Datasets:

JDWebProgrammer
/

arg-agi-augmented

This dataset is an augmented version of grids extracted from the ARC-AGI dataset (Abstraction and Reasoning Corpus). It focuses on individual grids rather than full tasks or games, providing an expanded collection for pretraining and testing models like autoencoders (AEs) or latent-space reasoners.

Source: Derived from the training split of ARC-AGI (all demonstration and test grids).
Augmentations: Each original grid is expanded with 5 transformations (horizontal flip, vertical flip, 90°/180°/270° rotations), resulting in 6 variants per grid (original + 5 augments).
Key Note: This is not the full games/tasks from ARC-AGI. It contains only the raw, augmented grids (as 2D lists of integers 0-10) for standalone use in perceptual pretraining or reconstruction testing. Use the original ARC-AGI for full few-shot reasoning tasks.

Dataset Structure

Format: Hugging Face Dataset object.
Splits: Single split (train) with one field:
- augmented_grids: List of 2D lists (grids). Each grid is [[int, ...], ...] (H x W, values 0-10).
Size: ~48,000 grids (from ~400 ARC training tasks × ~4 grids/task × 6 augments).
Metadata: See metadata.json for stats (original grids, augmentation factor).

Example grid entry:

augmented_grids[0] = [[0, 1, 0], [1, 0, 1], [0, 1, 0]]  # Example 3x3 grid

Usage

Load and use for AE pretraining:

from datasets import load_dataset
ds = load_dataset("JDWebProgrammer/arc-agi-augmented")
grids = ds['train']['augmented_grids']  # List of all grids

# Example: Batch grids for AE
def grid_to_tensor(grid):
    h, w = len(grid), len(grid[0])
    return torch.tensor(grid, dtype=torch.float).view(1, -1) / 10.0  # Normalize 0-1

batch = torch.cat([grid_to_tensor(g) for g in grids[:32]])  # Batch of 32
# Feed to AE: z = ae.encode(batch); recon = ae.decode(z)

Ideal for:

Pretraining perceptual models.
Testing reconstruction accuracy (compare original vs. augmented).
Data augmentation for fluid intelligence tasks (e.g., ARC-like pattern inference).

Generation

Extracted all input/output grids from ARC-AGI training split demos/tests.
Applied deterministic augmentations (flips/rotations) to expand variety without labels.
No synthetic generation — pure augmentation of real ARC data.

Limitations

Grids only (no task structure/context) — not for end-to-end ARC solving.
Augmentations preserve structure but may introduce artifacts (e.g., rotations on asymmetric grids).
Values 0-10 (ARC standard); normalize for models.

License

Based on ARC-AGI (CC BY-SA 4.0) — inherits same license.
Augmentations: MIT (free for research/commercial).

Citation

@misc{dataartist/arc-agi,
  title = {Augmented ARC-AGI Grids for Pretraining},
  author = {dataartist},
  year = {2025},
  url = {https://huggingface.co/datasets/your_username/arc-augmented-grids}
}

Generated for pretraining perceptual models on ARC-style puzzles. Not a substitute for full ARC tasks.