thegovind/azure-architecture-vqa

Azure Architecture Visual Question Answering Dataset

A comprehensive visual question answering (VQA) dataset for fine-tuning vision language models to become Azure Cloud Solution Architects. Created from the Azure Architecture Center.

Dataset Description

This dataset contains Q&A pairs paired with Azure architecture diagrams, designed for fine-tuning vision language models (like Qwen 3.5 VL) to understand and reason about cloud architecture patterns.

Source

• Azure Architecture Center: 373 architecture pages scraped
• Architecture Diagrams: 500+ high-quality architecture diagrams (PNG format)
• Q&A Generation: Generated using Qwen 2.5 72B Instruct via HuggingFace Inference API

Question Categories

Each architecture generates 5 Q&A pairs across these categories:

1. Architecture Overview (free_form): High-level design purpose and patterns
2. Component Identification (free_form): Azure services used and their roles
3. Design Decision (multi_choice): Design choices and best practices
4. Visual Understanding (free_form): Diagram layout, data flow, and connections
5. Scenario Application (multi_choice): When and how to apply the pattern

Architecture Categories Covered

• Reference Architectures (containers, networking, identity, etc.)
• Example Scenarios (IoT, data, mainframe, SAP, etc.)
• Design Patterns (CQRS, Event Sourcing, Gateway, etc.)
• Solution Ideas (analytics, AI/ML, hybrid, etc.)
• Best Practices and Guides
• Cloud migration patterns
• Microservices architectures
• Serverless patterns
• Hybrid and multi-cloud designs

Dataset Structure

Data Fields

Field	Type	Description
pid	string	Unique question ID
question	string	The question text
image	string	Image filename reference
decoded_image	Image	The architecture diagram (PIL Image)
choices	list[string]	Answer choices for multi_choice, empty for free_form
answer	string	The correct answer
question_type	string	"free_form" or "multi_choice"
answer_type	string	Answer format type
metadata	string	JSON with category, skills, source, page info
query	string	Formatted query with hints

Splits

Split	Rows
train	1,678
test	187

Intended Use

Fine-tuning Vision Language Models

This dataset is specifically designed for fine-tuning models like:

• Qwen 3.5 VL (0.8B, 3B, 8B variants)
• Other vision-language models that support image+text input

Target Capability

Train a model to act as an Azure Cloud Solution Architect that can:

• Analyze architecture diagrams
• Identify Azure services and their roles
• Explain design decisions and trade-offs
• Recommend architectures for given scenarios
• Understand data flow and system interactions

Example

from datasets import load_dataset

ds = load_dataset("thegovind/azure-architecture-vqa")

# View a sample
sample = ds['train'][0]
print(f"Question: {sample['question']}")
print(f"Answer: {sample['answer']}")
print(f"Type: {sample['question_type']}")
if sample['decoded_image']:
    sample['decoded_image'].show()

Citation

@dataset{azure_architecture_vqa_2026,
  title={Azure Architecture Visual Question Answering Dataset},
  author={thegovind},
  year={2026},
  url={https://huggingface.co/datasets/thegovind/azure-architecture-vqa},
  note={Created from Azure Architecture Center documentation}
}

License

This dataset is released under CC-BY-SA-4.0. The source content is from Microsoft's Azure Architecture Center documentation.