Cube Sudoku Benchmark
Cube Sudoku Benchmark contains 5,000 generated puzzles for testing constraint solvers and machine learning models. Each puzzle spans the six 9×9 faces of a cube. Standard Sudoku rules apply to each face. A cell along an edge belongs to two faces and a corner cell belongs to three. Every puzzle has exactly one solution across the whole cube.
Dataset and splits
| App difficulty | Train | Validation | Test | Total |
|---|---|---|---|---|
| Easy | 800 | 100 | 100 | 1,000 |
| Medium | 800 | 100 | 100 | 1,000 |
| Hard | 800 | 100 | 100 | 1,000 |
| Expert | 800 | 100 | 100 | 1,000 |
| Master | 800 | 100 | 100 | 1,000 |
| Total | 4,000 | 500 | 500 | 5,000 |
The archive contains train.jsonl, validation.jsonl and test.jsonl with one UTF-8 JSON object per line. It also contains topology.json for the shared geometry, manifest.json for dataset metadata and SHA-256 file checksums, and verify.py for validation.
| Field | Meaning |
|---|---|
puzzle_id |
Stable identifier derived from the ordered givens |
app_difficulty |
Requested app tier, from easy to master |
engine_rating |
Hardest technique used by the engine in its logical solution path, such as hidden_single or x_wing |
givens_count |
Number of clues in the initial state |
givens |
A 386-character digit string with 0 for blanks |
solution |
A 386-character digit string containing the solution |
Read the strings using the cell order in topology.json. For example, face_grids["x_neg"][0][0] gives the index of the top-left cell on the negative-X face. The same index refers to the same cell wherever it appears. The topology also provides each cell's 3D coordinates and the 150 distinct rows, columns, and boxes. Each of these 150 groups must contain the digits 1–9 exactly once.
Generation and validation
Every puzzle in the benchmark has a single valid solution and is graded by the hardest logical technique used in its reference solve path. Before inclusion, each puzzle was checked for solution validity, uniqueness, and difficulty rating.
Every released solution was checked against the givens and all 150 constraints. For each puzzle, a separate Z3 model was asked to find a solution that differed from the supplied solution in at least one cell. All 5,000 checks returned unsatisfiable. Run python3 verify.py to verify checksums, schema, constraints, split assignments, and solution families. With z3-solver installed in your Python environment, run python3 verify.py --z3 to repeat the independent uniqueness checks.
To avoid sharing closely related solutions across splits, we canonicalised every complete solution across the cube's 48 rotations and reflections. Within each orientation we renamed digits in order of first occurrence and then chose the lexicographically smallest result. The release contains 5,000 distinct canonical solutions. Within each app tier, we sort these solution groups by a versioned SHA-256 key and take the first 800 for training, the next 100 for validation, and the final 100 for testing.
The app tiers can overlap in their use of individual techniques. The engine rating describes a specific solve path, and the dataset contains no ratings from human solvers. The manifest lists the observed distribution of engine ratings within every app tier.
Evaluating solvers
Use the givens and public topology as input. Keep the solution for scoring. Report the percentage of completely solved cubes overall and for each app tier. For timed comparisons, record the solver version, hardware, time limit, and timeout count.
The fixed test split has public solutions for reproducibility. Use a fresh held-out set for models that may already have seen this release. The instances reflect a single generator and rule set, so results apply to this particular distribution of cube puzzles.
Citation
Patrick Schibli / Dream Factory (2026). Cube Sudoku Benchmark, version 1.0.0. https://www.dream-factory.app/cubesudoku/en/datasets
Licensed under CC BY 4.0. When sharing or adapting this dataset, please credit Patrick Schibli / Dream Factory and link to https://www.dream-factory.app/cubesudoku/en/datasets.
For prior research on Sudoku on a cube, see Lambert and Whitlock, Generalizing Sudoku to three dimensions (2010). The game that provides the engine is Cube Sudoku for iPhone and iPad.