---
license: apache-2.0
source_datasets:
- original
pretty_name: SudokuBench
dataset_info:
  features:
    - name: starting_cells
      dtype: int32
    - name: puzzle
      dtype: string
    - name: puzzle_pretty
      dtype: string
    - name: solution
      dtype: string
    - name: solution_pretty
      dtype: string
configs:
- config_name: eval
  default: true
  data_files:
    - split: train
      path:
        - "eval.parquet"
- config_name: kids
  data_files:
    - split: train
      path:
        - "kids.parquet"
- config_name: easy
  data_files:
    - split: train
      path:
        - "easy.parquet"
- config_name: medium
  data_files:
    - split: train
      path:
        - "medium.parquet"
- config_name: hard
  data_files:
    - split: train
      path:
        - "hard.parquet"
- config_name: insane
  data_files:
    - split: train
      path:
        - "insane.parquet"
- config_name: all
  data_files:
    - split: train
      path:
        - "kids.parquet"
        - "easy.parquet"
        - "medium.parquet"
        - "hard.parquet"
        - "insane.parquet"
---

# Dataset Card for SudokuBench

## Dataset Details

This dataset contains a list of sudoku puzzles and their solutions, all at
varying levels of difficulty.

The difficulties are based on the number of squares (also sometimes referred to
as cells) that are provided at the start of the puzzle.

The puzzles are guaranteed to have a single unique solution without any overlap.

Within a difficulty config, you will find `10,000` puzzles at every number of
available cells at the start of the board.

This means that within the `kids` category, you will find 10,000 sudoku boards
that have at least 63 filled squares. You will find 10,000 boards with 64 filled
squares, etc all the way up to having 10,000 boards with 80 filled squares.

I hope that this granularity provides for a clear understanding of where models
start to have problems.

- **Curated by:** Aaron Batilo

## Uses

### Direct Use

The intended use for SudokuBench is to be able to evaluate language models on
their ability to handle long context reasoning tasks.

## Dataset Structure

All puzzles have the following parquet format:

| starting_cells | puzzle | puzzle_pretty | solution | solution_pretty |
|----------------|--------|---------------|----------|-----------------|
| int            | str    | str           | str      | str             |

- **starting_cells**: How many cells are already filled (integer).
- **puzzle**: The puzzle string (compact format).
- **puzzle_pretty**: The puzzle string in a human-readable pretty format.
- **solution**: The solution string (compact format).
- **solution_pretty**: The solution string in a human-readable pretty format.

### Configs

The dataset is organized into multiple parquet files grouped by difficulty thresholds, each represented as a separate config:

| Config name   | Minimum clues | Number of examples | Description                                |
|---------------|---------------|--------------------|--------------------------------------------|
| `kids`        | 63            | 180000             | Very easy puzzles suitable for kids        |
| `easy`        | 45            | 180000             | Easy puzzles                               |
| `medium`      | 36            | 90000              | Medium difficulty puzzles                  |
| `hard`        | 27            | 90000              | Hard puzzles                               |
| `insane`      | 17            | 100000             | Insane difficulty puzzles                  |

Lastly, there's `eval`, which contains the first 200 puzzles of every single
difficulty from 80 already filled squares to 17 already filled squares. This is
sampling of a smaller number of puzzles is much more manageable for running
holistic evals, compared to running 10,000 attempts on every single difficulty
level.