Sudoku-Bench

🤗 [Sudoku-CTC-Reasoning dataset]
🐙 [Sudoku-Bench GitHub]
📝 [Blog Post]

Sudoku-Bench puzzle dataset

The SakanaAI/Sudoku-Bench puzzle dataset contains three subsets:

• challenge_100: A collection of 100 creative Sudoku puzzles.
  • test split: 100 puzzles
• nikoli_100: A collection of 100 beautiful handmade standard Sudoku puzzles designed by Nikoli.
  • test split: 100 puzzles
• ctc: A larger collection of puzzles featured as puzzles solves in the Cracking the Cryptic (CTC) YouTube channel.
  • test split: 2565 puzzles

Subset details

challenge_100 subset

The purpose of the challenge_100 subset is to evaluate the reasoning capabilities of LLMs on a diverse set of Sudokus.

The subset includes

• 15 4×4 puzzles (Sudoku variants)
• 15 6×6 puzzles (Sudoku variants)
• 50 9×9 puzzles (Sudoku variants)
• 20 9×9 puzzles (standard Sudoku) taken from the nikoli_100 set

The selection of puzzles covers a range of difficulty. The 9×9 puzzles are roughly evenly distributed across difficulty levels 1 through 5 (using the Logic Masters difficulty scale). Around 5 puzzles are more difficult than the standard 5-star difficulty and are considered a challenge to the best human solvers. Difficulty is not a reflection of how complex the puzzle appears, and is not necessarily related to the length of the ruleset. Difficulty is a measure of how much skill and time is required for a human solver and is more closely a reflection of the depth of the idea required to find the puzzle's break-in.

The 4×4 puzzles are significantly easier and most are rated 1-star difficulty. A subset of the 4×4 puzzles are quite simple and predominately test the model's ability to understand the constraints of Sudoku variant.

Taken as a whole, the challenge_100 includes a broad spectrum of difficulty and can be used to evaluate the performance of reasoning models of varying capabilities.

nikoli_100 subset

The nikoli_100 subset contains 100 beautiful handmade standard Sudoku puzzles designed by Nikoli, the Japanese puzzle company that popularized Sudoku.

Algorithimcally generated Sudoku puzzles tend to find only puzzles of a certain type, namely whose solution path is similarly algorithmic. Human setters are more capable of creating puzzles that require deeper reasoning and creativity in the solve: see this video, for an example.

ctc subset

The ctc subset contains 2565 puzzles featured as puzzles solves in the Cracking the Cryptic channel. The ctc subset can be used in conjunction with the reasoning traces in huggingface.co/datasets/SakanaAI/Sudoku-CTC-Reasoning. That is, you may wish to use the reasoning traces together with prompts derived from the content of the puzzle being solved, which the ctc subset can provide.

Puzzle details

Each puzzle in SakanaAI/Sudoku-Bench contains the fields:

Puzzle data

• puzzle_id: Identifier for the puzzle
• sudokupad_url: Link to play the puzzle on Sudokupad
• author: Creator of the puzzle
• title: Name of the puzzle
• rules: The puzzle rules
• initial_board: String representation of the starting grid (empty cells shown as '.')
• solution: String representation of the completed grid (81 digits for a 9×9 puzzle)
• rows: Number of rows in the puzzle
• cols: Number of columns in the puzzle
• visual_elements: JSON-encoded string containing detailed specifications for visual components like circles, lines, and other custom markings specific to the puzzle variant (see Sudoku-Bench/src/sudokupad_interaction/puzzle_tools for the extraction of the visual elements)
• encoded_puzzle: A compressed representation of the puzzle using SudokuPad's encoding scheme; for loading the puzzle directly in an offline SudokuPad (see Sudoku-Bench/src/sudokupad_interaction/README.md)

The puzzles from the ctc subset contain additional fields:

Video metadata

• youtube_id: The YouTube ID of the video from which the puzzle was solved
• sequential_number: The index of the puzzle in the video (for videos where multiple puzzles are solved; in most cases this is 1)
• date: The upload date of the video
• lgc_timestamp: The time in seconds when the phrase "let's get cracking" is said indicating the start of the solve in the video

Example puzzle: Parity Fish

The puzzle Parity Fish by Marty Sears is included in the challenge_100 dataset.

• puzzle_id: 'sxsm_MartySears_580c6fdbbba9bfb0e71ae19044f02d4c' (using SudokuPad's internal id field)
• sudokupad_url: 'https://sudokupad.app/wsj7iunsg6' (link to the puzzle on SudokuPad)
• author: 'Marty Sears'
• title: 'Parity Fish'
• rules: 'Normal sudoku rules apply; fill the grid with the digits 1-9 so that digits don\'t repeat in any row, column, and marked 3x3 box.\\nTwo cells adjacent along a red line must contain one even digit and one odd digit.\\nTwo cells connected by a white dot contain consecutive digits.\\nTwo cells connected by a black dot contain digits where one is double the other.',
• initial_board: '.................................................................................' (empty cells are represented as .)
• solution: '854369172976251834123478956419582367568937421237146598785694213691823745342715689'
• rows: 9
• cols: 9

Visual elements

The visual_elements field is a JSON-encoded string containing detailed specifications for visual components of the puzzle. In the Parity Fish puzzle, there are 24 visual elements: 5 black dots, 16 white dots, and 3 red lines. You can display the visual elements using the pretty_print_visual_elements function in src/eval.utils in the SakanaAI/Sudoku-Bench repo

import datasets
import json
from eval.utils import pretty_print_visual_elements

puzzle = datasets.load_dataset("SakanaAI/Sudoku-Bench", "challenge_100")['test'][23]  # Parity Fish puzzle
print(pretty_print_visual_elements(json.loads(puzzle['visual_elements'])))
# - shape: circle, color: white (stroke color: black), location: between r4c8 and r4c9
# - shape: circle, color: white (stroke color: black), location: between r5c8 and r5c9
# - shape: circle, color: white (stroke color: black), location: between r6c8 and r6c9
# - shape: circle, color: white (stroke color: black), location: between r5c1 and r5c2
# - shape: circle, color: white (stroke color: black), location: between r8c3 and r9c3
# - shape: circle, color: white (stroke color: black), location: between r7c1 and r8c1
# - shape: circle, color: white (stroke color: black), location: between r1c1 and r2c1
# - shape: circle, color: white (stroke color: black), location: between r7c7 and r7c8
# - shape: circle, color: white (stroke color: black), location: between r7c1 and r7c2
# - shape: circle, color: white (stroke color: black), location: between r9c8 and r9c9
# - shape: circle, color: white (stroke color: black), location: between r8c5 and r8c6
# - shape: circle, color: white (stroke color: black), location: between r1c4 and r2c4
# - shape: circle, color: white (stroke color: black), location: between r7c6 and r8c6
# - shape: circle, color: white (stroke color: black), location: between r2c7 and r3c7
# - shape: circle, color: white (stroke color: black), location: between r1c2 and r1c3
# - shape: circle, color: white (stroke color: black), location: between r1c5 and r2c5
# - shape: circle, color: black, location: between r3c2 and r4c2
# - shape: circle, color: black, location: between r4c7 and r4c8
# - shape: circle, color: black, location: between r2c3 and r3c3
# - shape: circle, color: black, location: between r9c2 and r9c3
# - shape: circle, color: black, location: between r8c8 and r9c8
# - line, color: red, coords: r3c2, r3c3, r3c4, r3c5, r3c6, r4c7, r5c8, r6c7, r7c6, r7c5, r7c4, r7c3, r7c2
# - line, color: red, coords: r4c1, r4c2, r5c3, r6c4, r7c4
# - line, color: red, coords: r6c1, r6c2, r5c3, r4c4, r3c5

The intermediate json.loads(puzzle['visual_elements'])) is a list of dictionaries, each of which is a verbose description extracted from the SudokuPad rendering engine. We encourage the user to adopt their own pretty_print_visual_elements function to display the visual elements in a way that is most useful for their application.

Please see src/sudokupad_interaction/puzzle_tools for more details on the visual_elements field.

Encoded puzzle

The encoded_puzzle field is a byte64 encoding of the puzzle using SudokuPad's internal encoding method.

The encoded_puzzle field can be used to obtain an alternate URL for the puzzle. Namely, https://sudokupad.app/wsj7iunsg6 and https://sudokupad.app/{parity fish's encoded_puzzle string} will load the same puzzle. Both URLs point to Sven's SudokuPad website. However, only the second method works when running SudokuPad locally and avoids a call to the SudokuPad puzzle database. To ensure longevity of the benchmark, we provide a local usage in src/sudokupad_interaction.

The encoded_puzzle field can be ignored if using the text-only approach outlined in src.eval in this repo as all relevant information has already been extracted.

Puzzle edge cases

Because of the wide array of puzzles solved, the ctc subset is provided "as-is". There are a number of edge cases that make a pure text representation of the puzzle incomplete:

1. Some puzzles have visual elements that are difficult to encode in the visual_elements field (see below for a description of the visual_elements field). For example, the delightful RatRun puzzles will not have a coherent textual description of the visual elements due to the visual complexity of the puzzle.
2. Other puzzles have the solution field omitted as many puzzle setters choose not to disclose the solution in SudokuPad.
3. A popular recent trend is the use of fog-of-war in Sudoku puzzles. For such puzzles, all hidden elements will be exposed in the visual_elements field meaning the puzzle will not be presented as intended by the puzzle setter.

Please consider filtering the ctc subset based on your needs.

Citation

@misc{seely2025sudoku-bench,
  title={{Sudoku-Bench}},
  author={Seely, Jeffrey and Imajuku, Yuki and Zhao, Tianyu and Cetin, Edoardo and Jones, Llion},
  howpublished = {\url{https://github.com/SakanaAI/Sudoku-Bench}},
  year={2025}
}