Hegarty Dataset

🏔️ Overview

The Hegarty Dataset is a comprehensive benchmark for evaluating spatial reasoning and perspective-taking abilities in vision-language models. Inspired by Piaget's classic Three Mountains Task, this dataset provides a systematic evaluation across three cognitive dimensions in both controlled lab and real-world settings.

📊 Dataset Structure

2×3 Design

• 2 Settings: Lab (controlled) and Real-world
• 3 Task Types:
  • Visual Access Reasoning: Determining object visibility from different viewpoints
  • Mental Reasoning: Understanding spatial relationships between objects
  • Perspective-Taking Reasoning: Understanding what others can see from their viewpoint

Dataset Statistics

Setting	Task Type	Samples	Description
Lab	Visual Access	~1,000	Controlled geometric scenarios
Lab	Mental Reasoning	~500	Spatial relationship tasks
Lab	Perspective-Taking	~300	Other-perspective understanding
Real	Visual Access	~2,000	Real-world ego-exo scenarios
Real	Mental Reasoning	~2,000	Natural spatial reasoning
Real	Perspective-Taking	~2,000	Real-world perspective-taking

🎯 Task Descriptions

Visual Access Reasoning (formerly Perspective-Taking L1)

• Question: "Can you see object X from viewpoint Y?"
• Format: Binary (Yes/No) or Multiple Choice
• Cognitive Ability: Basic visual perspective-taking

Mental Reasoning (formerly Spatiality)

• Question: "What is the spatial relationship between objects?"
• Format: Multiple Choice (4 options)
• Cognitive Ability: Spatial reasoning and mental rotation

Perspective-Taking Reasoning (formerly Perspective-Taking L2)

• Question: "What can the observer at position X see?"
• Format: Multiple Choice (4 options)
• Cognitive Ability: Theory of mind and advanced perspective-taking

💡 Usage

from datasets import load_dataset

# Load specific configuration
dataset = load_dataset("Hokin/hegarty", "lab_visual_access")

# Load all lab settings
lab_data = {
    'visual_access': load_dataset("Hokin/hegarty", "lab_visual_access"),
    'mental_reasoning': load_dataset("Hokin/hegarty", "lab_mental_reasoning"),
    'perspective_taking': load_dataset("Hokin/hegarty", "lab_perspective_taking")
}

# Load all real-world settings
real_data = {
    'visual_access': load_dataset("Hokin/hegarty", "real_visual_access"),
    'mental_reasoning': load_dataset("Hokin/hegarty", "real_mental_reasoning"),
    'perspective_taking': load_dataset("Hokin/hegarty", "real_perspective_taking")
}

📈 Model Performance

Current state-of-the-art models show significant gaps compared to human performance:

• Humans: 80-93% accuracy across all tasks
• Best AI Models: 30-55% accuracy
• Key Finding: Models perform better on visual access than mental reasoning or perspective-taking

🔗 Related Resources

• Paper: [Coming Soon]
• GitHub: grow-ai-like-a-child/hegarty
• Leaderboard: [Coming Soon]

📝 Citation

@dataset{three_mountain_2024,
  title={Hegarty: A Comprehensive Benchmark for Spatial Reasoning and Perspective-Taking},
  author={Hokin Deng and Contributors},
  year={2024},
  publisher={HuggingFace}
}

📄 License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

🤝 Contributing

We welcome contributions! Please see our GitHub repository for guidelines.

🏷️ Tags

spatial-reasoning perspective-taking visual-reasoning cognitive-science vision-language benchmark evaluation