CellARC

Detailed Introduction

CellARC is an open-source toolkit for dataset generation and loading of ARC-style cellular-automaton episodes. The project publishes dataset snapshots on the Hugging Face Hub (for example mireklzicar/cellarc_100k) and provides convenient APIs such as EpisodeDataset and EpisodeDataLoader to download, cache, and batch episodes. Built-in visualization helpers make it easy to replay CA rollouts and inspect episode cards.

Main Features

• Dataset snapshots: ready-to-use 100k dataset with fixed 100-episode subsets for fast iteration.
• Simulation & visualization: CA rollouts and episode card rendering for debugging and analysis.
• Optional generation stack: install cellarc[all] to enable JAX/Flax/CAX-based generation and advanced sampling tools.

Use Cases

• ML research: benchmark tasks for models on structured reasoning and CA dynamics.
• Teaching and reproducibility: classroom examples and baseline experiments with easy dataset access.
• Data analysis: tools for studying rule-space coverage, episode difficulty, and dataset statistics.

Technical Features

• Lightweight Python API: EpisodeDataset.from_huggingface and EpisodeDataLoader support on-demand downloads and caching for integration with training loops.
• Flexible storage: JSONL and Parquet artifacts with data_files.json and dataset_stats.json for quick split enumeration.
• Packaging & compatibility: published as a PyPI package; full generation/simulation features require Python 3.11+ and extra dependencies.