zeyuzy/puzzle-bench

Puzzle Bench

Difficulty-labeled evaluation datasets for Sudoku and Maze tasks, designed for benchmarking language models on combinatorial reasoning.

GitHub: zeyuzhangzyz/puzzle-bench

Dataset Overview

Config	Total	Train	Test	Difficulty Labels
sudoku_4x4	10,000	8,000	2,000	--
sudoku_9x9	52,806	42,244	10,562	easy / medium / hard
maze_5x5	10,000	8,000	2,000	--
maze_7x7	10,000	8,000	2,000	--
maze_10x10	10,000	8,000	2,000	--
maze_15x15	30,000	24,000	6,000	easy / medium / hard

Sudoku

sudoku_4x4

4x4 Sudoku puzzles generated via backtracking with unique-solution verification.

Column	Description
puzzle	16-character string (0 = empty cell)
solution	16-character solution
empty_count	Number of blank cells
source	Generator identifier

sudoku_9x9

9x9 Sudoku puzzles with solver-computed difficulty metrics. Mixed from multiple sources for balanced difficulty distribution.

Column	Description
puzzle	81-character string (0 = empty cell)
solution	81-character solution
empty_count	Number of blank cells
steps_count	Solver step count (MRV + backtracking)
backtrack_count	Number of backtracks
max_depth	Maximum recursion depth
difficulty	easy / medium / hard
source	Source dataset identifier

Difficulty	Count	Criterion
easy	29,842	0 backtracks (pure logic)
medium	10,000	1-1,000 backtracks
hard	12,964	1,000+ backtracks

Maze

Mazes encoded as binary wall strings with BFS-computed path metrics. Algorithms used: dfs, wilson, prim, kruskal, rdiv.

Column	Description
maze	Binary string encoding walls
start	Start coordinates row,col
goal	Goal coordinates row,col
grid_size	Grid dimension (5/7/10/15)
algorithm	Generation algorithm
solution_length	BFS shortest path length
bfs_nodes	BFS nodes expanded
source	Generator identifier
difficulty	easy / medium / hard (maze_15x15 only, by solution_length tercile)

maze_15x15 difficulty breakdown

Difficulty	Train	Test	Total
easy	8,000	2,000	10,000
medium	8,000	2,000	10,000
hard	8,000	2,000	10,000

Usage

from datasets import load_dataset

# 4x4 Sudoku
sudoku_4x4 = load_dataset("zeyuzy/puzzle-bench", "sudoku_4x4")

# 9x9 Sudoku, hard difficulty only
sudoku_9x9 = load_dataset("zeyuzy/puzzle-bench", "sudoku_9x9")
hard = sudoku_9x9["test"].filter(lambda x: x["difficulty"] == "hard")

# Maze with train/test split
maze_5x5 = load_dataset("zeyuzy/puzzle-bench", "maze_5x5")

# Maze 15x15 with difficulty labels
maze_15x15 = load_dataset("zeyuzy/puzzle-bench", "maze_15x15")
hard_maze = maze_15x15["test"].filter(lambda x: x["difficulty"] == "hard")

Citation

@software{puzzle-bench,
  title={Puzzle Bench},
  author={Zhang, Zeyu},
  url={https://github.com/zeyuzhangzyz/puzzle-bench},
  year={2026}
}