Data Preparation#

Data preparation is the foundation of successful computer vision models. Ultralytics Platform provides comprehensive tools for managing your training data, from upload through annotation to analysis.

Overview#

The Data section of Ultralytics Platform helps you:

• Upload images, videos, and dataset files (ZIP, TAR including .tar.gz/.tgz, NDJSON)
• Import from a URL by pasting a direct link to an archive or NDJSON export, or from Roboflow
• Connect Google Cloud Storage, Amazon S3, or Azure Blob Storage and use your data in place without uploading a copy
• Keep pixels on premise with Enterprise On Premise CPU/GPU workers
• Annotate with manual drawing tools and SAM-powered smart labeling — choose from SAM 2.1 or the new SAM 3
• Manage classes by renaming, recoloring, merging, and deleting them across the whole dataset
• Analyze your data with statistics, visualizations, and embedding-based clustering
• Export in NDJSON format for local training

Workflow#

graph LR
    A[Upload]:::start --> B[Annotate]:::proc
    B --> D[Train]:::out
    B --> C[Analyze]:::proc

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

Supported Tasks#

Ultralytics Platform datasets support all 7 YOLO task types:

Key Features#

Smart Storage#

Ultralytics Platform manages storage efficiently:

• Deduplication: Identical images in the same data region are stored once
• Integrity: Uploads are verified for data integrity
• Efficiency: Optimized storage and fast processing

Dataset URIs#

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

yolo train data=ul://username/datasets/my-dataset

This allows training on the platform's datasets from any machine with your API key configured.

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
model.train(data="ul://username/datasets/my-dataset", epochs=100)

Dataset Versioning#

Create immutable NDJSON snapshots of your dataset for reproducible training. Each version captures image counts, class counts, and annotation counts at the time of creation. See Versions Tab for details.

Dataset Tabs#

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

Classes appears when the dataset has images and its task has classes, while Charts appears whenever it has images. Errors appears only when processing failures exist. Versions appears when you have edit access, or in read-only mode when versions already exist.

Clustering#

Explore your dataset as an interactive 2D scatter plot where visually similar images sit close together — useful for surfacing clusters, duplicates, and outliers, and for inspecting how splits or classes are distributed across your data. Lasso a region of the plot to filter the gallery to those images. Analysis needs between 20 and 200,000 non-errored images. See Clustering for details.

Statistics and Visualization#

The Charts tab provides automatic analysis including:

• Split Distribution: Donut chart of train/val/test image counts
• Top Classes: Donut chart of the 10 most frequent annotation classes
• Image Dimensions: Histogram of image width and height distribution (in pixels)
• Image Dimensions 2D: 2D heatmap of width vs height with aspect ratio guide lines
• Annotation Locations: 2D heatmap of bounding box center positions
• Points per Instance: Polygon vertex or keypoint count distribution (segment/pose datasets)

See the Charts tab for the full list.

Quick Links#

• Datasets: Upload, manage, and export your training data
• Annotation: Label data with manual and AI-assisted tools
• Cloud Training: Train models on your annotated datasets
• Dataset URI: Use ul:// URIs to train from anywhere

FAQ#