Download Data

Downloading data from Datature Vi enables you to create backups, develop locally, process data externally, or migrate to other platforms. You can export complete datasets with all assets and annotations, or download only the annotation data for lightweight transfers.

This guide helps you choose the right export method and format for your specific needs.

💡

Two ways to download data

• Full dataset — Assets (images/videos) + annotations in organized folders
• Annotations only — Just the annotation data in Vi JSONL or TFRecord format

Choose based on whether you need the actual image files or just the annotation metadata.

Export options

Which export method should I use?

Choose the right download method based on your use case:

Download full dataset

Download your complete dataset including all assets (images, videos) and their annotations in a training-ready folder structure.

What's included

Best for

• Backup and disaster recovery — Complete copy of your dataset
• Local development — Train models on your own infrastructure
• Data migration — Move datasets to other platforms or systems
• External processing — Use custom tools or pipelines
• Offline access — Work without internet connectivity
• Version control — Create snapshots of dataset versions

Export formats

Full dataset exports include:

• Assets — Original image/video files in their native formats (JPEG, PNG, etc.)
• Annotations — Vi JSONL format with all annotation data
• Folder structure — Organized directories for training workflows:

  dataset-name/
  ├── assets/
  │   ├── train/
  │   └── validation/
  └── annotations/
      └── annotations.jsonl

Learn how to download full datasets →

Download annotations only

Export just the annotation data without asset files for lightweight transfers, analysis, or format conversion.

What's included

• Annotation metadata — All phrase grounding and VQA annotation data (freeform coming soon)
• Bounding box coordinates — Spatial information for phrase grounding
• Text data — Captions, phrases, questions, and answers
• Asset references — File paths and IDs (without the actual files)
• Dataset metadata — Information about splits and structure

Best for

• Lightweight backups — Small file size, fast transfer
• Annotation analysis — Analyze labels without processing images
• Format conversion — Convert Vi annotations to other formats
• Label review — Check annotation quality and distribution
• Programmatic access — Load and parse annotation data in scripts
• Sharing labels — Send annotations to team members without large files

Export formats

Choose from two annotation-only formats:

Learn how to download annotations →

Comparison: Full dataset vs annotations only

Download methods

You can download data through the web interface or programmatically with the Vi SDK:

Web interface downloads

Best for: One-time exports, small to medium datasets, visual workflows

Features:

• Point-and-click export
• Real-time progress tracking
• Visual format selection
• Browser-based downloads
• No coding required

Access:

1. Navigate to dataset in Explorer
2. Click download button in dataset menu
3. Choose export format and options
4. Monitor download progress

Programmatic downloads with Vi SDK

Best for: Large datasets, automated workflows, repeated exports, CI/CD integration

Features:

• Efficient batch processing
• Automated retry logic
• Progress tracking in code
• Integration with pipelines
• Scheduled exports

Example:

import vi

client = vi.Client(
    secret_key="your-secret-key",
    organization_id="your-org-id"
)

# Download full dataset
result = client.datasets.download(
    dataset_id="dataset-id",
    save_dir="./data"
)

# Download annotations only (recommended convenience method)
result = client.annotations.download(
    dataset_id="dataset-id",
    save_dir="./annotations"
)

# Alternative: Download annotations via datasets API
result = client.datasets.download(
    dataset_id="dataset-id",
    save_dir="./annotations",
    annotations_only=True
)

Learn more about Vi SDK →

Export formats explained

Vi JSONL format

JSON Lines format with one annotation record per line:

{"asset_id": "asset-123", "filename": "image001.jpg", "annotations": [...]}
{"asset_id": "asset-124", "filename": "image002.jpg", "annotations": [...]}

Advantages:

• Human-readable and easy to understand
• Simple to parse with standard JSON libraries
• One record per line makes streaming efficient
• Easy to convert to other formats
• Works with text editors and command-line tools

Use cases:

• Annotation analysis and statistics
• Format conversion to COCO, YOLO, etc.
• Custom processing scripts
• Quality assurance reviews
• Label distribution analysis

TFRecord format

TensorFlow's binary format optimized for training:

Advantages:

• Highly efficient binary format
• Optimized for TensorFlow data pipelines
• Faster I/O compared to text formats
• Compressed storage for large datasets
• Built-in support in TensorFlow ecosystem

Use cases:

• TensorFlow model training
• Performance-critical data loading
• Large-scale training workflows
• Production ML pipelines
• Distributed training systems

Common use cases

Best practices

Download size and time estimates

Plan your downloads by understanding approximate sizes and times:

Full dataset downloads

Annotation-only downloads

💡

Tip: Download times vary based on your internet connection speed, server load, and asset file sizes. Use Vi SDK for more reliable downloads of large datasets with automatic retry logic.

Troubleshooting

Next steps

Related resources

• Manage datasets — Overview of all dataset management operations
• Rename a dataset — Update dataset names
• Delete a dataset — Permanently remove datasets
• Vi SDK Getting Started — Programmatic data access
• Vi SDK Datasets API — Complete dataset API reference