arc-agi-community/arc-agi-2

ARC-AGI-2是一个用于衡量人工智能一般智力的任务数据集，包含1000个训练任务和120个公开评估任务。训练任务用于展示任务格式和ARC-AGI使用的核心知识先验，可用于训练AI模型。公开评估任务用于测试之前未见过的AI模型。此外，还有两个不包含在仓库中的私有测试集，分别用于远程托管商业模型的测试和ARC Prize竞赛中的自包含模型测试。任务文件以Parquet格式存储，包含示范输入/输出对和测试输入/输出对，每个任务都需要在两次尝试内构造出正确的输出网格。

ARC-AGI-2 is a task dataset for measuring general intelligence in artificial intelligence, containing 1,000 training tasks and 120 public evaluation tasks. The training tasks are used to demonstrate the task format and the core knowledge priors used by ARC-AGI, which can be used for training AI models. The public evaluation tasks are used to test AI models that have never seen these tasks before. In addition, there are two private test sets not included in the repository, one for testing remotely hosted commercial models with low leakage probability and the other for testing self-contained models during the ARC Prize competition with near-zero leakage probability. The task files are stored in Parquet format, containing demonstration input/output pairs and test input/output pairs, and each task requires constructing the correct output grid within two attempts.