KotshinZ/arc-agi-augmented-100

ARC-AGI Augmented Dataset

This dataset is an augmented version of the Abstraction and Reasoning Corpus (ARC-AGI), processed for training neural networks (such as Transformers or Neural Cellular Automata).

Dataset Details

• Original Source: ARC-AGI Benchmark
• License: MIT
• Augmentation Method:
  • Dihedral Transformations: 8 symmetries (rotations/flips).
  • Color Permutation: Random permutation of colors 1-9 (0 is fixed as background).
  • Translational Padding: Randomly positioning the grid within a 30x30 canvas.
• Seed: 42
• Augmentation Factor: 100 per puzzle.

Statistics

• Total Original Puzzles: 1189
• Total Augmented Examples: 350417

Data Structure

Each row in the dataset represents a single input/output example pair (flattened).

• input_ids: Flattened array (int32) of the input grid.
  • Values: 0 (Pad), 1 (EOS), 2-11 (Colors 0-9).
  • Dimensions: 30 x 30 flattened.
• labels: Flattened array (int32) of the output grid.
• puzzle_id: Integer ID for the puzzle.
• original_puzzle_id: The hex string ID from the original ARC dataset (e.g., 007bbfb7).
• group_id: Identifies the augmentation group. All examples with the same group_id are variations of the same puzzle.

Usage

from datasets import load_dataset

dataset = load_dataset("KotshinZ/arc-agi-augmented-100")
print(dataset["train"][0])