Train a Model

Train custom vision-language models (VLMs) by fine-tuning state-of-the-art architectures on your annotated data. Transform general-purpose models into specialized AI that understands your domain, terminology, and visual patterns.

💡

New to VLM training?

Start with the quickstart guide to train your first model in under an hour. This section provides comprehensive documentation for production training workflows.

📋

Prerequisites

• Prepared dataset with images and annotations
• Training project set up in Datature Vi
• Understanding of VLM concepts — phrase grounding or VQA
• Compute Credits for GPU training resources

Get started with quickstart →

Training workflow overview

Training a VLM on Datature Vi follows a structured workflow with five key stages:

1. Create Workflow
   ↓
2. Configure System Prompt
   ↓
3. Configure Dataset
   ↓
4. Configure Model
   ↓
5. Launch Training Run

Each stage builds on the previous one, culminating in a reusable workflow that you can execute across multiple training runs with consistent configuration.

Why workflows?

Workflows are reusable training configurations that capture your complete setup:

• System prompt — Natural language instructions that guide model behavior
• Dataset configuration — Train-test splits, shuffling, and data distribution
• Model selection — Architecture choice and training parameters
• Reproducibility — Run the same configuration multiple times for experimentation

Once configured, workflows enable rapid iteration and experimentation without reconfiguring settings each time.

Training process stages

Stage 1: Create a workflow

Workflows are the foundation of your training configuration. They encapsulate all settings required to train a model and can be reused across multiple training runs.

What you'll configure:

• System prompt configuration — Instructions that guide model behavior
• Dataset selection and splitting — Choose data source and configure train-test splits
• Model architecture and settings — Select VLM and configure training parameters

Complete guide: Create a Workflow →

Stage 2: Configure system prompt

System prompts are natural language instructions that define what your VLM should learn during training. They shape model behavior for both phrase grounding and VQA tasks.

Key decisions:

• Task definition — What should the model detect or answer?
• Domain terminology — Industry-specific language and concepts
• Output format — How should the model structure responses?
• Edge case handling — Behavior for ambiguous or out-of-scope inputs

Example system prompts:

You are an AI assistant specialized in identifying printed circuit board (PCB)
components in manufacturing images. Given a text description of a component
type (e.g., "capacitor", "resistor", "IC chip"), locate and mark all instances
of that component in the image with precise bounding boxes.

Focus on:
- Component type identification based on visual characteristics
- Clear distinction between similar components (e.g., capacitors vs. resistors)
- Accurate bounding box placement around component boundaries
- Detection of partially visible or occluded components

If the described component is not present in the image, respond with no
bounding boxes.

You are a quality control AI assistant for automotive manufacturing. Answer
questions about vehicle assembly images with clear, specific responses based
only on what is visible in the image.

Your responses should:
- Be concise and factual (1-2 sentences maximum)
- Reference specific visual evidence when possible
- Indicate uncertainty when details are unclear or ambiguous
- Use standard automotive terminology

For yes/no questions, provide reasoning in parentheses. For count questions,
provide the exact number visible. For defect questions, describe the type and
location clearly.

Complete guide: Configure Your System Prompt →

Stage 3: Configure dataset

Dataset configuration determines how your training data is split, shuffled, and distributed across training, validation, and test sets. Proper configuration ensures your model learns effectively and generalizes to new data.

Configuration options:

Common split strategies:

• 80/20 split — Standard for most datasets (1000+ images)
• 90/10 split — When validation data is limited (< 500 images)
• 70/20/10 split — Add dedicated test set for final evaluation

Complete guide: Configure Your Dataset →

Stage 4: Configure model

Model configuration involves two key choices: selecting the right architecture for your task and tuning training settings to optimize performance.

Model architecture selection

Choose from 4 state-of-the-art VLM architectures (with more coming soon):

Architecture selection factors:

• Task complexity — Larger models (7B+) for complex reasoning; smaller models (1-2B) for speed
• Compute budget — Larger models require more GPU memory and training time
• Inference speed — Smaller models deploy faster with lower latency
• Domain requirements — Specialized architectures for OCR, multilingual, or document tasks

View complete architecture comparison →

Model settings and hyperparameters

Configure training behavior and optimization parameters:

Core settings:

• Training mode — LoRA fine-tuning (efficient) or full fine-tuning (maximum customization)
• Batch size — Number of samples per training step (affects GPU memory and convergence)
• Learning rate — Controls speed and stability of model updates
• Epochs — Number of complete passes through training data

Complete guide: Configure Your Model →

Stage 5: Launch training run

After configuring your workflow, launch a training run with specific hardware and checkpoint settings.

Training run configuration:

1. Advanced settings — Checkpoint frequency and evaluation options
2. Hardware configuration — GPU type and quantity selection
3. Dataset validation — Automatic checks ensure data readiness
4. Review summary — Verify configuration before launch

GPU selection guidance:

Compute costs:

Training consumes Compute Credits based on GPU type and duration:

Example: 4× A10G GPUs for 2 hours
  Usage multiplier: 10.0 credits/minute
  Duration: 120 minutes
  Total cost: 1,200 Compute Credits

Complete guide: Configure Training Settings →

View GPU pricing and specifications →

After training completes

Once training finishes, you can evaluate performance, compare runs, and deploy your model.

Training best practices

Common questions

Related resources

• Create a training project — Set up your training environment
• Prepare your dataset — Upload images and create annotations
• Phrase grounding concepts — Understanding text-guided object detection
• Visual question answering concepts — Understanding VQA tasks
• Resource usage — Understanding Compute Credits and GPU pricing
• Monitor training runs — Track training progress in real-time
• Manage workflows — Edit, delete, and organize workflows
• Create a workflow — Define training configuration
• Configure your model — Select model architecture and settings
• Configure training settings — Set checkpoint strategy and GPU
• Evaluate a model — Assess model performance and quality
• Manage runs — Kill or delete runs
• Quickstart — End-to-end training tutorial