Amaze

Dataset: https://huggingface.co/datasets/piekenius123/Amaze

Amaze is a benchmark for Edting-as-Reasoning task (EAR). It features four maze shapes: circle, hexagon, square, and triangle. Each sample provides: an unmarked maze image (original_img), a maze image with start and end points marked (m_original_img), a blue solution path image (sol_img), a binary path mask (mask_img), a cell segmentation map (cell_map), and metadata (JSON) for describing the maze structure and difficulty.

The test set covers various sizes from 3×3 to 16×16 (50 samples for each size), while the training set mainly consists of 3×3 mazes (1024 samples), and validation set consists of 3×3 mazes (256 samples).

Browse samples by shape / split / maze size, then view images + metadata.