GEOM-3-28: A Synthetic Concept-Annotated Dataset for Concept Bottleneck Models

GEOM-3-28 is a synthetic image dataset developed for benchmarking Concept Bottleneck Models (CBMs) under controlled conditions. The dataset contains 28,000 images equally distributed across 28 classes, with 1,000 images per class. Each image shows a scene composed of simple geometric objects generated through a rule-based process. For this version we have figures with 3 shapes: Triangles, Circles and Squares. Every sample is accompanied by complete concept annotations describing the presence, size and spatial relationships between the objects. Presence: triangle_present, circle-present, square_presentLargest: triangle_largest, circle_largest, square_largestSpatial: triangle_above_circle, triangle_left_circle, triangle_above_square, triangle_left_square, circle_above_square, circle_left_square The dataset also includes the official training, validation and test splits. Both PNG and SVG versions of every generated image are provided. The complete dataset generation pipeline is publicly available in the accompanying GitHub repository, allowing the dataset to be reproduced or extended.