Prepare Your Dataset

📍

Quickstart: Step 1 of 3

This is the first step in the quickstart guide. After preparing your dataset, you'll train a VLM and deploy it.

Before you can train a VLM, you need a dataset with images and annotations. Follow these three focused steps to get your data ready for training.

⏱️ Time to complete: ~15 minutes

📚 What you'll learn: Dataset creation, image uploading, and annotation basics

📋

Prerequisites

You'll need a Datature Vi account. Sign up for free if you haven't already.

Three steps to prepare your dataset

Follow these steps in order to set up your data for training:

Quick overview

What you'll do

1. Create a dataset — Choose between Phrase Grounding, Visual Question Answering, or Freeform (coming soon), then configure storage
2. Upload images — Use drag-and-drop or SDK to add your images
3. Add annotations — Import existing labels or create them manually

What you'll need

• Your images in supported formats (.jpg, .png, etc.)
• Annotations (optional - you can create them in Datature)
• About 15 minutes

💡

Need more detail?

This quickstart covers the essentials. For comprehensive guides, see:

• Complete dataset creation guide
• Complete image upload guide
• Complete annotation guide

Tips for success

• Start with 20-50 images for initial testing
• Ensure annotation filenames match your image filenames exactly
• Use Multi-Region storage for best performance
• Check out our AI-assisted annotation tools to speed up labeling
• Learn about Phrase Grounding vs Visual Question Answering concepts

Dataset types explained

Not sure which dataset type to choose? Here's a quick guide:

What's next?

Once you've completed all three steps, your dataset will be ready for training.

Train Your VLM →

Create a training workflow and start fine-tuning your vision-language model on your prepared dataset.

Related resources

• Create a dataset — First step: configure dataset settings
• Upload images — Second step: add images
• Add annotations — Third step: import labels
• Train a VLM — Next: train your model
• Annotate data — Create phrase grounding and VQA annotations
• Upload data guide — Detailed upload instructions
• Manage datasets — Organize and maintain datasets
• Concepts — Understanding VLM concepts
• Vi SDK — Programmatic uploads with Python
• View insights — Check dataset quality
• Quickstart overview — Back to main quickstart
• Contact us — Get help from the Datature team