sudoku-extreme

Sudoku-Extreme consists of 1,000 base puzzles augmented 1,000\u00d7 via symmetry transformations (rotations, reflections, block permutations), yielding 1,000,000 training examples. Each puzzle is a 9\u00d79 grid represented as a sequence of 81 tokens over vocabulary size 10, where 0 denotes empty cells and digits 1-9 represent clues. The dataset is split 80\/20 into 800,000 training and 200,000 validation puzzles. All puzzles are rated \extreme\ difficulty, requiring advanced solving techniques beyond basic constraint propagation. The test set contains 423,168 puzzles generated from different base puzzles to ensure distribution shift. Input puzzles typically provide 20-25 initial clues, requiring logical deduction across interdependent row, column, and 3\u00d73 block constraints. This dataset challenges models to perform multi-step reasoning and constraint satisfaction, making it suitable for evaluating recursive reasoning architectures.