Minimal ARC

A repository for working with the Abstraction and Reasoning Corpus (ARC) dataset.

Apply to join the Trelis ARC AGI 2 Team here.

Check out Trelis' video explaining ARC here.

Very Basic Examples

If you're new to ARC AGI, read through these very simple examples:

• Domain Specific Language Approach = combine basic operations
• Neural Net Approach = train on examples
• LLM guided program search = write python solvers using LLMs

Then, move to the comprehensive examples in the dsl, llmgs and ttt folders.

Scripts

clean_arc_data.py

This script filters ARC data files to keep only examples with single test inputs/outputs.

Functionality:

• Removes examples with multiple tests or solutions
• Preserves the original data structure
• Outputs statistics about removed examples

Usage:

uv venv
cd arc-data
uv run python clean_arc_data.py

Output:

• Creates an arc-data-cleaned directory
• Saves filtered versions of the original files
• Prints statistics about the cleaning process

Folders

• arc-data: Original ARC data files
• arc-data-cleaned: Cleaned ARC data files
• dsl: Domain Specific Language approach (create basic programs and try to combine them to solve the training examples. Deterministic approach.).
• llmgs: LLM guided search (get an LLM to keep writing programs until one passes on train examples).
• ttt: Test-time training (train a neural net to predict outputs, then add depth first search).