Sudoku Validation

Synthetic benchmark for evaluating Concept Bottleneck Models on a sudoku validity task. Each example is an image of a 9x9 sudoku board; the binary label indicates whether the board is a valid solved sudoku (1) or has at least one row/column/block conflict (0). The 27 binary concepts each encode the validity of one row, column, or 3x3 block.

Generated from

This dataset is the exact output of the concept-benchmark Python package — seed and config are pinned for bit-identical reproduction.

# pip install concept-benchmark==0.3.1
from concept_benchmark.sudoku import DatasetGenerator
dataset = DatasetGenerator(seed=171, n_boards=1000).generate()

📓 Quickstart notebook — load from the Hub → train a label predictor → run selective classification, end-to-end in one notebook.

🔗 PyPI · GitHub

Quick start (datasets library)

from datasets import load_dataset
ds = load_dataset("juliannski/sudoku")
row = ds["train"][0]
row["input"]       # PIL.Image (sudoku board render)
row["concepts"]    # 27-dim binary list (9 rows + 9 cols + 9 blocks)
row["label"]       # 0 = invalid, 1 = valid

Schema

Field	Type	Notes
input	datasets.Image()	rendered sudoku board PNG
concepts	Sequence(int8, length=27)	binary row/col/block validity; index order below
label	ClassLabel("0", "1")	0 = invalid, 1 = valid
label_name	string	display alias

The 27 concept slots are ordered: 9 row-validity flags, then 9 column-validity flags, then 9 block-validity flags.

Splits

Split	n
train	640
validation	160
test	200

Stratified on label; seed=171. Roughly balanced (~50/50 valid/invalid by construction).

Companion datasets

• juliannski/robots-true-concepts — robot image classification benchmark, 7 concepts.
• juliannski/robots-human-concepts — robot benchmark with 12 finer-grained concepts.

License

MIT.