uq-bench Real World Datasets

Twenty real-world tabular regression datasets for benchmarking uncertainty quantification methods, one Parquet file per dataset. Each subset is one table of feature columns plus a __target__ column. Features are either binary with dtype int64 or continuous with dtype double.

Datasets

name	n	d	description	source	license
energy_efficiency	768	10	Building energy load	UCI 242	CC BY 4.0
concrete	1030	8	Concrete compressive strength	UCI 165	CC BY 4.0
insurance	1338	8	Medical insurance charges	pycaret	public domain
airfoil	1503	5	Airfoil self-noise level	UCI 291	CC BY 4.0
debutanizer	2394	7	Debutanizer butane concentration	Fortuna 2007	no license
allstate	5000	1037	Insurance claim severity	OpenML 42571	public
qsar	5742	500	Compound bioactivity	OpenML 3050	CC BY-SA 3.0
parkinsons	5875	18	Parkinsons symptom score	UCI 189	CC BY 4.0
computer	8192	21	Computer system activity	OpenML 197	public
kin8nm	8192	8	Simulated robot arm kinematics	OpenML 189	public
powerplant	9568	4	Power plant energy output	UCI 294	CC BY 4.0
sulfur	10081	5	Sulfur recovery output	OpenML 23515	no license
naval_propulsion	11934	24	Naval gas turbine decay	UCI 316	CC BY 4.0
ailerons	13750	40	Aircraft ailerons control	OpenML 296	public
miami_housing	13932	28	Miami house sale price	OpenML 43093	public
elevator	16599	18	Aircraft elevator control	OpenML 216	public
ca_housing	20640	8	California housing value	StatLib	public domain
superconductor	21263	79	Superconductor critical temperature	UCI 464	CC BY 4.0
protein	45730	9	Protein structure RMSD	UCI 265	CC BY 4.0
diamond	53940	23	Diamond price	OpenML 42225	public

Provenance and processing

Packaged from the CLEAR and PCS-UQ benchmarks of the Bin Yu group at UC Berkeley. Processing is inherited from those pipelines: categorical features are one-hot encoded with drop_first; allstate is a seeded 5000-row subsample with about 310 all-zero dummy columns; qsar keeps the top 500 fingerprint bits by variance; debutanizer features are pre-scaled to [0, 1]. Values are otherwise unscaled and targets are raw.

Licensing

Datasets are aggregated under several licenses; per-dataset licenses are in the table. qsar is CC BY-SA 3.0 from ChEMBL, so reuse inherits the share-alike term. sulfur and debutanizer have no formal data license; they are companion data from Fortuna et al., "Soft Sensors for Monitoring and Control of Industrial Processes", redistributed for academic research with attribution. allstate originates from a Kaggle competition and is listed Public on OpenML.

Citation

@misc{uqbench2026,
  title  = {uq-bench},
  author = {Anzalone, S.},
  year   = {2026},
  note   = {arXiv TODO}
}

@article{agarwal2025pcsuq,
  title   = {PCS-UQ: Uncertainty Quantification via the Predictability-Computability-Stability Framework},
  author  = {Agarwal, A. and Xiao, M. and Barter, R. and Ronen, O. and Fan, B. and Yu, B.},
  year    = {2025},
  journal = {arXiv preprint arXiv:2505.08784}
}

@inproceedings{azizi2026clear,
  title     = {CLEAR: Calibrated Learning for Epistemic and Aleatoric Risk},
  author    = {Azizi, I. and Bodik, J. and Heiss, J. and Yu, B.},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026}
}