Kanzoet97/Sumo

Sudoku-Bench

📝 [Leaderboard] 📝 [Technical Report] 📝 [Blog Post]
🤗 [Sudoku-Bench puzzle dataset] 🤗 [Sudoku-CTC-Reasoning dataset]

Sudoku-CTC-Reasoning dataset

The Sudoku-CTC-Reasoning dataset contains the reasoning traces of 1351 puzzles featured in the Cracking the Cryptic YouTube channel, and thus provides rich learning signals for training LMs to learn reasoning in a Sudoku game or for a broader range of reasoning-intensive tasks.

This dataset is provided with permission from Cracking the Cryptic.

Data statistics

• Videos with reasoning traces: 1351
• Total actions: 3539008
• Total actions (excluding highlights): 818145
• Total words: 8921707

There will be updates to the dataset with more reasoning traces, so please stay tuned.

Combined ASR and Action Dataset

For each video we provide action_data and asr_data. The action_data is extracted from the youtube video using a video-to-actions pipeline (described below) and consists of a sequence of actions taken in the SudokuPad app as the host solves the puzzle. The asr_data is extracted from the youtube video using Whisper.

Loading the dataset

The dataset has two subsets. We document the raw subset here, and the processed subset is described in the Sudoku-Bench's data_processing README.

import datasets

dataset = datasets.load_dataset('SakanaAI/Sudoku-CTC-Reasoning', 'raw')

Each entry of dataset contains the following fields:

• youtube_id: the id of the youtube video
• action_data: the Sudokupad actions from the video
• asr_data: the audio transcript of the puzzle
• puzzle_id: the id of the puzzle
• puzzle_data: the puzzle data following the format of the Sudoku puzzle datasets.

Actions format

The action_data for each video is a list of action groups. Each action group has the following fields:

• idx: the index of the action group in the video
• frame: the frame number of the corresponding frame in the video
• time: the time in seconds of the action group using the video's time axis
• actions: a list of strings of (serialized actions) taken on the SudokuPad board that occurred between the previous frame and the current frame.
• serialized_state: the serialized state of the SudokuPad board.

Typically each action group contains a single action.

Serialized action format

The serialized action 'cd:+:7:r3c5' denotes "add a candidate 7 to cell r3c5". In general we use the following action types:

Action type:

• vl: value (i.e. the actual proposed value of the cell)
• cd: candidate or center pencilmark
• pm: corner pencilmark
• co: color
• sl: select
• ds: deselect

Operation type:

• +: add to the current list
• -: remove from the current list

Value:

• 1-9: for value, candidates, and pencilmarks
• 0-9: for color, with mapping here.

Coordinates:

• rxcy: row and column of the action. In sl and ds actions, rxcy is a comma-separated list.

Serialized state format

The serialized state can be loaded into SudokuPad by

import requests

# with sudokupad_interaction/app.py running
response = requests.put("http://localhost:8000/set_state", json={"serialized_state": serialized_state})

The format of serialized state follows that used internally by SudokuPad: For each cell, use / to separate value, candidate, pencilmark, color, highlight, pen-tool, respectively, with trailing /s removed.

Example: A serialized_state of {"cells":["6","/1,2,4,5", ... indicates that r1c1 has a value of 6, and r1c2 has candidates (center small digits)1, 2, 4, 5.

ASR format

The asr_data is the output of Whisper ASR using model.transcribe(audio_file, language="en", task="transcribe", word_timestamps=True) for model = whisper.load_model("turbo"). Please see Whisper's documentation for details.

Video-to-actions summary

Extracting sudokupad actions from the video is a multi-step process:

1. Detect the x, y, height, width of the Sudoku board in the video
2. Detect the x, y, height, width of the corresponding Sudoku board in the SudokuPad app with the same puzzle loaded
3. Using 1. and 2. and the location of individual cell rendering in the browser window in SudokuPad (the lines that make up the typically 9x9 grid), determine the corresponding cell locations in the youtube video.
4. Take a sequence of keyframes from the video cropped to the Sudoku board. A keyframe is where a pixel-wise change occured above a threshold. A keyframe is a candidate for when an action was taken.
5. For each keyframe, use a trained resnet classifier to map the pixel input to a multiclass prediction each of the (typically 81) SudokuPad cell states, which includes the colors, pencilmarks (corner small digits), candidates (center small digits), and current value (center large digit). The resnet was trained on synthetic data from the SudokuPad app. The exact images to feed into the resnet are determined by the cell locations from 1. and 2. Combine the individual cell state predictions to obtain a single board state for this keyframe.
6. From the sequence of states, determine the state-diffs to get the sequence of actions, saved as action_data.

The code for the video-to-actions pipeline itself is not open-sourced as part of Sudoku-Bench.

References

• CTC YouTube channel
• CTC catalogue

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}
}