captcha_data/README.md

metadata

license: mit
task_categories:
  - visual-question-answering
  - image-classification
  - video-classification
language:
  - en
tags:
  - captcha
  - visual-reasoning
  - spatial-reasoning
  - 3d-understanding
  - temporal-reasoning
  - benchmark
  - multimodal
size_categories:
  - 1K<n<10K

NextGen-CAPTCHAs

A comprehensive benchmark dataset of next-generation CAPTCHA challenges designed to evaluate visual reasoning, spatial understanding, and temporal processing capabilities of AI systems.

Dataset Description

This dataset contains 27 distinct CAPTCHA task types with 3,037 total files including images (PNG), animated GIFs, and ground truth annotations. The challenges span multiple cognitive dimensions including:

• 3D Spatial Reasoning: Understanding objects from different viewpoints
• Temporal Processing: Tracking motion and changes over time
• Visual Pattern Recognition: Counting, matching, and identifying patterns
• Logical Reasoning: Following paths, folding shapes, understanding shadows

Task Categories

Category	Files	Media Type	Description
3D_Viewpoint	202	Image	Match 3D wireframe objects viewed from different angles
Backmost_Layer	110	Image	Identify the backmost layer in overlapping shapes
Box_Folding	333	Image	Predict how 2D nets fold into 3D boxes
Color_Counting	34	Image	Count colors in sketches (grid selection)
Dice_Roll_Path	23	Image	Track dice faces after rolling along a path
Dynamic_Jigsaw	201	GIF	Complete jigsaw puzzles with animated pieces
Hole_Counting	62	Image	Count holes in topological shapes
Illusory_Ribbons	82	Image	Reason about illusory/impossible ribbon configurations
Layered_Stack	62	Image	Understand layered/stacked object arrangements
Mirror	10	Image	Identify correct mirror reflections
Multi_Script	202	Image	Recognize text across multiple writing systems
Occluded_Pattern_Counting	52	Image	Count patterns with partial occlusion
Red_Dot	1	Image	Track red dot position
Rotation_Match	182	Image	Match rotated shapes to reference
Shadow_Direction	106	Image	Determine shadow direction consistency
Shadow_Plausible	32	Image	Judge if shadows are physically plausible
Spooky_Circle	21	GIF	Track circles with spooky/illusory motion
Spooky_Circle_Grid	22	GIF	Grid-based spooky circle challenges
Spooky_Jigsaw	381	GIF	Jigsaw with illusory animated pieces
Spooky_Shape_Grid	33	GIF	Grid of shapes with illusory animations
Spooky_Size	21	GIF	Track size changes in illusory animations
Spooky_Text	21	GIF	Text with spooky/illusory effects
Static_Jigsaw	3	Image	Traditional static jigsaw puzzles
Structure_From_Motion	22	Image/Video	Infer 3D structure from motion cues
Subway_Paths	102	Image	Navigate complex subway/metro path networks
Temporal_Object_Continuity	21	Video	Track object identity through occlusions
Trajectory_Recovery	81	Image	Reconstruct movement trajectories

Data Format

Each task category contains:

• Media files: PNG images or GIF animations
• ground_truth.json: Annotations with prompts, options, and correct answers

Example Ground Truth Entry

{
  "color_counting_0000": {
    "prompt": "Click all sketches with 3 or fewer colors (ignore white)",
    "description": "Grid with 4 sketch(es) matching: less equal 3 colors",
    "options": ["6_bird_0.png", "6_car_0.png", "3_sun_1.png", ...],
    "answer": [2, 3, 7, 15],
    "grid_size": [4, 4],
    "difficulty": 5,
    "media_type": "image"
  }
}

Statistics

• Total Tasks: 27 categories
• Total Files: 3,037
• Image Files (PNG): 1,628
• Animation Files (GIF): 1,367
• Dataset Size: ~897 MB

Usage

from datasets import load_dataset

dataset = load_dataset("YaxinLuo/NextGen-CAPTCHAs")

Or download directly and load ground truth:

import json

with open("Color_Counting/ground_truth.json") as f:
    challenges = json.load(f)

for challenge_id, data in challenges.items():
    print(f"Prompt: {data['prompt']}")
    print(f"Answer indices: {data['answer']}")

Intended Use

This dataset is intended for:

• Benchmarking multimodal AI systems on visual reasoning tasks
• Research on spatial and temporal understanding in AI
• Evaluating robustness of vision-language models
• Studying CAPTCHA security against automated solvers

Citation

If you use this dataset, please cite:

@dataset{nextgen_captchas_2025,
  author = {Yaxin Luo},
  title = {NextGen-CAPTCHAs: A Benchmark for Visual Reasoning},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/YaxinLuo/NextGen-CAPTCHAs}
}

License

apache-2.0