Datasets#

Ultralytics Platform datasets provide a streamlined solution for managing your training data. After upload, the platform processes images, labels, and statistics automatically.

A dataset is ready to train once processing has completed and it has at least one image in the train split, at least one image in either the val or test split, and at least one labeled image. The dataset header shows a Ready badge when all three conditions are met, and a Not Ready badge otherwise — click the badge to see exactly which condition is missing.

Collect Images with Agents#

Use Agents to add images to a dataset when a model detection meets a condition, such as confidence within a chosen range. Connect the condition to a Dataset block and select a destination you can edit. Images arrive unlabeled in the train split; existing copies are skipped. Review and label them with the existing annotation editor.

Upload Dataset#

Ultralytics Platform accepts multiple upload formats for flexibility.

Supported Formats#

Video Codec Support#

The file extension alone isn't enough: a video can still fail if its codec cannot be decoded during processing.

ffmpeg -i input.mov \
  -c:v libx264 -pix_fmt yuv420p \
  -c:a aac -movflags +faststart \
  output.mp4

Preparing Your Dataset#

The Platform supports Ultralytics YOLO, COCO, LabelMe JSON, semantic PNG masks, depth datasets, Ultralytics NDJSON and Labelbox NDJSON, and raw (unannotated) uploads:

my-dataset/
├── images/
│   ├── train/
│   │   ├── img001.jpg
│   │   └── img002.jpg
│   └── val/
│       ├── img003.jpg
│       └── img004.jpg
├── labels/
│   ├── train/
│   │   ├── img001.txt
│   │   └── img002.txt
│   └── val/
│       ├── img003.txt
│       └── img004.txt
└── data.yaml

# data.yaml
path: .
train: images/train
val: images/val

names:
    0: person
    1: car
    2: dog

If labels reference class IDs but no class names are supplied, Platform remaps the IDs to a dense 0-indexed sequence and names each class after its source ID (class0, class3, …), which you can rename later in the Classes tab.

For task-specific format details, see supported tasks and the Datasets Overview.

Upload Process#

To create a dataset:

1. Navigate to Annotate in the sidebar
2. Click New Dataset
3. Pick a data source tab (see Data Sources below)
4. Add a name — the URL slug is derived automatically and can be edited — plus an optional description
5. Select the task type (see supported tasks), an optional license (see available licenses), and visibility (public or private)
6. Click Create & Upload for local files, Create & Import for a URL or connected source, or Create Dataset to start empty

To add files to an existing dataset, open its dataset page and either drag the files onto the gallery or click the upload icon in the page header. The upload icon opens your browser's native file picker directly because the dataset task is already defined.

Data Sources#

The New Dataset dialog offers four sources:

Depth datasets can only be created from Upload or URL; the Cloud and On Premise tabs are disabled for the depth task.

Before the Upload Starts#

Platform validates your files in the browser before uploading anything, so common problems surface immediately rather than after a long transfer. Archives are checked for corruption, emptiness, password protection, and YAML syntax errors, and oversized files are listed by name.

Two dialogs may then appear:

After upload, the platform processes your data automatically:

graph LR
    A[Upload]:::start --> B[Validate]:::proc
    B --> C[Normalize]:::proc
    C --> D[Thumbnail]:::proc
    D --> E[Parse Labels]:::proc
    E --> F[Statistics]:::out

    classDef start fill:#4CAF50,color:#fff
    classDef proc fill:#2196F3,color:#fff
    classDef out fill:#9C27B0,color:#fff

1. Validation: Format and size checks
2. Normalization: Large images resized (max 4096px, min dimension 28px), grayscale expanded to RGB, transparency flattened onto white, and EXIF orientation applied; TIFF originals are stored as uploaded
3. Thumbnails: 256px WebP previews generated
4. Label Parsing: YOLO, COCO, LabelMe, and Ultralytics or Labelbox NDJSON labels extracted
5. Statistics: Class distributions and image dimensions computed

from ultralytics.data.utils import check_det_dataset

check_det_dataset("path/to/data.yaml")

Browse Images#

View your dataset images in multiple layouts.

Open the Clustering panel from the gallery toolbar to explore your dataset as an interactive 2D scatter plot.

Sorting and Filtering#

Images can be sorted and filtered for efficient browsing:

Fullscreen Viewer#

Click any image to open the fullscreen viewer with:

• Navigation: Arrow keys or thumbnail previews to browse
• Image information: Review Platform-generated properties, custom metadata, and embedded file metadata such as EXIF
• Custom metadata: Owners and editors can add or replace a JSON object, including nested values up to 500,000 serialized characters and top-level keys up to 128 characters
• Annotations: Toggle annotation overlay visibility
• Class Breakdown: Per-class label counts with color indicators
• Annotate: When you have edit access, annotation controls are active immediately when the fullscreen viewer opens on desktop
• Blur faces: Right-click an annotation to preview and apply face blurring to the image when available
• Download: Download the original image file
• Delete: Delete the image from the dataset
• Zoom: Cmd/Ctrl+Scroll, Cmd/Ctrl++, or Cmd/Ctrl+= to zoom in, and Cmd/Ctrl+- to zoom out
• Reset view: Cmd/Ctrl + 0 or the reset button to fit the image to the viewer
• Pan: Hold Space and drag, or drag with the middle mouse button, to pan the canvas at any zoom level
• Pixel view: Toggle pixelated rendering for close inspection
• Depth curtain: On depth datasets, a draggable divider wipes between the RGB image and its colorized depth map

Filter by Split#

Filter images by their dataset split:

Clustering#

The Clustering panel projects your dataset into an interactive 2D scatter plot where visually similar images sit close together. Use it to surface clusters, spot duplicates and outliers, and inspect how splits or classes are distributed across your data — without leaving the gallery. Open it from the scatter-chart icon in the gallery toolbar on any dataset page. The panel is desktop-only, and viewers without edit access see the icon once the dataset has been analyzed.

Running Analysis#

Start an analysis:

1. Open a dataset and click the scatter-chart icon in the gallery toolbar
2. Click Analyze Dataset
3. Wait for the progress bar to finish — results appear in the same panel

Analysis runs in the background in two stages, Computing embeddings and Clustering, and can take a few minutes depending on the size of your dataset. You can close the panel or leave the page and come back later.

Visualization#

Once analysis completes, the panel shows a 2D scatter of all analyzed images with a legend and a point counter. Gallery filters (split, class, labeled/unlabeled) dim out-of-filter points so you can focus on the subset you care about — the counter then reads visible / total points.

Color By#

Change how data points are shaded with the Color by dropdown in the panel toolbar. Switch view modes at any time — the plot re-colors instantly so you can see how splits, classes, or image properties are distributed across your clusters:

Lasso and Click Selection#

Draw a free-form selection around a region to highlight points on the plot, or click a point to select every point drawn in the same color (only that point when coloring by width, height, size, or annotations); click an empty area of the plot to clear the selection. The gallery filters down to the matching images, so you can inspect, relabel, move, or delete them using the usual image operations.

Pan and Zoom#

Navigate large scatters directly from your mouse and keyboard, or with the zoom buttons at the bottom-left of the plot:

Re-analyzing#

If your dataset changes after analysis — new images arrive, or the analyzed count no longer matches the dataset — or the analysis predates the Clusters color option, a Re-analyze button appears at the top of the panel for owners and editors.

Click Re-analyze to recompute embeddings and the 2D projection from scratch.

Find Similar Images#

The same embeddings power similarity search across public datasets and your own and team datasets. In a dataset you can edit, right-click an image in Grid or Compact view (or a single selected row in Table view) and choose Find similar images. The dialog shows the source image beside up to 24 of the nearest images, each labeled with its similarity score, excluding images your dataset already holds and copies of the selected image in other datasets. Click an image to preview it full size and see its source dataset, license, and similarity score. Select images with their checkboxes (or select all), then click Add N to dataset: they are added to the train split as unlabeled images, counted against your storage, and ready for annotation.

An image without an embedding — in a dataset not yet analyzed, or added since the last analysis — is embedded when you open the dialog, so you do not need to run a Clustering analysis first. The dialog is unavailable on connected datasets. A model's per-image validation diagnostics run the same search from its worst-performing images.

Dataset Tabs#

Each dataset page can show up to six tabs, depending on the dataset state and your permissions:

Images Tab#

The default view showing the image gallery with annotation overlays. Supports grid, compact, and table view modes. Drag and drop files here to add more images.

Classes Tab#

This tab appears when the dataset has images and its task has classes.

Manage annotation classes for your dataset:

• Class histogram: Bar chart showing annotation count per class, sorted by frequency, with a linear/log scale toggle
• Class table: Sortable, searchable table with index, name, annotation count, and image count
• Edit class names: Click any class name to rename it inline
• Edit class colors: Click a color swatch to change the class color
• Add new class: Use the input at the bottom to add classes
• Merge classes: Select two or more rows and click Merge into one
• Delete classes: Select one or more rows and click Delete

Merge Classes#

Merging consolidates duplicate or overlapping labels — for example folding car, automobile, and vehicle into one class:

1. Select two or more classes using the row checkboxes
2. Click Merge into one in the table header, or right-click the selection
3. Choose which of the selected classes is the target the others merge into
4. Confirm

Every annotation belonging to the source classes is reassigned to the target class, and the source classes are removed. No annotations are deleted, so the dataset's total annotation count is unchanged.

Delete Classes#

1. Select one or more classes using the row checkboxes
2. Click Delete in the table header, or right-click the selection
3. Confirm

Deleting a class removes the class and all of its annotations. For classification datasets the labels are removed but the images remain, becoming unannotated.

Charts Tab#

This tab appears when the dataset has images.

Automatic statistics computed from your dataset:

Charts appear in this order, and each one is omitted when the dataset has no data for it — a raw image dataset shows no annotation charts, and Points per Instance only appears for segment and pose data:

Models Tab#

View all models trained on this dataset in a searchable table:

Errors Tab#

This tab appears only when one or more files fail processing.

Images that failed processing are listed here with:

• Error banner: Total count of failed images and guidance
• Error table: Filename, user-friendly error description, fix hints, and preview thumbnail
• Common errors include corrupted files, unsupported formats, images too small (min 28px), and unsupported color modes

Versions Tab#

Create immutable NDJSON snapshots of your dataset for reproducible training. Each version captures image counts, class counts, annotation counts, and the storage it added at the time of creation.

To create a version:

1. Open the Versions tab
2. Optionally enter a description (e.g., "Added 500 training images" or "Fixed mislabeled classes")
3. Click New Version
4. The new version appears in the table. If the dataset matches an existing version, for example right after restoring it, that version is reused and takes the new description if you entered one

Each version is numbered sequentially (v1, v2, v3...) and is immutable — versions cannot be edited or removed, only their descriptions can be changed. Use the row actions to download or restore any version at any time.

Compare Versions#

Click the compare icon on any version after v1 to see what changed since an earlier version. Compare versions starts from the previous version; pick another one in the From menu. Chips grouped under Images, Annotations, Classes, and Settings count the images added, removed, modified, and moved to another split and the annotations added and removed, and name the classes added, removed, or renamed and any dataset settings that changed. Each changed image is listed with its badge. Select an image to see it Before and After, each side with the labels that version stored. When the images and annotations are identical, the dialog reports No changes, or No image changes when only class definitions or dataset settings differ.

Export Dataset#

Export your dataset for offline use with an NDJSON download from the dataset header or the Versions tab.

To export:

1. Click the Download button (download icon) in the dataset header
2. Download the current NDJSON snapshot directly
3. Use the Versions tab when you want an immutable numbered snapshot you can re-download later

The NDJSON format stores one JSON object per line. The first line contains dataset metadata, followed by one line per image:

{"type": "dataset", "task": "detect", "name": "my-dataset", "description": "...", "bytes": 12345678, "url": "https://platform.ultralytics.com/...", "class_names": {"0": "person", "1": "car"}, "version": 1, "created_at": "2026-01-15T10:00:00Z", "updated_at": "2026-02-20T14:30:00Z"}
{"type": "image", "file": "img001.jpg", "url": "https://...", "width": 640, "height": 480, "split": "train", "metadata": {"location": {"site": "factory-1"}, "reviewed": true}, "annotations": {"boxes": [[0, 0.5, 0.5, 0.2, 0.3]]}}
{"type": "image", "file": "img002.jpg", "url": "https://...", "width": 1280, "height": 720, "split": "val"}

The optional image-level metadata object is preserved when an NDJSON file is imported into Platform. You can inspect or edit it from the image's fullscreen information panel. For programmatic archive uploads, the Ingest Dataset Data API accepts the equivalent imageMetadata path map.

Pose datasets also carry a kpt_shape field in the dataset header line, inferred from the annotations when it is not already set.

Depth exports declare "task": "depth" and "depth_scale": 1000 in the dataset line. Each image line includes a depth.url for its paired plain uint16 PNG. Ultralytics downloads image and depth URLs concurrently and writes a training YAML with the same scale, so the NDJSON file can be passed directly to model.train(data=...).

See the Ultralytics NDJSON format documentation for full specification.

Image Operations#

Quick Actions#

Right-click an image in Grid or Compact view, or a single selected row in Table view, to access quick actions. Available actions depend on your edit permissions and the dataset source:

Bulk Move to Split#

Reassign selected images to a different split within the same dataset:

1. Switch to Table view
2. Select images using checkboxes
3. Right-click to open the context menu
4. Choose Move to split > Train, Validation, or Test

You can also drag and drop images onto the split filter tabs in grid view. If moving an image would collide with an identical image already in the target split, Platform asks whether to skip, keep both, or replace.

Split Redistribution#

Redistribute all images across train, validation, and test splits using custom ratios:

1. Click the split bar in the dataset toolbar to open the Redistribute Splits dialog
2. Adjust split percentages using any of the methods below
3. Review the live image count preview to confirm the distribution
4. Click Apply to randomly reassign all images according to your percentages

The dialog provides three ways to set your target split ratios:

A live preview shows exactly how many images will land in each split before you apply.

Bulk Delete#

Delete multiple images at once:

1. Select images in the table view
2. Right-click and choose Delete, or press Cmd/Ctrl+Delete
3. Confirm deletion

Copy and Move Images#

Copy or move images from one dataset you can edit into another, including a dataset in a different workspace:

1. In the source dataset, right-click an image in Grid or Compact view and choose Copy or Cut, or select images in Table view and press Cmd/Ctrl+C or Cmd/Ctrl+X. Esc clears the clipboard.
2. Open the destination dataset, right-click an image and choose Paste, or press Cmd/Ctrl+V.

Pasted images keep their labels and splits, and images the destination already holds in the same split are skipped. Cut removes the pasted images from the source dataset; Copy leaves it unchanged. The source and destination must have the same task and compatible image channels, pose keypoint settings, and depth scale, even when the copied images have no labels. An empty destination can inherit unset image-channel and pose settings. Images cannot be pasted into a connected dataset.

Classes are matched by name, ignoring case except for one- and two-character names, and a destination without classes takes the source's class list. When a pasted image uses a class the destination does not have, the Map classes dialog asks you to map each such class to a dataset class or a new class, or to clear its Include checkbox to drop that class's labels; the images are pasted either way.

Generate Similar Images#

Create new training images from one you already have. In a dataset you can edit, right-click an image in Grid or Compact view (or a single selected row in Table view) and choose Generate similar images. Platform describes new scenes inspired by the image and generates one image for each, so the results share the source's subjects and style without copying it.

Ultralytics Image 9B and Krea 2 Turbo are also listed, but you can select them only after you turn on Early access. Krea 2 Turbo is the slowest option. Source proportions are preserved where supported. Very narrow images may need a larger longest edge, and output dimensions are rounded and limited to the generator's supported sizes.

Click Generate, or press ⌘/Ctrl+Enter in Instructions. Images appear as they finish, all selected, and Stop ends a running batch while keeping the images that already arrived. To try other instructions or settings, change them and click Generate more: each new batch appears above the earlier ones, which keep their selection. Click an image to view it full size, clear the checkbox of any you don't want, and click Add N to dataset. The kept images are uploaded as JPEGs named after the source image, without labels, counted against your storage, and ready for annotation. They use the active split filter: choose Train before generating to add them to train. With All selected, normal upload split assignment applies, including automatic validation splitting when needed. Cancel discards the previews without adding them. Generated images follow the dataset's upload face-blurring setting. The action is unavailable on connected datasets.

Blur Faces#

In a dataset you can edit, blur the faces in its images, for example to protect the privacy of people in your data. Blur Faces is not available for connected datasets or for datasets with more than three image channels.

• One image: right-click the image in Grid or Compact view (or a single selected row in Table view), or right-click an annotation in the fullscreen viewer, and choose Blur faces.
• Whole dataset: open More actions (⋯) on the dataset page and choose Blur faces.

The dialog first previews the detected faces on up to six images (or on the one image) without changing them. Adjust Confidence (default 0.25) and Box scale (0.5–1.5, default 1, which scales each face box around its center) to re-run the preview, then click Apply to replace the original pixels of every image in which faces are found. Images without detected faces are left unchanged, labels and splits are kept, and some faces may be missed, so review the result. Blurring does not automatically create a version snapshot. Blurring a whole dataset costs $1.00 per 1,000 processed images, with a minimum of $0.01 per run (billed as Auto-Annotation), and the dialog shows the estimate before you apply; previews and single-image blurring are free.

To blur faces in images as they are uploaded, turn on Blur faces when you create a dataset from Upload or URL, or Blur future uploads in the whole-dataset Blur faces dialog. The Blur future uploads switch saves immediately, even if you close the dialog without clicking Apply. Images uploaded to the dataset afterward are blurred during processing, at no charge; changing this setting does not blur images already in the dataset.

Dataset URI#

Reference Platform datasets using the ul:// URI format (see Using Platform Datasets):

ul://username/datasets/dataset-slug

You can also paste a dataset or model web URL directly (e.g. https://platform.ultralytics.com/username/datasets/dataset-slug); it is automatically rewritten to the ul:// URI. Passing a list of datasets fine-tunes one base model across each in series, for example model.train(data=["ul://username/datasets/a", "ul://username/datasets/b"]).

Use this URI to train models from anywhere:

export ULTRALYTICS_API_KEY="YOUR_API_KEY"
yolo train model=yolo26n.pt data=ul://username/datasets/my-dataset epochs=100

Available Licenses#

The Platform supports the following licenses for datasets:

Visibility Settings#

Control who can see your dataset:

Visibility is set when creating a dataset in the New Dataset dialog using a toggle switch. To change it later, click the Public or Private badge next to the dataset name in the page breadcrumb; making a dataset public asks for confirmation. On Premise datasets are always private. Public datasets are visible on the Explore page.

Edit Dataset#

Dataset metadata is edited inline directly on the dataset page — no dialog needed:

• Name: Click the dataset name to edit it. Changes auto-save on blur or Enter. Names are limited to 100 characters.
• Description: Click the description (or "Add a description..." placeholder) to edit. Changes auto-save. Descriptions are limited to 1,000 characters.
• Task type: Click the task badge to select a different task type.
• License: Click the license selector to change the dataset license.
• Icon: Click the dataset icon to upload your own image, promote one of the dataset's own images to the icon, or pick a letter and color.

Custom Metadata#

Open More actions and select Information to review two sections:

• Ultralytics Metadata: Read-only Platform details such as the dataset ID, owner, task, image and annotation counts, storage region, and timestamps
• Custom Metadata: Your own JSON object for provenance, capture conditions, customer IDs, governance, or other contextual data

Workspace viewers can inspect metadata, while members with edit access can replace the custom metadata object. The serialized metadata object is limited to 500,000 characters, and each top-level key is limited to 128 characters. Save an empty object ({}) to clear custom metadata.

Clone Dataset#

When viewing a public dataset you do not own, click Clone Dataset to open the clone dialog. Review the destination workspace, name, visibility, and license, then confirm the clone. The copy includes all images, annotations, and class definitions. Public source datasets stay public by default in workspaces whose default visibility is public; Enterprise workspace clones default to private. If the original dataset has a copyleft license, the clone inherits it and the license selector is locked.

The clone dialog keeps Clone Dataset disabled while the URL is already used in the target workspace, and cloning requires enough remaining storage quota to hold the copy.

Star and Share#

• Star: Click the star button to bookmark a dataset. The star count is visible to all users.
• Share: For public datasets, click the share button to copy a link, grab an embed snippet, or share to social platforms.

Delete Dataset#

Delete a dataset you no longer need:

1. Open More actions (the … button) in the dataset header
2. Choose Delete Dataset
3. Confirm in the dialog: "This will move [name] to trash. You can restore it within 30 days."

The same menu holds Information, which opens the metadata dialog described in Custom Metadata, and Refresh, which re-reads the dataset from the server.

Train on Dataset#

Start training directly from your dataset:

1. Click New Model on the dataset page
2. Select a project or create new
3. Configure training parameters
4. Start training

graph LR
    A[Dataset]:::start --> B[New Model]:::proc
    B --> C[Select Project]:::proc
    C --> D[Configure]:::proc
    D --> E[Start Training]:::out

    classDef start fill:#4CAF50,color:#fff
    classDef proc fill:#2196F3,color:#fff
    classDef out fill:#9C27B0,color:#fff

See Cloud Training for details.

FAQ#