datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
sudoku
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… See the full description on the dataset page: https://huggingface.co/datasets/juliannski/sudoku.robots-human-concepts
Robots — Human Concepts
Synthetic benchmark for evaluating Concept Bottleneck Models (CBMs) under finer-grained, human-annotated concepts. Same underlying robot images and labels as juliannski/robots-true-concepts, but the foot_shape ground-truth concept is replaced by 6 one-hot subtypes that a human annotator would actually see, modelling concept specification mismatch between annotators and the latent labeling rule.
Generated from
This dataset is the exact… See the full description on the dataset page: https://huggingface.co/datasets/juliannski/robots-human-concepts.robots-true-concepts
Robots — True Concepts
Synthetic benchmark for evaluating Concept Bottleneck Models (CBMs). Robot images are generated deterministically with pycairo; binary labels follow a known disjunction-style rule over the 7 ground-truth concepts.
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.robots import… See the full description on the dataset page: https://huggingface.co/datasets/juliannski/robots-true-concepts.