Sudoku Validation

Synthetic benchmark for evaluating Concept Bottleneck Models on sudoku validation. Each example is an image of a 9x9 board; the label says whether the board is valid. The 27 binary concepts encode the validity of each row, column and block.

Generated from

This dataset is the exact output of the concept-benchmark Python package, with the train/validation/test split of the paper: 1,000 boards at 50 px per cell (454 x 454 px), each written in its own handwriting, with printed starters and candidate marks.

# pip install concept-benchmark==0.4.1
from concept_benchmark.sudoku import DatasetGenerator
dataset = DatasetGenerator(seed=171).generate_splits()

The images were regenerated with 0.4.1, which gives every board its own layout of printed digits; in the 0.4.0 images (revision e0a31cfb950230abcc253a545ac260514d1ca2c7) every valid board shared one layout, so an image model could tell valid from invalid boards without reading a digit. Digits, concepts, labels and splits are the same in both.

📓 Quickstart notebook — load from the Hub and train a model, 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
row["concepts"]    # 27 binary values, ordered as concept_names below
row["label"]       # 0 or 1
row["label_name"]  # "invalid" or "valid"

Schema

Field	Type	Notes
input	datasets.Image()	PNG, embedded
concepts	Sequence(int8, length=27)	binary; index order = concept_names below
label	ClassLabel("invalid", "valid")
label_name	string	display alias for label

concept_names (column order in concepts):

['row_valid_1', 'row_valid_2', 'row_valid_3', 'row_valid_4', 'row_valid_5', 'row_valid_6', 'row_valid_7', 'row_valid_8', 'row_valid_9', 'col_valid_1', 'col_valid_2', 'col_valid_3', 'col_valid_4', 'col_valid_5', 'col_valid_6', 'col_valid_7', 'col_valid_8', 'col_valid_9', 'block_valid_1', 'block_valid_2', 'block_valid_3', 'block_valid_4', 'block_valid_5', 'block_valid_6', 'block_valid_7', 'block_valid_8', 'block_valid_9']

Splits

Split	n
train	600
validation	200
test	200

Seed 171. 50% of the training examples are valid.

License

MIT.