JUHE API Marketplace
DATASET
Open Source Community

ARC (Abstraction and Reasoning Corpus)

ARC (Abstraction and Reasoning Corpus) is a dataset jointly created by Gwangju Institute of Science and Technology and Korea University, intended to evaluate and improve AI systems' abstract reasoning capabilities. It contains various complex grid‑editing tasks with large action spaces and diverse task types. The dataset was generated using the Gymnasium environment by defining specific action and state spaces to simulate ARC challenges. ARC is primarily used in reinforcement learning to develop and test AI models capable of solving complex reasoning problems.

Updated 7/30/2024
arXiv

Description

ARCLE - ARC Learning Environment

Overview

ARCLE is a reinforcement‑learning environment for training Abstraction and Reasoning Corpus and similar datasets, built on Farama’s Gymnasium.

Environments

ARCLE includes the following environments:

  • RawARCEnv: Environment that only supports coloring, resizing answer grid, and submission actions.
  • ARCEnv: Implements ARC’s browser‑test interface, supporting coloring, flood‑fill, copy, and paste operations.
  • O2ARCv2Env: Implements the O2ARCv2 interface, extending ARCEnv with move, rotate, flip, crop, etc.

Version History

0.2.6

  • Environment changes:
    • O2ARCv2Env now accessible via O2ARCEnv or O2ARCEnv‑v2.
    • Unified observation type to np.int8.
    • Unified observation space to Box or MultiBinary dict for proper flattening.
    • Removed over‑strict constraints from AbstractARCEnv; users only need to override create_operations.
  • Added action wrappers (BBoxWrapper, PointWrapper) to switch selection mode.
  • Loader now supports configurable RNG.
  • Organized examples.
  • Fixed terminated not being emitted when max_trial is infinite (-1).

0.2.5

  • Python 3.8 support.
  • Environment changes:
    • Renamed ARCEnv to RawARCEnv.
    • Removed MiniARCEnv; use RawARCEnv with loader=MiniARCLoader().
    • Added new ARCEnv with action space matching ARC test interface.
    • All environments fully observable; state‑related variables now in current_state dict.
    • Added n‑trial mode via gym.make(); -1 denotes unlimited trials.
    • Customizable Submit action.
  • Fixed boundary issues for flood‑fill, paste, copy.

0.2.2

  • Fixed issue where total reward (all‑or‑nothing) was only given on submission.

0.2.1

  • O2ARCv2Env‑v0:
    • Changed ResizeGrid to CropGrid.
    • Enforced FloodFill to select a single pixel.
    • Treated black pixels as background for CopyI, CopyO, Paste.
    • action[selection] now accepts integers.

0.2.0

  • Renamed ArcEnv‑v0 and MiniArcEnv‑v0 to ARCEnv‑v0 and MiniArcEnv‑v0.
  • Added O2ARCv2Env‑v0.
  • Added O2ARC actions:
    • Uses selection and selected.
    • Supports coloring, flood‑fill, move, rotate, flip, copy, paste, clipboard.

0.1.1

  • Updated imagery.
  • Fixed minor bugs.

0.1.0

  • Initial build.
  • Released ArcEnv‑v0 and MiniArcEnv‑v0.

AI studio

Generate PPTs instantly with Nano Banana Pro.

Generate PPT Now

Access Dataset

Login to Access

Please login to view download links and access full dataset details.

Topics

Abstract Reasoning
Reinforcement Learning

Source

Organization: arXiv

Created: 7/30/2024

Power Your Data Analysis with Premium AI Models

Supporting GPT-5, Claude-4, DeepSeek v3, Gemini and more.

Enjoy a free trial and save 20%+ compared to official pricing.