AdaReasoner/AdaEval-Jigsaw-COCO

Jigsaw COCO Dataset

Dataset Description

This dataset contains jigsaw puzzle tasks based on COCO images. The task is to identify which image piece correctly completes a given jigsaw puzzle.

Dataset Summary

• Total samples: 1000
• Task: Visual spatial reasoning with jigsaw puzzles
• Splits:
• test: 1000 samples

Task Description

Given a jigsaw puzzle image with one missing piece and three candidate pieces (A, B, C), the model needs to identify which piece correctly completes the puzzle.

Each sample consists of:

1. A question image showing the jigsaw puzzle with a missing piece
2. Three choice images (labeled A, B, C) showing candidate pieces
3. A text question asking which piece fits
4. The correct answer (A, B, or C)

Data Fields

• idx: Unique identifier for each sample
• question_image: PIL Image of the jigsaw puzzle with missing piece
• choice_images: List of 3 PIL Images (choices A, B, C)
• question_text: Text description of the task
• correct_answer: The correct choice (A, B, or C)
• choices_info: JSON string containing choice metadata
• category: Task category (jigsaw_coco)

Usage

Load Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("hitsmy/AdaEval-Jigsaw-COCO")

# Access test split
test_data = dataset["test"]

# Get a sample
sample = test_data[0]
print(f"Question: {sample['question_text']}")
print(f"Correct Answer: {sample['correct_answer']}")

# Display images
sample['question_image'].show()
for i, choice_img in enumerate(sample['choice_images']):
    print(f"Choice {chr(65+i)}:")
    choice_img.show()

#  Evaluation

import re

def evaluate_answer(prediction, ground_truth):
    """
    Evaluate if the prediction matches the ground truth
    
    Args:
        prediction: Model's predicted answer
        ground_truth: Correct answer (A, B, or C)
    
    Returns:
        bool: True if correct, False otherwise
    """
    # Normalize strings
    pred = prediction.strip().upper()
    gold = ground_truth.strip().upper()
    
    # Direct match
    if pred == gold:
        return True
    
    # Extract letter from various formats
    patterns = [
        r'\\boxed{([A-C])}',
        r'\b([A-C])\b',
        r'(?:answer|option|choice)\s*:?\s*([A-C])',
    ]
    
    for pattern in patterns:
        match = re.search(pattern, pred, re.IGNORECASE)
        if match and match.group(1).upper() == gold:
            return True
    
    return False

# Example usage
correct_count = 0
total_count = 0

for sample in test_data:
    prediction = model.predict(sample)  # Your model prediction
    is_correct = evaluate_answer(prediction, sample['correct_answer'])
    
    if is_correct:
        correct_count += 1
    total_count += 1

accuracy = correct_count / total_count
print(f"Accuracy: {accuracy:.2%}")
Dataset Statistics
Distribution by Answer
The dataset is designed to have a balanced distribution across the three choices (A, B, C).

Image Characteristics
Images are sourced from COCO dataset
Jigsaw puzzles have varying complexity
Missing piece locations vary across samples
Dataset Creation
Source Data
Base images: COCO dataset
Jigsaw generation: Programmatic puzzle generation with controlled piece sizes and shapes
Annotations
Each sample is annotated with:

Question image with missing piece
Three candidate pieces (one correct, two distractors)
Correct answer label
Quality Control
All images are verified to exist before inclusion
Incomplete samples are filtered out
Image quality is validated
Limitations and Bias
Dataset is based on COCO images, which may have inherent biases
Jigsaw puzzle difficulty may vary across samples
Primarily focused on spatial reasoning rather than semantic understanding
Citation
If you use this dataset, please cite:

@dataset{jigsaw_coco,
  title={Jigsaw COCO Dataset},
  author={Your Name},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/hitsmy/AdaEval-Jigsaw-COCO}
}
License
Apache 2.0

Contact
For questions or issues, please open an issue on the dataset repository.