CLEVR-Sudoku

Dataset Summary

CLEVR-Sudoku is a challenging visual puzzle dataset requiring both visual object perception and reasoning capabilities. Each sample contains:

• A partially filled Sudoku puzzle presented with CLEVR-based imagery.
• Separate “options” images that illustrate how specific object properties map to digits.
• Metadata specifying puzzle attributes and the ground-truth solution.

Designed to encourage visual reasoning and pattern recognition, this dataset provides 6 different subsets (no further train/val/test splits). Each subset contains 1,000 puzzles, leading to 6,000 total puzzles.

---

Supported Tasks

• Visual Reasoning: Models must interpret compositional 3D object scenes to solve Sudoku logic.
• Pattern Recognition: Identify how object attributes (shape, color, size, etc.) correlate with digit placements.

---

Dataset Structure

Data Instances

A typical data instance has the following structure:

{
    "sudoku": 9x9 array of images or None,
    "options": 9x5 or 9x10 array of images,
    "attributes": dict that maps each digit to an attribute combination,
    "id": integer identifier for the puzzle,
    "solution": 9x9 array of integers (full sudoku)
}

Data Splits / Subsets

There are 6 different subsets, 3 for CLEVR-Easy (CLEVR-Easy-K10, CLEVR-Easy-K30, CLEVR-Easy-K50) and 3 for CLEVR (CLEVR-4-K10, CLEVR-4-K30, CLEVR-4-K50). For CLEVR-Easy only color and shape are relevant for the digit mapping while for CLEVR also material and size relevant are. K indicates how many cells are empty in the sudoku, i.e. K10 means that there are 10 empty cells. Each subset contains 1,000 puzzles. Currently, there are no separate training/validation/test splits within each subset.

Dataset Creation

Curation Rationale

Goal: Combine the compositional “CLEVR”-style visual complexity with the logical constraints of Sudoku. This setup pushes models to perform both visual recognition (understanding shapes, colors, etc.) and abstract reasoning (solving Sudoku).

Source Data

Synthetic Generation: The images are created in a CLEVR-like manner, programmatically generated with variations in shape, color, position, etc. Sudoku Logic: Each puzzle is automatically generated and there exists exactly one solution to the puzzle.

Annotations

Automatic Generation: Since the dataset is synthetic, the puzzle solutions are known programmatically. No human annotation is required for the puzzle solutions. Attributes: Each digit (1–9) is associated with one or more visual properties (e.g., color = "red", shape = "cube"). These are also generated systematically.

Personal and Sensitive Information

None: The dataset is purely synthetic, containing no personal or demographic data.

Usage

Loading the Dataset

Example usage:

from datasets import load_dataset

dataset = load_dataset("AIML-TUDA/CLEVR-Sudoku", "CLEVR-Easy-K10")
print(dataset[0])

The second argument ("CLEVR-Easy-K10", "CLEVR-Easy-K30", etc.) corresponds to the 6 different subsets. Each subset has 1,000 puzzles.

Citation

If you use or reference this dataset in your work, please cite the following:

@article{stammer2024neural,
  title={Neural Concept Binder},
  author={Stammer, Wolfgang and W{\"u}st, Antonia and Steinmann, David and Kersting, Kristian},
  journal={Advances in Neural Information Processing Systems},
  year={2024}
}