MultiInteract-Bench Dataset

A Benchmark Dataset for Evaluating Web Interaction Reconstruction from Image Sequences

📋 Overview

MultiInteract-Bench is a comprehensive dataset designed to evaluate the capabilities of multimodal large language models in reproducing web-based interactions from image sequences. The dataset contains real-world web interface snapshots showing progressive states of web applications through user interactions.

Key Features

• Multi-turn Interactions: Each task includes a sequence of web page states showing the progression of user interactions
• Real-world Applications: Covers popular web applications like Spotify, Stripe, and more
• Comprehensive Metadata: Each task includes detailed metadata describing interaction steps
• High-quality Images: PNG format screenshots with clear visual elements
• Diverse Scenarios: Includes music players, payment forms, and various web UI patterns

📊 Dataset Structure

Task Format

Each task in the dataset follows this structure:

task_name_timestamp/
├── metadata.json           # Task metadata and interaction descriptions
├── step_00.png           # Initial state (before any interaction)
├── step_01.png           # State after step 1 interaction
├── step_02.png           # State after step 2 interaction
└── ...                   # Additional interaction steps

Metadata Structure

Each metadata.json file contains:

{
  "id": "task_name_timestamp",
  "description": "Brief description of the web application",
  "steps": [
    {
      "step_index": 0,
      "description": "Initial state description",
      "image": "step_00.png"
    },
    {
      "step_index": 1,
      "description": "First interaction description",
      "image": "step_01.png"
    }
  ]
}

📦 Dataset Contents

This dataset includes:

• Total Tasks: Multiple real-world web interaction scenarios
• Steps per Task: Typically 5-7 interaction steps
• Image Format: PNG
• Image Resolution: High-resolution screenshots
• Applications: Various popular web platforms

🎯 Use Cases

MultiInteract-Bench is designed for:

1. Model Evaluation: Benchmarking multimodal LLMs on web interaction reconstruction
2. Web Development: Testing automated web page generation systems
3. UI/UX Research: Studying web interface patterns and interactions
4. Computer Vision: Evaluating image-to-code generation capabilities
5. Agent Systems: Training and testing web automation agents

🚀 Quick Start

Download the Dataset

# Using huggingface-cli
huggingface-cli download zionzionzion/MultiInteract-Bench --repo-type dataset

# Or download the zip file directly
wget https://huggingface.co/datasets/zionzionzion/MultiInteract-Bench/resolve/main/dataset_multi_turn.zip
unzip dataset_multi_turn.zip

Load in Python

import json
from pathlib import Path

# Load a specific task
task_path = "dataset_multi_turn/Spotify_1766618072"
with open(f"{task_path}/metadata.json", "r") as f:
    metadata = json.load(f)

print(f"Task: {metadata['id']}")
print(f"Description: {metadata['description']}")
print(f"Number of steps: {len(metadata['steps'])}")

# Access images
for step in metadata['steps']:
    image_path = f"{task_path}/{step['image']}"
    print(f"Step {step['step_index']}: {step['description']}")
    print(f"  Image: {image_path}")

🔧 Related Repository

For the complete evaluation framework including:

• Model reproduction scripts
• Visual metrics calculation
• Automated screenshot capture
• Statistical analysis tools

Please visit our GitHub repository.

Evaluation Metrics

The associated repository implements 8 evaluation metrics:

1. CLIP Similarity - Semantic alignment (0-1, higher is better)
2. LPIPS Distance - Perceptual similarity (0-∞, lower is better)
3. Style Loss - Artistic style consistency (0-∞, lower is better)
4. Text Similarity - Text content preservation (0-1, higher is better)
5. Color Histogram Similarity - Color distribution (0-1, higher is better)
6. Dominant Color Similarity - Primary color consistency (0-1, higher is better)
7. DINO Similarity - Structural layout (0-1, higher is better)
8. SSIM - Structural fidelity (0-1, higher is better)

📊 Dataset Statistics

Metric	Value
Total Tasks	Multiple scenarios
Total Images	5-7 per task
Image Format	PNG
Metadata Format	JSON
Languages	English, Chinese

📝 Citation

If you use MultiInteract-Bench in your research, please cite:

@dataset{multinteract_bench_2026,
  title = {MultiInteract-Bench: A Benchmark Dataset for Evaluating Web Interaction Reconstruction from Image Sequences},
  author = {Yang, Tiankun},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/zionzionzion/MultiInteract-Bench}
}

📧 Contact

For questions, issues, or suggestions regarding this dataset, please contact:

Email: yangtiankun25@mails.ucas.cn

📄 License

This dataset is provided under the MIT License. See the LICENSE file for details.

🔗 Links

• GitHub Repository
• Dataset Download
• HuggingFace Space (if applicable)

---

Note: This dataset is intended for research and educational purposes. Please respect the terms of service of the web applications from which screenshots were captured.