ARC-AGI Dataset

Table of contents

• ARC-AGI Dataset
  • 2024 ARC-AGI Dataset
    • Data format
    • Visualization

2024 ARC-AGI Dataset

Let us begin with understanding the data.

Data format

All the data is JSON encoded, the structures of JSON are as followings (* stands for training , evaluation and test) :

• *_challenges

  {
    "task_id": {
        "test": [
            {
                "input": [[...], ..., [...]]
            }
        ],
        "train": [
            {
                "input": [[...], ..., [...]],
                "output": [[...], ..., [...]]
            },
            ...
        ]
    }, 
    ...
  }
  

• *_solutions

  {
    "task_id": [[...], ..., [...]],
    ...
  }
  

• submission format

  {
    "task_id": [
        {
            "attempt_1": [[...], ..., [...]],
            "attempt_2": [[...], ..., [...]]
        },
    ],
    ...
  }
  

All the arrays have integer digits in range 0 to 9, and each element stands index of some hex_colors,

hex_colors = ['#000000', '#0074D9','#FF4136','#2ECC40','#FFDC00',
     '#AAAAAA', '#F012BE', '#FF851B', '#7FDBFF', '#870C25']

Visualization

Let’s visualize the data (I choose f35d900a and fcb5c309) Basically, we visaulize the data in 2d-grid with colors, however, we could treat the color values not as color channel but as positional values in z-axis, thinking that we applied one-hot encoding. The third one is colored 3-d scatter, just for highlights the z-level.

	f35d900a	fcb5c309
Basic
3d scatter
3d scatter (colored)