access_key)
>>> dataset = load_from_disk('s3://my-private-datasets/imdb/train', storage_options=s3.storage_options)
>>> print(len(dataset))
25000

Saving dataset to S3 using S3Filesystem and Dataset.save_to_disk().

>>> from datasets import load_dataset
>>> from datasets.filesystems import S3Filesystem

>>> dataset = load_dataset("imdb")
>>> s3 = S3FileSystem(key=aws_access_key_id, secret=aws_secret_access_key)
>>> dataset.save_to_disk('s3://my-private-datasets/imdb/train', storage_options=s3.storage_options)

datasets.filesystems.extract_path_from_uri

< >

( dataset_path: str )

Parameters

  • dataset_path (str) — Path (e.g. dataset/train) or remote uri (e.g. s3://my-bucket/dataset/train) of the dataset directory.

Preprocesses dataset_path and removes remote filesystem (e.g. removing s3://).

datasets.filesystems.is_remote_filesystem

< >

( fs: AbstractFileSystem )

Parameters

Checks if fs is a remote filesystem.

Fingerprint

class datasets.fingerprint.Hasher

< >

( )

Hasher that accepts python objects as inputs.

< > Update on GitHub

←All about metrics Builder classes→