Skip to content

deepbox/datasets

Built-in & External Datasets

Bundled reference datasets such as Iris and Wine, CSV, text and image loaders that use the network, and Kaggle helpers.

  • Data sources

Overview

  • loadIris(), loadWine(), loadDigits(), loadBreastCancer(), loadDiabetes() and loadLinnerud() are bundled and need no network. Each returns { data, target, featureNames, targetNames, description }. data follows the global default dtype (float32), so pass { dtype: "float64" } to get the exact reference values, as scikit-learn returns them. Classification targets are int32. The page also lists smaller teaching datasets such as loadHousingMini and loadStudentPerformance.
  • Remote helpers need network access and throw when the download fails: fetchCSVDataset, fetchMNIST, fetchCIFAR10, fetch20Newsgroups, fetchIMDB and the Kaggle functions, which read credentials with readKaggleCredentials. fetch20Newsgroups and fetchIMDB download the official archives (the default timeout is 10 minutes). Pass allowSyntheticFallback: true to get placeholder text instead of an error.
  • fetchCSVDataset({ url, targetColumn }) and parseCSV(text) accept numeric CSV only. Every data cell must be a finite number, so text columns and empty cells are rejected. The target is the last column unless you set targetColumn. fetchCSVDataset stops at maxBytes (100 MB) and timeout (30 seconds) by default.

Types

12

ImageDataset

type

Result of an image dataset fetch.

  • readonly data: Tensor

    Flattened image tensor of shape (nSamples, nPixels), dtype float32, pixel values scaled to [0, 1].

  • readonly target: Tensor

    Class index tensor of shape (nSamples,), dtype int32.

  • readonly nClasses: number

    Number of classes

  • readonly classNames: string[]

    Class label names

  • readonly imageDims: readonly [ number, number, number ]

    Image dimensions [height, width, channels]

  • readonly description: string

    Dataset description

ImageFetchOptions

type

Options shared by fetchMNIST and fetchCIFAR10.

  • readonly baseUrl?: string

    Base URL override for custom mirrors.

  • readonly maxSamples?: number

    Maximum number of samples to load (positive integer).

  • readonly split?: "train" | "test"

    Which split to load: 'train' or 'test'.

  • readonly signal?: AbortSignal

    Abort signal forwarded to fetch.

KaggleCredentials

type

Kaggle API credentials.

  • readonly username: string

    Kaggle username.

  • readonly key: string

    Kaggle API key.

KaggleDatasetInfo

type

Metadata about a Kaggle dataset.

  • readonly id: string
  • readonly title: string
  • readonly subtitle: string
  • readonly totalBytes: number
  • readonly fileCount: number
  • readonly lastUpdated: string
  • readonly downloadCount: number
  • readonly usabilityRating: number

KaggleDownloadResult

type

Result of a Kaggle dataset download.

  • readonly data: Uint8Array

    Raw file data as Uint8Array (may be a ZIP archive).

  • readonly filename: string

    The filename.

  • readonly contentType: string

    Content type from the response.

  • readonly totalBytes: number

    Number of bytes in data (at most maxBytes when that option is set).

KaggleFetchOptions

type

Options for fetching a Kaggle dataset.

  • readonly credentials?: KaggleCredentials

    API credentials.

  • readonly file?: string

    Specific file within the dataset to download.

  • readonly maxBytes?: number

    Maximum number of bytes to keep (non-negative integer).

  • readonly apiBaseUrl?: string

    API base URL override (mirrors, proxies, tests).

  • readonly signal?: AbortSignal

    Abort signal forwarded to fetch.

Dataset

type

Standard dataset structure returned by all built-in dataset loaders.

  • data: Tensor
  • target: Tensor
  • featureNames: string[]
  • targetNames?: string[]
  • description: string
  • images?: Tensor

DatasetLoadOptions

type

Options accepted by the loaders for the bundled reference datasets (loadIris, loadWine, loadDigits, loadBreastCancer, loadDiabetes and loadLinnerud).

  • readonly dtype?: "float32" | "float64"

    Floating-point dtype of data (and of target for regression datasets).

FetchCSVDatasetOptions

type

Options for fetchCSVDataset.

  • readonly url: string

    URL of the CSV file to fetch.

  • readonly targetColumn?: number

    0-based column index to use as the target variable.

  • readonly header?: boolean

    Whether the first row is a header.

  • readonly separator?: string

    Single-character column separator.

  • readonly maxBytes?: number

    Maximum bytes to download (streaming limit).

  • readonly timeout?: number

    Fetch timeout in milliseconds.

  • readonly abortSignal?: AbortSignal

    AbortSignal for cancellation.

RemoteDataset

type

Result of a remote dataset fetch.

  • readonly data: Tensor
  • readonly target: Tensor
  • readonly featureNames: string[]
  • readonly targetName: string
  • readonly description: string

TextDataset

type

Result of a text dataset fetch.

  • readonly texts: string[]

    Array of text documents

  • readonly target: Tensor

    Integer (int32) label tensor of shape (nSamples,)

  • readonly nClasses: number

    Number of classes

  • readonly classNames: string[]

    Class label names

  • readonly description: string

    Dataset description

  • readonly isSynthetic: boolean

    Whether this is synthetic/fallback data (true when remote fetch failed)

TextFetchOptions

type

Options shared by fetch20Newsgroups and fetchIMDB.

  • readonly subset?: "train" | "test" | "all"

    Which subset: 'train', 'test', or 'all' (default: 'all')

  • readonly maxSamples?: number

    Maximum number of samples to load (a positive integer).

  • readonly baseUrl?: string

    Base URL of a JSON mirror; the file name is appended to it.

  • readonly archiveUrl?: string

    URL of a gzip-compressed tar archive with the official layout (for example a local mirror of the default archive).

  • readonly allowSyntheticFallback?: boolean

    When true, return synthetic placeholder data on fetch failure.

  • readonly timeout?: number

    Fetch timeout in milliseconds (0 disables the timeout).

Functions

36

fetchCIFAR10

function fetchCIFAR10(options?: ImageFetchOptions): Promise<ImageDataset>;

Fetch the CIFAR-10 image classification dataset.

Downloads the binary version from a public mirror and returns images (32×32×3 RGB) with labels across 10 categories. Each row of data keeps the CIFAR channel-major layout (1024 red values, then 1024 green, then 1024 blue).

Parameters

optionsoptional
ImageFetchOptions
Configuration options

Returns

Promise<ImageDataset>ImageDataset with CIFAR-10 data (float32 pixels in [0, 1], int32 labels)

Throws

Example
import { fetchCIFAR10 } from 'deepbox/datasets'; const cifar = await fetchCIFAR10();console.log(cifar.data.shape);   // [50000, 3072]console.log(cifar.target.shape); // [50000]

fetchMNIST

function fetchMNIST(options?: ImageFetchOptions): Promise<ImageDataset>;

Fetch the MNIST handwritten digits dataset.

Downloads the dataset from a public mirror and returns 60,000 training images (28×28 grayscale) with labels 0-9.

Parameters

optionsoptional
ImageFetchOptions
Configuration options

Returns

Promise<ImageDataset>ImageDataset with MNIST data (float32 pixels in [0, 1], int32 labels)

Throws

Example
import { fetchMNIST } from 'deepbox/datasets'; const mnist = await fetchMNIST();console.log(mnist.data.shape);   // [60000, 784]console.log(mnist.target.shape); // [60000]

fetchKaggleDataset

function fetchKaggleDataset(datasetId: string, options?: KaggleFetchOptions): Promise<KaggleDownloadResult>;

Download a Kaggle dataset or specific file.

Without file the result is the whole dataset as a ZIP archive.

Parameters

datasetId
string
Dataset identifier in "owner/dataset-name" format
optionsoptional
KaggleFetchOptions
Fetch options

Returns

Promise<KaggleDownloadResult>Download result with raw data

Throws

Example
import { fetchKaggleDataset } from 'deepbox/datasets'; const result = await fetchKaggleDataset('zillow/zecon', {  credentials: { username: 'myuser', key: 'mykey' },  file: 'Zip_zhvi_uc_sfrcondo_tier_0.33_0.67_sm_sa_month.csv'});console.log(result.totalBytes);

fetchKaggleDatasetInfo

function fetchKaggleDatasetInfo(datasetId: string, options?: KaggleFetchOptions): Promise<KaggleDatasetInfo>;

Fetch metadata for a Kaggle dataset.

Parameters

datasetId
string
Dataset identifier in "owner/dataset-name" format
optionsoptional
KaggleFetchOptions
Fetch options (credentials)

Returns

Promise<KaggleDatasetInfo>Dataset metadata

Throws

Example
import { fetchKaggleDatasetInfo } from 'deepbox/datasets'; const info = await fetchKaggleDatasetInfo('zillow/zecon', {  credentials: { username: 'myuser', key: 'mykey' }});console.log(info.title, info.totalBytes);

listKaggleFiles

function listKaggleFiles(datasetId: string, options?: KaggleFetchOptions): Promise<{ name: string; totalBytes: number; }[]>;

List files in a Kaggle dataset.

Parameters

datasetId
string
Dataset identifier in "owner/dataset-name" format
optionsoptional
KaggleFetchOptions
Fetch options

Returns

Promise<{ name: string; totalBytes: number; }[]>Array of file names and sizes

Throws

readKaggleCredentials

function readKaggleCredentials(): Promise<KaggleCredentials | undefined>;

Read Kaggle credentials.

Looks, in order, at the KAGGLE_USERNAME / KAGGLE_KEY environment variables, then at kaggle.json in $KAGGLE_CONFIG_DIR (if set) or in ~/.kaggle. File access needs Node.js; in browsers the result is undefined.

Returns

Promise<KaggleCredentials | undefined>Credentials if found, undefined otherwise

searchKaggleDatasets

function searchKaggleDatasets(query: string, options?: KaggleFetchOptions): Promise<KaggleDatasetInfo[]>;

Search Kaggle datasets.

Parameters

query
string
Search query string
optionsoptional
KaggleFetchOptions
Fetch options (credentials)

Returns

Promise<KaggleDatasetInfo[]>Array of dataset metadata

Throws

loadBreastCancer

function loadBreastCancer(options?: DatasetLoadOptions): Dataset;

Load the Breast Cancer Wisconsin (Diagnostic) dataset (UCI).

The real, canonical measurements: 569 samples, 30 features, 2 classes (212 malignant, 357 benign). Deterministic: the committed reference values are always returned unchanged.

Parameters

optionsoptional
DatasetLoadOptions
Optional DatasetLoadOptions (e.g. { dtype: "float64" }).

Returns

DatasetA Dataset with data shape [569, 30] and target shape [569] (int32).

Throws

  • InvalidParameterErrorIf options.dtype is not "float32" or "float64".

loadConcentricRings

function loadConcentricRings(): Dataset;

Load the synthetic Concentric Rings classification dataset.

150 samples, 2D, 3 concentric circle classes. Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [150, 2] and target shape [150] (int32).

loadCropYield

function loadCropYield(): Dataset;

Load the synthetic Crop Yield regression dataset.

200 samples, 3 features (rainfall, fertilizer, temperature), target: yield (tons/ha). Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [200, 3] and target shape [200].

loadCustomerSegments

function loadCustomerSegments(): Dataset;

Load the synthetic Customer Segments clustering dataset.

200 samples, 3 features (age, income, spending score), 4 natural clusters. Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [200, 3] and target shape [200] (int32).

loadDiabetes

function loadDiabetes(options?: DatasetLoadOptions): Dataset;

Load the Diabetes regression dataset (Efron et al., 2004).

The real, canonical data: 442 samples, 10 baseline features (age, sex, BMI, blood pressure, and six blood-serum measurements; each column mean-centered and scaled to unit L2 norm), with a continuous disease-progression target (real values, genuinely correlated with the features). Deterministic: the committed reference values are always returned unchanged.

Parameters

optionsoptional
DatasetLoadOptions
Optional DatasetLoadOptions (e.g. { dtype: "float64" }).

Returns

DatasetA Dataset with data shape [442, 10] and target shape [442].

Throws

  • InvalidParameterErrorIf options.dtype is not "float32" or "float64".

loadDigits

function loadDigits(options?: DatasetLoadOptions): Dataset;

Load the Optical Recognition of Handwritten Digits dataset (UCI).

The real, canonical data: 1797 samples, 64 features (8×8 pixels, integer intensities in [0, 16]), 10 classes (digits 0–9). Deterministic: the committed reference values are always returned unchanged.

Parameters

optionsoptional
DatasetLoadOptions
Optional DatasetLoadOptions (e.g. { dtype: "float64" }).

Returns

DatasetA Dataset with data shape [1797, 64], target shape [1797] (int32), and images shape [1797, 8, 8] (a view of data).

Throws

  • InvalidParameterErrorIf options.dtype is not "float32" or "float64".

loadEnergyEfficiency

function loadEnergyEfficiency(): Dataset;

Load the synthetic Energy Efficiency regression dataset.

200 samples, 3 features (insulation, window area, orientation), target: energy usage (kWh). Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [200, 3] and target shape [200].

loadFitnessScores

function loadFitnessScores(): Dataset;

Load the synthetic Fitness Scores multi-output regression dataset.

100 samples, 3 exercise features, 3 fitness targets (strength, endurance, flexibility). Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [100, 3] and target shape [100, 3].

loadFlowersExtended

function loadFlowersExtended(): Dataset;

Load the synthetic Flowers Extended classification dataset.

180 samples, 6 features, 4 species. Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [180, 6] and target shape [180] (int32).

loadFruitQuality

function loadFruitQuality(): Dataset;

Load the synthetic Fruit Quality classification dataset.

150 samples, 5 features, 3 fruit classes. Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [150, 5] and target shape [150] (int32).

loadGaussianIslands

function loadGaussianIslands(): Dataset;

Load the synthetic Gaussian Islands classification dataset.

200 samples, 3D, 4 separated Gaussian clusters. Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [200, 3] and target shape [200] (int32).

loadHousingMini

function loadHousingMini(): Dataset;

Load the synthetic Housing-Mini regression dataset.

200 samples, 4 features (size, rooms, age, distance), target: price (thousands). Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [200, 4] and target shape [200].

loadIris

function loadIris(options?: DatasetLoadOptions): Dataset;

Load the Iris dataset (R. A. Fisher, 1936).

The real, canonical measurements: 150 samples, 4 features, 3 classes (50 setosa, 50 versicolor, 50 virginica). Deterministic: the committed reference values are always returned unchanged.

Parameters

optionsoptional
DatasetLoadOptions
Optional DatasetLoadOptions (e.g. { dtype: "float64" }).

Returns

DatasetA Dataset with data shape [150, 4] and target shape [150] (int32).

Throws

  • InvalidParameterErrorIf options.dtype is not "float32" or "float64".

loadLeafShapes

function loadLeafShapes(): Dataset;

Load the synthetic Leaf Shapes classification dataset.

150 samples, 8 geometric features, 5 plant species. Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [150, 8] and target shape [150] (int32).

loadLinnerud

function loadLinnerud(options?: DatasetLoadOptions): Dataset;

Load the Linnerud multi-output regression dataset (Tenenhaus, 1998).

The real, canonical data: 20 middle-aged men, 3 exercise features (chin-ups, sit-ups, jumps) and 3 physiological targets (weight, waist, pulse). Deterministic: the committed reference values are always returned unchanged.

Parameters

optionsoptional
DatasetLoadOptions
Optional DatasetLoadOptions (e.g. { dtype: "float64" }).

Returns

DatasetA Dataset with data shape [20, 3] and target shape [20, 3].

Throws

  • InvalidParameterErrorIf options.dtype is not "float32" or "float64".

loadMoonsMulti

function loadMoonsMulti(): Dataset;

Load the synthetic Moons-Multi classification dataset.

150 samples, 2D, 3 interleaving rotated moon classes. Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [150, 2] and target shape [150] (int32).

loadPerfectlySeparable

function loadPerfectlySeparable(): Dataset;

Load the synthetic Perfectly Separable classification dataset.

100 samples, 4 features, 2 linearly separable classes. Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [100, 4] and target shape [100] (int32).

loadPlantGrowth

function loadPlantGrowth(): Dataset;

Load the synthetic Plant Growth regression dataset.

200 samples, 3 features (sunlight, water, soil quality), target: height (cm). Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [200, 3] and target shape [200].

loadSeedMorphology

function loadSeedMorphology(): Dataset;

Load the synthetic Seed Morphology classification dataset.

150 samples, 4 features, 3 seed types. Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [150, 4] and target shape [150] (int32).

loadSensorStates

function loadSensorStates(): Dataset;

Load the synthetic Sensor States classification dataset.

180 samples, 6 sensor readings, 3 hidden operating modes. Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [180, 6] and target shape [180] (int32).

loadSpiralArms

function loadSpiralArms(): Dataset;

Load the synthetic Spiral Arms classification dataset.

150 samples, 2D, 3 spiral classes. Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [150, 2] and target shape [150] (int32).

loadStudentPerformance

function loadStudentPerformance(): Dataset;

Load the synthetic Student Performance classification dataset.

150 samples, 3 integer features, 3 outcome classes. Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [150, 3] and target shape [150] (int32).

loadTrafficConditions

function loadTrafficConditions(): Dataset;

Load the synthetic Traffic Conditions classification dataset.

150 samples, 3 features, 3 traffic level classes. Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [150, 3] and target shape [150] (int32).

loadWeatherOutcomes

function loadWeatherOutcomes(): Dataset;

Load the synthetic Weather Outcomes multi-output regression dataset.

150 samples, 3 features, 2 targets (rain probability, wind speed). Deterministic: always returns the same data.

Returns

DatasetA Dataset with data shape [150, 3] and target shape [150, 2].

loadWine

function loadWine(options?: DatasetLoadOptions): Dataset;

Load the Wine recognition dataset (UCI).

The real, canonical measurements: 178 samples, 13 chemical-analysis features, 3 cultivar classes (59 / 71 / 48). Deterministic: the committed reference values are always returned unchanged.

Parameters

optionsoptional
DatasetLoadOptions
Optional DatasetLoadOptions (e.g. { dtype: "float64" }).

Returns

DatasetA Dataset with data shape [178, 13] and target shape [178] (int32).

Throws

  • InvalidParameterErrorIf options.dtype is not "float32" or "float64".

fetchCSVDataset

function fetchCSVDataset(options: FetchCSVDatasetOptions): Promise<RemoteDataset>;

Fetch a CSV dataset from a remote URL and parse it into tensors.

The file must be numeric: every cell has to be a finite decimal number. See parseCSV for the accepted format.

Parameters

options
FetchCSVDatasetOptions
Fetch configuration

Returns

Promise<RemoteDataset>Parsed remote dataset

Throws

  • InvalidParameterErrorIf an option is invalid
  • DeepboxErrorIf the request fails, times out, is aborted or exceeds maxBytes
  • DataValidationErrorIf the CSV is malformed or contains non-numeric or missing values
Example
import { fetchCSVDataset } from 'deepbox/datasets'; const ds = await fetchCSVDataset({  url: 'https://example.com/iris.csv',  targetColumn: 4,  timeout: 5000,  maxBytes: 10 * 1024 * 1024, // 10 MB});console.log(ds.data.shape, ds.target.shape);

parseCSV

function parseCSV(text: string, options?: { readonly targetColumn?: number; readonly header?: boolean; readonly separator?: string; }): RemoteDataset;

Parse a CSV string into a RemoteDataset.

Details

Supports quoted fields (including embedded separators, line breaks and doubled quotes) and \n, \r\n or \r line endings. Blank lines are skipped. Every data cell must be a finite decimal number (12, -0.5, 1e-3); empty cells, NaN, Infinity, hexadecimal and text are rejected. Every row must have as many cells as the header (or as the first row when header is false).

Exported for testing and for users who already have CSV text.

Parameters

text
string
CSV text
optionsoptional
{ readonly targetColumn?: number; readonly header?: boolean; readonly separator?: string; }

Returns

RemoteDatasetThe features, the target column and the column names

Throws

  • InvalidParameterErrorIf an option is invalid or targetColumn is out of range
  • DataValidationErrorIf the text is empty, a row has the wrong number of cells, or a cell is missing or non-numeric

fetch20Newsgroups

function fetch20Newsgroups(options?: TextFetchOptions): Promise<TextDataset>;

Fetch the 20 Newsgroups text classification dataset.

Details

By default the official "bydate" archive (the one scikit-learn downloads, about 14 MB) is fetched from https://ndownloader.figshare.com/files/5975967 and unpacked: 11 314 training and 7 532 test posts in 20 newsgroups. Labels follow the alphabetical order of the newsgroup names; posts are ordered by newsgroup and then by file name, the training split before the test split for subset: 'all', and decoded as ISO-8859-1. Headers, footers and quotes are kept (scikit-learn's remove=()) and nothing is shuffled. With maxSamples the same number of posts is taken from every newsgroup.

Pass archiveUrl to read the same archive from another location. Pass baseUrl to read a JSON mirror instead: <baseUrl>/20newsgroups_<subset>.json, an object with data (array of strings), target (array of integer labels) and target_names (array of class names), of which the first maxSamples documents are kept.

Parameters

optionsoptional
TextFetchOptions
Configuration options

Returns

Promise<TextDataset>TextDataset with newsgroup texts and int32 labels

Throws

  • InvalidParameterErrorIf an option is invalid
  • DeepboxErrorIf the download fails or the response has the wrong layout, and allowSyntheticFallback is not set
Example
import { fetch20Newsgroups } from 'deepbox/datasets'; const news = await fetch20Newsgroups({ subset: 'train' });console.log(news.texts.length);   // 11314console.log(news.classNames[0]);  // 'alt.atheism'

fetchIMDB

function fetchIMDB(options?: TextFetchOptions): Promise<TextDataset>;

Fetch the IMDB movie review sentiment dataset.

Details

By default the official archive (about 80 MB) is fetched from https://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz and unpacked: 25 000 training and 25 000 test reviews, half of them negative (label 0) and half positive (label 1). The unlabeled unsup reviews are not read. Reviews are ordered train before test for subset: 'all', negative before positive, then by file name, and decoded as UTF-8 (the HTML line breaks of the original files are kept). With maxSamples the same number of reviews is taken from each class.

Pass archiveUrl to read the same archive from another location. Pass baseUrl to read a JSON mirror instead: <baseUrl>/imdb_<subset>.json, an object with data (array of review strings) and target (array of labels, 0 = negative, 1 = positive), of which the first maxSamples documents are kept.

Parameters

optionsoptional
TextFetchOptions
Configuration options

Returns

Promise<TextDataset>TextDataset with review texts and int32 labels

Throws

  • InvalidParameterErrorIf an option is invalid
  • DeepboxErrorIf the download fails or the response has the wrong layout, and allowSyntheticFallback is not set
Example
import { fetchIMDB } from 'deepbox/datasets'; const imdb = await fetchIMDB({ subset: 'train', maxSamples: 1000 });console.log(imdb.texts.length);   // 1000console.log(imdb.classNames);     // ['negative', 'positive']
datasets-builtin.ts
import { loadDigits, loadIris, parseCSV } from "deepbox/datasets"; const iris = loadIris({ dtype: "float64" });console.log(iris.data.shape, iris.target.shape, iris.target.dtype);console.log(iris.featureNames, iris.targetNames);console.log(loadDigits().data.shape); // parseCSV reads numeric CSV text that you already have.const parsed = parseCSV("x1,x2,label\n1,2,0\n3,4,1");console.log(parsed.data.shape, parsed.target.toString()); // Remote loaders need network access:// const remote = await fetchCSVDataset({ url: "https://example.com/data.csv" });// const news = await fetch20Newsgroups({ subset: "train" });

Found a mistake? Open an issue.