README.md

metadata

task_categories:
  - image-text-to-text
tags:
  - multimodal
  - mllm
  - geometric-reasoning
  - visual-question-answering
  - shape-recognition
  - chain-of-thought
  - mathematics
  - reasoning
language:
  - en
dataset_info:
  features:
    - name: shape
      dtype: string
    - name: background_color
      dtype: string
    - name: image
      dtype: image
    - name: metadata
      dtype: string
  splits:
    - name: regular_polygons
      num_bytes: 3950491.492
      num_examples: 1948
    - name: regular_polygon_pairs
      num_bytes: 17922128.490000002
      num_examples: 5090
    - name: abstract_shapes
      num_bytes: 1522583
      num_examples: 403
    - name: heptagons_with_visual_cues
      num_bytes: 6340402.2
      num_examples: 1400
    - name: arrow_on_plus_with_visual_cues
      num_bytes: 9327783.92
      num_examples: 1540
  download_size: 26192011
  dataset_size: 39063389.102
configs:
  - config_name: default
    data_files:
      - split: regular_polygons
        path: data/regular_polygons-*
      - split: regular_polygon_pairs
        path: data/regular_polygon_pairs-*
      - split: abstract_shapes
        path: data/abstract_shapes-*
      - split: heptagons_with_visual_cues
        path: data/heptagons_with_visual_cues-*
      - split: arrow_on_plus_with_visual_cues
        path: data/arrow_on_plus_with_visual_cues-*
library_name:
  - pytorch

Forgotten Polygons: Multimodal Large Language Models are Shape-Blind

This dataset is part of the work "Forgotten Polygons: Multimodal Large Language Models are Shape-Blind".
📖 Read the Paper
💾 GitHub Repository

Overview

This dataset is designed to evaluate the shape understanding capabilities of Multimodal Large Language Models (MLLMs).

Sample Usage

This dataset is designed to be used with the evaluation code provided in the GitHub Repository. To evaluate MLLMs on various tasks using this dataset, follow the instructions in the evaluation folder of the repository.

For example, to run a shape identification task using LLaVA-1.5:

# Navigate to the 'evaluation' folder in the cloned GitHub repository
cd Shape-Blind/evaluation
# Run the evaluation script
python3 evaluate_MLLMs.py --model_version llava-1.5 --task shape_id --dataset_size full

Dataset Splits

Each split corresponds to a different reasoning task and shape identification challenge.

🟢 Regular Polygons (regular_polygons)

• Task: Shape Identification & Sides Counting
• Description: Consists of images of regular polygons (e.g., triangles, pentagons, hexagons, etc.).
• Example Queries:
  • "What shape is in the image?"
  • "How many sides does the shape in the image have?"

🟡 Regular Polygon Pairs (regular_polygon_pairs)

• Task: Multi-Shape Reasoning
• Description: Images contain two distinct polygons. The task involves identifying both shapes, counting their sides, and summing the total.
• Example Query:
  • "What are the two shapes in the image, and how many sides do they have in total?"

🔵 Abstract Shapes (abstract_shapes)

• Task: Complex Shape Recognition
• Description: Features irregular and merged polygons, stars, arrows, and abstract geometric figures.
• Example Query:
  • "How many sides does this shape have?"

🟣 Heptagons with Visual Cues (heptagons_with_visual_cues)

• Task: Visually-Cued Chain-of-Thought (VC-CoT) Reasoning
• Description: Evaluates VC-CoT prompting by annotating it with visual cues on top of heptagon images.
• We chose heptagons because it was the most difficult regular polygon for MLLMs.
• The annotations range from ordered numbers and letters, to random numbers and letters.
• Example Query:
  • "Observe the shape and list the numbers you see. How many sides does the shape have?"

🔴 Arrow on Plus with Visual Cues (arrow_on_plus_with_visual_cues)

• Task: VC-CoT with Alternative Visual Cues
• Description: Similar to the heptagons_with_visual_cues split but using arrow-on-plus shapes instead.
• Example Query:
  • "Count the total number of numbers associated with the shape’s sides."

Citation

If you use this dataset, please cite:

Forgotten Polygons: Multimodal Large Language Models are Shape-Blind
Arxiv: 2502.15969