Create a dataset

Instructions for uploading a dataset to Labelbox via the application.

Import specifications: image | video | text | conversational text | geospatial | audio | documents | HTML | DICOM

Python tutorials: datasets | data rows | data row metadata

Data warehouse : Databricks (no code)| Databricks connector| Google BigQuery connector| Snowflake connector


In Labelbox, a data row represents an asset and all of its relevant information. A dataset is a collection of data rows imported to Labelbox.

Key definitions

TermDefinition
Data rowContains all of the following information for a single asset:
- URL to your cloud-hosted file
- Metadata
- Media attributes (e.g., data type, size, etc.)
- Attachments (files that provide context for your labelers)
DatasetA set of data rows from a single domain or source
AssetA single cloud-hosted file to be labeled (e.g., an image, a video, or a text file).
AttachmentSupplementary information you can attach to an asset that provides contextual information used as an aid during labeling. Learn more about attachments and image layers.
Global keyA customer-specified ID for each data row asset. It is an optional field, but it is a good practice to use global keys to map your external database/file path to your Labelbox assets for easy retrieval.

Global keys are uniquely enforced at the Catalog (organization) level, so it helps prevent duplicate data upload. This is the preferred ID to use to identify all your assets.

Supported data types

NameKindsImport specs
ImagesPNG, JPEG, BMPImage import format
VideoMP4Video import format
TextTXT (UTF-8)Text import format
Conversational textTXT (UTF-8)Conversational text import format
Geospatial imageryTile Map ServerGeospatial import format
Simple tiledTile Map ServerSimple tiled import format
AudioMP3, WAV, M4AAudio import format
DocumentsPDFDocuments import format
HTMLHTMLHTML import format
DICOMDCMDICOM import format

Create a dataset via the app UI

Although uploading a dataset via the Python SDK is recommended, you can still upload datasets via to Labelbox via the app UI.

Only images, PDFs, audio files, raw text and video files and video can currently be uploaded direclty from local storage.

  1. Sign in to Labelbox

  2. From the main menu, select Catalog and then use New dataset to create a new dataset

  3. Select files via Choose files to upload

A number of limits may affect how you choose to import your data. To learn more see Limits.

Best practices

It is best to put data from a single domain or source into a single dataset. Organizing your data this way will make it easier to set up your labeling workflows. For example, it would be easiest to organize a set of images coming from a particular type of medical device into a single dataset. You can then use metadata to better organize and filter the Data Rows within that dataset.

When naming datasets, remember that names:

  • Can contain letters (upper and lowercase), numbers, spaces, and the following punctuation symbols: _-.,()/. Other characters are not supported.
  • Are limited to 256 characters.

For best results, use clear names that explain the source and purpose of your dataset. For example medical-device-type-1 helps identify this dataset as data relating to a particular version of a device. Use the dataset description to include more context.

Append to an existing dataset

You can add data rows to a dataset in the UI. Go to Catalog, select your dataset from the left, then click Append to dataset.

For instructions on how to append to a dataset using the Python SDK, see Dataset.

Copy the dataset ID

Each dataset has a unique dataset ID. You can find this dataset ID in the UI of Labelbox:

  1. Go to Catalog
  2. Select your dataset
  3. Copy the ID from the URL

Delete dataset

To delete a dataset, use Catalog to select the dataset and then select Delete dataset from the Dataset menu.

A confirmation dialog asks you to verify the request. To do so, type delete and then select the Delete dataset button.

Delete datasets with care, this is a permanent action that cannot be undone.

When you delete a dataset, its data rows are also deleted, along with any annotations, metadata, classifications, and other details associated with each row. Should you accidentally delete a dataset, you will need to import the original data into a new dataset and create associated details.