# ARC Prize 2024 — Data

来源：https://www.kaggle.com/competitions/arc-prize-2024/data

采集时间：2026-10-19T22:21:34Z

The objective of this competition is to create an algorithm that is capable of solving abstract reasoning tasks. Critically, these are *novel* tasks: tasks that the algorithm has never seen before. Hence, simply memorizing a set of reasoning templates will not suffice.

The format is different from the previous competition, so please read this information carefully, and refer to supplementary documentation as needed.

When looking at a task, a "test-taker" has access to inputs and outputs of the demonstration pairs (train pairs), plus the input(s) of the test pair(s). The goal is to construct the output grid(s) corresponding to the test input grid(s), using **2 trials** for each test input. "Constructing the output grid" involves picking the height and width of the output grid, then filling each cell in the grid with a symbol (integer between 0 and 9, which are visualized as colors). Only **exact** solutions (all cells match the expected answer) can be said to be correct.

Any additional information, as well as an interactive app to explore the objective of this competition is found at the  [ARCPrize.org](http://ARCPrize.org/play). _**It is highly recommended that you explore the interactive app, as the best way to understand the objective of the competition**._

## Task files

The information is stored in two files:
- **arc-agi_training-challenges.json**: contains input/output pairs that demonstrate reasoning pattern to be applied to the "test" input for each task. This file and the corresponding **solutions** file can be used as training for your models.
- **arc-agi_training-solutions.json**: contains the corresponding task "test" outputs (ground truth).
- **arc-agi_evaluation-challenges.json**: contains input/output pairs that demonstrate reasoning pattern to be applied to the "test" input for each task. This file and the corresponding **solutions** file can be used as validation data for your models.
- **arc-agi_evaluation-solutions.json**: contains the corresponding task "test" outputs (ground truth).
- **arc-agi_test-challenges.json**: this file contains the tasks that will be used for the leaderboard evaluation, and contains "train" input/output pairs as well as the "test" input for each task. Your task is to predict the "test" output. **Note:** The file shown on this page is a placeholder using tasks from **arc-agi_evaluation-challenges.json**. When you submit your notebook to be rerun, this file is swapped with the actual test challenges.
- **sample_submission.json**: a submission file in the correct format

Each task contains a dictionary with two fields:
- `"train"`: demonstration input/output pairs. It is a list of "pairs" (typically 3 pairs).
- `"test"`: test input - your model should predict the output.

A "pair" is a dictionary with two fields:
- `"input"`: the input "grid" for the pair.
- `"output"`: the output "grid" for the pair.

A "grid" is a rectangular matrix (list of lists) of integers between 0 and 9 (inclusive). The smallest possible grid size is 1x1 and the largest is 30x30.

The data on this page should be used to develop and evaluate your models. When notebooks are submitted for rerun, they are scored using 100 unseen tasks found in the rerun file named **arc-agi_test_challenges.json**. The rerun tasks will contain `train` pairs of inputs and outputs as well as the tasks `test` input. Your algorithm must predict the `test` output. The majority of the 100 tasks used for leaderboard score only have one `test` input that will require a corresponding output prediction, although for a small number of tasks, you will be asked to make predictions for two `test` inputs.