arcagi2kaggle

Open-source campaign for the ARC Prize 2026 — ARC-AGI-2 Kaggle competition.

Read the public research site: ARC/RESEARCH.

Approach: the public NVARC-lineage test-time-training baseline (Qwen3-4B, per-task LoRA fine-tuning, DFS decoding, augmentation-consistency selection) extended with neurosymbolic components — falsification-based symbolic predictors that constrain decoding, verified program induction as a diverse second attempt, and improved candidate selection.

Layout

• arc/ — evaluation harness: task loading, submission validation, official-metric scoring, deterministic 5-fold splits (fold 4 is a shadow fold, held out between milestones).
• symbolic/ — falsification-first symbolic predictors (output size, palette) used to prune and verify neural candidates. Rules only fire when consistent with every demonstration pair.
• kaggle_notebook/ — the submission notebook, its kernel metadata, baseline analysis, and launch runbook.
• scripts/ — utilities, including the secure Kaggle data downloader (python scripts/secure_kaggle_download.py; interactive, token never touches disk).
• data/ (gitignored) — official competition data. Not redistributed, per competition rules; fetch it with the download script.
• .tools/ (gitignored) — vendored Python packages (Kaggle CLI etc.), used via PYTHONPATH.
• sources/ — read-only synced reference material.

Conventions

• Credentials never touch disk: no kaggle.json, no tokens in files or code. Kaggle auth is passed via KAGGLE_API_TOKEN in process environment only.
• Experiments are measured on the 120-task evaluation set locally before any leaderboard submission. One scored submission per day — every one is a controlled measurement.

License

MIT-0 (see LICENSE). Competition data is Apache 2.0, © ARC Prize Foundation, and is not included in this repository.