BOCoDe: Engineering-Centered Benchmarking for Bayesian Optimization

Companion dataset for the paper BOCoDe: Engineering-Centered Benchmarking for Bayesian Optimization and the BOCoDe library (pip install bocode).

BOCoDe is a benchmark of 307 black-box optimization problems — 159 engineering, 80 hyperparameter-optimization (HPO), and 68 synthetic — spanning five optimization classes (single-/multi-objective, unconstrained/constrained, mixed-variable), with 31 reference Bayesian-optimization and evolutionary baselines behind one PyTorch evaluate() API. Docs: https://bocode.readthedocs.io.

Most BOCoDe problems are functions (simulators, analytical models, control policies) and live in the Python package. This dataset hosts the parts of the benchmark that are data:

Configurations

config	rows	what it is
problem_catalog (default)	583	One row per registered BOCoDe problem: name, optimization class, domain, dimension, #objectives, #constraints, variable types, source citation. in_paper_benchmark=True marks the 307 problems benchmarked in the paper. Browse it in the viewer to explore the whole suite.
materials_agnp	3,295	Silver-nanoparticle synthesis experiments (5 inputs → optical score)
materials_autoam	100	Additive-manufacturing print quality (4 inputs → score)
materials_crossed_barrel	1,800	3D-printed crossed-barrel toughness (4 inputs → toughness)
materials_hoip	480	Hybrid organic–inorganic perovskite stability (composition → stability)
materials_p3ht	233	P3HT/CNT thin-film conductivity (5 inputs → conductivity)
materials_perovskite	139	Perovskite composition screening (3 inputs → objective)
lcbench	70,000	LCBench tabular HPO: 35 datasets × 2,000 MLP configs (7 named hyperparameters → validation accuracy)
hpob	1,888,555	HPO-B tabular HPO: 92 (search-space × dataset) tasks; normalized hyperparameters (x, names in variable_names) → accuracy
nasbench201	15,625	NAS-Bench-201 architecture accuracies on CIFAR-10/100 and ImageNet16-120

from datasets import load_dataset

catalog = load_dataset("rosenyu/BOCoDe", "problem_catalog")
agnp    = load_dataset("rosenyu/BOCoDe", "materials_agnp")

To optimize over these (and the other 290+ function-based problems), use the library — every problem, dataset-backed or not, is one API:

import bocode
problem = bocode.get_problem("AgNP")()
x = problem.sample(5, seed=0)
obj, cons = problem.evaluate(x)   # maximization; constraints feasible when <= 0

Large binary assets (MOPTA08/Mazda executables, the SVM dataset) are fetched on demand by the library from rosenyu/bocode-data.

Licensing & provenance

This is a redistribution of community benchmark data, unchanged except for format (CSV/NPZ → Parquet). Per-source licenses and citations:

config	upstream	license
materials_*	PV-Lab Benchmarking (Liang et al., npj Comput. Mater. 2021)	MIT
lcbench	LCBench (Zimmer et al., 2021)	Apache-2.0
hpob	HPO-B (Arango et al., NeurIPS D&B 2021)	upstream terms (CC-BY-style; see repo)
nasbench201	NAS-Bench-201 (Dong & Yang, ICLR 2020)	MIT
problem_catalog	BOCoDe metadata	MIT

If you use a config, please cite its upstream source alongside BOCoDe.

Citation

@article{yu2026bocode,
  title={BOCoDe: Engineering-Centered Benchmarking for Bayesian Optimization},
  author={Yu, Rosen Ting-Ying and Hatterer, Christophe and Narayanan, Advaith and Picard, Cyril and Ahmed, Faez},
  journal={arXiv preprint arXiv:2608.15073},
  year={2026}
}