gemma/diffusion/hackable_diffusion_adapter/data/sudoku/README.md

Sudoku Onboarding and Evaluation on HD Text Diffusion

This folder contains the dataset ingestion and preprocessing for the Sudoku puzzle solving dataset.

Sudoku solving is a structured conditional generation task where the model reads a space-separated 81-character puzzle representation and learns to:

1. Preserve the unmasked "clue" digits provided in the initial puzzle context.
2. Iteratively denoise and fill in the missing cells (marked by 0).

Directory Structure

• prepare_sudoku_dataset.sh: Script that configures Kaggle credentials and fetches the dataset.
• convert_sudoku.py: Preprocessing script to format CSV raw datasets into Bagz files.
• sudoku_data.py: Dataset and transform configuration using standard Kauldron pipelines.

Expected output

• sudoku_train.bagz: Converted training records containing the source puzzles and reference solutions.
• sudoku_eval.bagz: Converted evaluation records.