ARC Prize 2025 — Full(er) Baseline

A readable, dependency-free solver package for ARC Prize 2025 (ARC‑AGI‑2). It learns a variety of human-ish transforms, backs off to short program search, and writes the required submission.json with two attempts per test input.

Works locally with the public ARC‑AGI‑2 repo and on Kaggle (no‑internet, notebook‑only).

Features

• Grid toolkit: rotations, flips, transpose, integer scale up/down, nz‑bbox crop, translation, connected components, border drawing, tiling.
• Learners (exact fits over train pairs):
  • rotation / flip / transpose
  • global color permutation (bijective) across all training pairs
  • same‑shape translation with 0 fill
  • crop‑to‑nonzero bounding box
  • largest/smallest component extraction
  • detect & add uniform border (thickness 1–3)
  • integer scale up/down (nearest‑neighbor up; block‑mode down)
  • small compositions (recolor∘rotation, rotation∘recolor)
• Short program search (length ≤ 2 from a small library) to catch simple multi‑step rules.
• Always emits two predictions per test input, as required by Kaggle evaluation.

Quick start (Kaggle)

from arc25_full import generate_submission
eval_dir = "/kaggle/input/arc-prize-2025/arc-agi_evaluation_challenges.json"  # adjust if needed
generate_submission(eval_dir, out_path="submission.json")

Then Commit → Submit to Competition.

Local dev (with ARC‑AGI‑2)

1. Clone https://github.com/ai-art-dev99/ARC-Prize-2025
2. Run:

python -m arc25_full.solver --help  # or use the API

or via Python:

from arc25_full import generate_submission
generate_submission("/path/to/ARC-AGI-2/data/evaluation", "submission.json", limit=10)

Notes

• This starter is intentionally conservative and fast. Extend by adding richer transforms and a smarter search policy if you want more coverage.
• Keep the notebook offline and within Kaggle efficiency limits.