Datasets:

Ardea
/

arc_agi_v1

license: apache-2.0
task_categories:
  - table-question-answering
tags:
  - arc
  - agi
  - arc-agi
pretty_name: ARC AGI v1
size_categories:
  - 1K<n<10K

ARC-AGI-V1 Dataset (A Take On Format)

This dataset is a reorganized version of the ARC-AGI v1 (Abstraction and Reasoning Corpus) benchmark, formatted for HuggingFace Datasets.

Dataset Structure

The original ARC-AGI dataset has been transformed from its file-based JSON structure into a standardized HuggingFace dataset with two splits:

train (400 examples): Tasks from the original training directory
test (400 examples): Tasks from the original evaluation directory

Original Structure

The original ARC-AGI dataset consisted of:

A training directory with JSON files (one per task)
An evaluation directory with JSON files (one per task)
Each JSON file named with a task ID (e.g., 007bbfb7.json)
Each file containing:
- train: Array of input/output example pairs for learning the pattern
- test: Array of input/output pairs representing the actual task to solve

Transformed Structure

Each row in this dataset represents a single ARC-AGI task with the following schema:

{
    "id": string,           // Task ID from the original filename
    "list": [               // Combined training examples and test inputs
        [                   // Training example inputs (from original 'train')
            [[int]], [[int]], ...
        ],
        [                   // Training example outputs (from original 'train')
            [[int]], [[int]], ...
        ],
        [                   // Test inputs (from original 'test')
            [[int]], [[int]], ...
        ]
    ],
    "label": [              // Test outputs (from original 'test')
        [[int]], [[int]], ...
    ]
}

Field Descriptions

id: The unique task identifier from the original filename
list: A nested list containing three components in order:
1. Example inputs (list[0]): All input grids from the original train array
2. Example outputs (list[1]): All output grids from the original train array (paired with example inputs)
3. Test inputs (list[2]): All input grids from the original test array
label: The correct output grids for the test inputs (from original test array outputs)

Data Format

Each grid is represented as a 2D array of integers (0-9), where:

Values range from 0 to 9 (representing different colors/states)
Grid dimensions vary from 1×1 to 30×30
Each integer represents a colored cell in the grid

Example

{
    "id": "007bbfb7",
    "list": [
        [
            [[0, 7, 7],                            // Example input 1
            [7, 7, 7],                             // 
            [0, 7, 7]],                            // 
            [[4, 0, 4], [0, 0, 0], [0, 4, 0]],     // Example input 2
            [[0, 0, 0], [0, 0, 2], [2, 0, 2]]      // Example input 3
        ],
        [
            [[0, 0, 0, 0, 7, 7, 0, 7, 7],          // Example output 1
            [0, 0, 0, 7, 7, 7, 7, 7, 7],
            [0, 0, 0, 0, 7, 7, 0, 7, 7],
            [0, 7, 7, 0, 7, 7, 0, 7, 7],
            [7, 7, 7, 7, 7, 7, 7, 7, 7],
            [0, 7, 7, 0, 7, 7, 0, 7, 7],
            [0, 0, 0, 0, 7, 7, 0, 7, 7],
            [0, 0, 0, 7, 7, 7, 7, 7, 7],
            [0, 0, 0, 0, 7, 7, 0, 7, 7]],
            [[], [], [], [], [], [], [], [], []],  // etc..
        ],
        [
            [[7, 0, 7], [7, 0, 7], [7, 7, 0]]      // Test input 1
        ]
    ],
    "label": [
        [[7, 0, 7, 0, 0, 0, 7, 0, 7],              // Test output 1 (ground truth)
        [7, 0, 7, 0, 0, 0, 7, 0, 7],
        [7, 7, 0, 0, 0, 0, 7, 7, 0],
        [7, 0, 7, 0, 0, 0, 7, 0, 7],
        [7, 0, 7, 0, 0, 0, 7, 0, 7],
        [7, 7, 0, 0, 0, 0, 7, 7, 0],
        [7, 0, 7, 7, 0, 7, 0, 0, 0],
        [7, 0, 7, 7, 0, 7, 0, 0, 0],
        [7, 7, 0, 7, 7, 0, 0, 0, 0]]
    ]
}

Usage Philosophy

pprint(dataset['train']['list'][0][0][0]) pprint(dataset['train']['list'][0][1][0]) print('') pprint(dataset['train']['list'][0][2][0]) pprint(dataset['train']['label'][0][0])

This ARC-AGI dataset format allows (me at least) to think about the tasks in this way:

Learn from examples: Study the input/output pairs:
- input: dataset['train']['list'][0][0][0]
- output: dataset['train']['list'][0][1][0]
- input: dataset['train']['list'][0][0][1]
- output: dataset['train']['list'][0][1][1]
- where:
  - 1st num: task number
  - 2nd num: either 0: example input || 1: example output
  - 3rd num: which example?
Then 'Get the tests':
- dataset['train']['list'][0][2][0]
Apply the pattern: Use the learned rule to make your two guesses
Evaluate performance: Compare model predictions against the label field
- dataset['train']['label'][0][0]

Training Split

Contains all tasks from the original training directory
Intended for model training and development
Both example pairs and test solutions are provided

Test Split

Contains all tasks from the original evaluation directory
Intended for final model evaluation
In competition settings, test labels may be withheld

Dataset Features

Features({
    'id': Value('string'),
    'list': List(List(List(List(Value('int64'))))),
    'label': List(List(List(Value('int64'))))
})

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("ardea/arc_agi_v1")

# Access splits
train_data = dataset['train']
test_data = dataset['test']

# Example: Get a single task
task = train_data[0]
task_id = task['id']
example_inputs = task['list'][0]
example_outputs = task['list'][1]
test_inputs = task['list'][2]
test_outputs = task['label']

# Example: Get a task by id
task = list(filter(lambda t: t['id'] == '007bbfb7', train_data))

Transparency

I've left the script I used on the original dataset here as arc_to_my_hf.py

Citation

If you use this dataset, please cite the original ARC-AGI work:

@misc{chollet2019measure,
    title={On the Measure of Intelligence},
    author={François Chollet},
    year={2019},
    eprint={1911.01547},
    archivePrefix={arXiv},
    primaryClass={cs.AI}
}

License

This dataset maintains the Apache 2.0 license from the original ARC-AGI corpus.