README.md

metadata

task_categories:
  - image-feature-extraction
  - image-text-to-text
tags:
  - vlm
  - mechanistic-interpretability
  - compositional-reasoning
  - synthetic-data
pretty_name: Toy Shapes Dataset
dataset_info:
  features:
    - name: num_shapes
      dtype: int64
    - name: shape_0
      dtype: string
    - name: color_0
      dtype: string
    - name: position_0
      dtype: string
    - name: shape_1
      dtype: string
    - name: color_1
      dtype: string
    - name: position_1
      dtype: string
    - name: shape_2
      dtype: string
    - name: color_2
      dtype: string
    - name: position_2
      dtype: string
    - name: shape_3
      dtype: string
    - name: color_3
      dtype: string
    - name: position_3
      dtype: string
    - name: shape_4
      dtype: string
    - name: color_4
      dtype: string
    - name: position_4
      dtype: string
    - name: image
      dtype: image
  splits:
    - name: train
      num_bytes: 1559700
      num_examples: 500
  download_size: 1354975
  dataset_size: 1559700
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
arxiv: 2508.16652

This dataset is used in the work "Do VLMs Have Bad Eyes? Diagnosing Compositional Failures via Mechanistic Interpretability" Paper

For more details, please refer to the GitHub repository.