Hi everyone,

I'm excited to share my ARC-AGI Starter Notebook for the ARC-AGI competition, which implements a zero-shot learning approach using the Qwen2.5-1.5B-Instruct. This notebook serves as a baseline solution, processing input grids, generating predictions, and creating submission files in both JSON and CSV formats. It includes utility functions for grid-to-text conversion, visualization with Matplotlib, and robust error handling with fallback strategies.

Key features:

• Zero-Shot Learning: Leverages Qwen2 for pattern analysis without relying on training data.
• Grid Processing: Converts 2D grids to text for model input and back to grids for output.
• Visualization: Displays input and predicted grids side-by-side for better understanding.
• Flexible Output: Generates submissions matching the competition's required format.

I’d love to hear your thoughts, suggestions, or any improvements you might recommend! Has anyone else experimented with zero-shot approaches for ARC-AGI? Let’s discuss how we can enhance this baseline or tackle specific challenges in the dataset.

Looking forward to your feedback!

And don't forget to star the repo if you find it helpful.

You can find the notebook link👇👇.

https://github.com/mubrij/My-ARC-AGI-Starter-Notebook-with-Zero-Shot-Learning-Approach/blob/main/arc_agi_starter%20(1).ipynb