Manage Datasets

Effective dataset management is essential for maintaining high-quality training data and keeping your workspace organized. Datature Vi provides comprehensive tools to rename, download, analyze, and clean up your datasets throughout the model development lifecycle.

This guide covers all aspects of dataset management, from basic operations like renaming and downloading to advanced maintenance tasks like bulk asset cleanup and insight analysis.

💡

Complete dataset workflow

Create dataset → Upload data → Annotate → Manage datasets (you are here) → Train model

Core dataset operations

Datature Vi provides five essential categories of dataset management operations:

Quick reference

Common dataset management tasks and where to find them:

Renaming datasets

Keep your workspace organized with descriptive, meaningful dataset names.

When to rename

• Improved clarity — Make dataset purpose clear to team members
• Project evolution — Update names as content or scope changes
• Standardization — Apply consistent naming conventions
• Version management — Distinguish between dataset versions

Key features

• Safe operation — Dataset ID remains unchanged; all integrations continue working
• No downtime — Active training runs and workflows are unaffected
• Instant updates — Name changes appear immediately across the platform
• Unlimited changes — Rename as often as needed

Learn how to rename datasets →

Downloading data

Export your datasets and annotations for backup, local development, or external processing.

Export options

Common use cases

Download methods

• Web interface — Point-and-click export with progress tracking
• Vi SDK — Programmatic downloads with client.get_dataset()
• Automated workflows — Schedule periodic backups via SDK

Explore download options →

Viewing dataset insights

Analyze your dataset composition, annotation distribution, and quality metrics to ensure optimal training data.

Available insights

• Asset statistics — Total count, file types, size distribution
• Annotation counts — Phrase grounding pairs, VQA pairs, distribution
• Class balance — See if your annotations are evenly distributed
• Quality indicators — Identify potential issues or gaps
• Split information — Training vs validation set breakdown

When to check insights

Learn how to view insights →

Managing assets

Maintain dataset quality by removing poor quality images, duplicates, or unnecessary assets.

Asset management operations

Common scenarios

Important considerations

⚠️

Asset deletion is permanent

• Cannot be undone — Deleted assets are removed immediately
• Annotations lost — All annotations associated with deleted assets are also removed
• Always backup first — Download dataset before large-scale deletions
• No recovery option — Only way to restore is re-uploading from backups

Learn about asset management →

Deleting datasets

Permanently remove datasets you no longer need to keep your workspace organized and manage storage.

When to delete

• Completed projects — Remove datasets after project completion
• Test datasets — Clean up experimental or prototype datasets
• Duplicate data — Remove redundant datasets
• Storage optimization — Free up space for new projects
• Workspace cleanup — Maintain organized, clutter-free environment

Safety measures

Before deleting a dataset:

1. Create a backup — Download the full dataset
2. Check dependencies — Verify no active training runs depend on this data
3. Inform team — Notify collaborators about the planned deletion
4. Review carefully — Ensure you're deleting the correct dataset

🔒

Deletion is permanent and irreversible

• All assets (images, videos) are deleted
• All annotations are permanently removed
• Dataset ID becomes invalid
• Training runs lose reference to source data
• Cannot be recovered through any means
• SDK queries for deleted dataset ID will fail

Always export a backup first if there's any chance you'll need the data later.

Learn how to delete datasets →

Best practices for dataset management

Dataset management workflow

Follow this recommended workflow for maintaining healthy datasets:

1. Regular monitoring

• Check insights weekly — Review statistics and distributions
• Track annotation progress — Monitor annotation completion rates
• Identify quality issues — Spot problems early

2. Periodic cleanup

• Remove poor quality data — Delete blurry, corrupted, or off-topic images
• Eliminate duplicates — Use bulk operations for efficiency
• Archive old versions — Download and remove superseded datasets

3. Backup strategy

• Before major changes — Always download before bulk deletions
• Monthly backups — Regular exports for important datasets
• Version snapshots — Download before significant modifications

4. Organization maintenance

• Standardize names — Apply consistent naming conventions
• Delete test data — Remove experimental datasets after use
• Document structure — Maintain team documentation on dataset organization

Programmatic dataset management

For advanced users, the Vi SDK provides programmatic access to all dataset management operations:

import vi

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

# List all datasets
for dataset in client.datasets:
    print(f"Dataset: {dataset.name} (ID: {dataset.dataset_id})")

# Download a dataset
result = client.get_dataset(
    dataset_id="dataset-id",
    save_dir="./backups"
)

# Delete a dataset
client.datasets.delete("dataset-id")

Learn more about Vi SDK →

Troubleshooting

Next steps

Related resources

• Create a dataset — Set up new datasets for different task types
• Upload data — Add images and annotations to datasets
• Annotate data — Create and edit annotations
• Vi SDK getting started — Programmatic dataset management
• Resource usage — Monitor Data Rows consumption
• Rename a dataset — Update dataset names
• Delete a dataset — Remove unused datasets
• Download data — Export full datasets or annotations
• View insights — Analyze dataset statistics
• Manage assets — Delete assets and perform bulk actions
• Train a model — Use datasets for VLM training
• Quickstart — Complete training workflow