ExaTune Landscapes is a collection of exhaustive hyperparameter search results produced by running every configuration in a dense Cartesian-product grid for seven dataset–model-family pairs. Each row corresponds to one evaluated configuration and records the cross-validated performance metric(s), per-fold timing, and (where applicable) supplementary metrics. Together the seven landscapes span 301,516 fully evaluated configurations across five scikit-learn/XGBoost model families and six publicly available tabular datasets, making this the largest openly available collection of complete hyperparameter fitness landscapes for tabular ML.

The primary intended use is empirical analysis of hyperparameter landscape topology: spatial autocorrelation (Moran's I), ruggedness, local optima detection, fitness-distance correlation, parameter importance, and pairwise interaction strength. The exhaustive enumeration eliminates sampling bias that confounds landscape analyses derived from random or Bayesian search trajectories.

Curated by: Kevin Dick and Katie Lonsway, Children's Hospital of Eastern Ontario Research Institute (CHEO-RI) / BORN Ontario
Funded by: N/A
License: MIT

Dataset Sources

Repository: https://huggingface.co/datasets/Dick-AI-Lab/ExaTune_landscapes
Paper: Dick, K. et al. ExaTune: Exhaustive Hyperparameter Landscape Exploration at Scale. NeurIPS 2025 (under review).
Framework: https://github.com/Dick-AI-Lab/ExaTune

Uses

Direct Use

Landscape topology analysis: Compute Moran's I, ruggedness index, local optima count, fitness-distance correlation, and funnel index on any of the seven landscapes using the companion exatune Python library.
Hyperparameter importance benchmarking: Evaluate variance-decomposition (η²), correlation-based, or range-based importance methods against a ground-truth exhaustive baseline.
Algorithm comparison: Benchmark hyperparameter optimisation algorithms (random search, Bayesian optimisation, evolutionary methods) by replaying their trajectories against the pre-computed landscape rather than re-running expensive model training.
Meta-learning: Derive landscape features as meta-features for algorithm selection or warm-starting.
Cost–performance analysis: Study the relationship between computational cost (node-hours) and model performance across grid size and model family.

Out-of-Scope Use

The landscapes should not be used as training data for general tabular prediction tasks; the rows describe hyperparameter configurations, not independent data instances.
Landscapes with a narrow performance range (e.g., Diabetes Ridge Regression R² ∈ [0.25, 0.28]) may not be informative for importance or interaction studies; landscape metrics should be interpreted in context of the range-to-noise ratio.
Causal claims about which hyperparameter causes performance differences cannot be made from observational grid data alone.

Dataset Structure

The dataset contains seven Parquet files, one per experiment, stored under ExaTune/. Each file is named {Dataset}_{ModelFamily}_{ApproxGridSize}.parquet.

Files and Summary Statistics

File	Task	Model Family	Grid \|Λ_G\|	CV Folds	Score Range	Compute
`California-Housing_Gradient-Boosting_58K.parquet`	Regression	Gradient Boosting	51,478	5	R² [0.213, 0.848]	165 node-h
`Diabetes_Ridge-Regression_8k.parquet`	Regression	Ridge Regression	4,158	5	R² [0.252, 0.279]	<0.1 node-h
`Gestational_diabetes_XGBoost_4k.parquet`	Binary Classification	XGBoost	4,680	5	Acc [0.962, 0.974]	<0.1 node-h
`Maternal_Risk_XGBoost_4k.parquet`	Multiclass Classification	XGBoost	4,680	5	Acc [0.697, 0.869]	<0.1 node-h
`Titanic_XGBoost_100k.parquet`	Binary Classification	XGBoost	112,320	5	Acc [0.739, 0.883]	1.5 node-h
`UCI-Adult_random-forest_92K.parquet`	Binary Classification	Random Forest	92,160	5	Acc [0.808, 0.859]	25 node-h
`Wine_KNN_32k.parquet`	Multiclass Classification	K-Nearest Neighbours	32,040	5	Acc [0.629, 0.781]	<0.1 node-h

Total: 301,516 evaluated configurations.

Schema (all files)

Column	Type	Description
`job_id`	string	Unique SLURM job identifier
`config_hash`	string	SHA256 of the hyperparameter configuration dict
`timestamp`	datetime	UTC timestamp of job completion
`success`	bool	Whether the configuration evaluated without error (all 100%)
`cv_scores`	list[float]	Per-fold held-out scores (length = CV folds)
`mean_score`	float	Mean of `cv_scores` (primary optimisation metric)
`std_score`	float	Standard deviation of `cv_scores`
`mean_train_score`	float	Mean in-fold training score
`fit_time_mean`	float	Mean wall-clock fit time per fold (seconds)
`score_time_mean`	float	Mean wall-clock score time per fold (seconds)
`hyperparameters.*`	numeric/string	One column per hyperparameter (see below)
`additional_metrics.*`	float	Supplementary metrics (mean and std; see below)

Hyperparameter Grids

California Housing — Gradient Boosting n_estimators ∈ {25, 50, …, 300} (12 levels) × max_depth ∈ {3, …, 15} (13 levels) × learning_rate ∈ {0.01, 0.02, …, 0.30} (30 levels) × subsample ∈ {0.5, 0.55, …, 1.0} (11 levels)

Diabetes — Ridge Regression alpha ∈ {0.15, 0.30, …, 9.90} (66 levels) × max_iter ∈ {100, 300, 500} × solver ∈ {auto, cholesky, lsqr, sag, saga, sparse_cg, svd} × tol ∈ {1e-4, 1e-3, 1e-2}

Gestational Diabetes / Maternal Risk — XGBoost (4k grids) learning_rate (30 levels) × max_depth (13 levels) × n_estimators (12 levels); colsample_bytree, gamma, min_child_weight, subsample held fixed.

Titanic — XGBoost (100k grid) n_estimators (12) × max_depth (13) × learning_rate (30) × subsample ∈ {0.7, 0.9} × colsample_bytree ∈ {0.7, 0.9} × min_child_weight ∈ {1, 3, 5, 7, 9, 11}; gamma fixed at 0.

UCI Adult — Random Forest n_estimators (12) × max_depth (15 levels: 3–17) × min_samples_split (16 levels: 2–17) × min_samples_leaf (16 levels: 1–16) × max_features ∈ {log2, sqrt}

Wine — K-Nearest Neighbours n_neighbors ∈ {1, …, 89} (89 levels) × weights ∈ {uniform, distance} × metric ∈ {euclidean, manhattan, minkowski} × p ∈ {1, 2, 3} × leaf_size ∈ {10, 20, 30, 40, 50} × algorithm ∈ {auto, ball_tree, brute, kd_tree}

Supplementary Metrics

Regression landscapes include additional_metrics.neg_mean_squared_error, neg_mean_absolute_error, and neg_root_mean_squared_error (mean and std). Classification landscapes include additional_metrics.f1_macro, f1_micro, f1_weighted, precision_macro, recall_macro, and (where available) roc_auc_ovr (mean and std).

Dataset Creation

Curation Rationale

Benchmark suites for hyperparameter optimisation algorithms typically evaluate algorithms by running them, making fair comparison dependent on budget, parallelism, and random seed. This dataset eliminates that confound by pre-computing the entire performance surface, enabling replay-based algorithm evaluation and direct landscape topology analysis.

The seven experiments were chosen to span a range of landscape characteristics: grid size (4k–112k), task type (regression, binary and multiclass classification), model family (tree ensembles, linear, distance-based), and computational cost per configuration (milliseconds to minutes), providing a heterogeneous testbed for both landscape analysis and algorithm benchmarking.

Source Data

Data Collection and Processing

All landscapes were generated with ExaTune on Digital Research Alliance of Canada (Compute Canada) HPC clusters using SLURM job arrays. Each configuration was evaluated with scikit-learn's cross_validate using 5-fold stratified cross-validation (classification) or standard cross-validation (regression). Per-fold scores and wall-clock times were collected for every configuration; no configurations were pruned or early-stopped. Results were aggregated from per-job Parquet outputs and stored with Snappy compression.

Source datasets are standard public benchmarks:

California Housing — StatLib (Pace & Barry 1997), via sklearn.datasets
Diabetes — Efron et al. (2004), via sklearn.datasets
Gestational Diabetes — Publicly available clinical dataset
Maternal Risk — UCI ML Repository (Ahmed et al. 2020)
Titanic — Kaggle Titanic survival prediction
UCI Adult (Census Income) — Kohavi (1996), UCI ML Repository
Wine Quality — Cortez et al. (2009), UCI ML Repository

Who are the source data producers?

The hyperparameter landscapes were generated by Kevin Dick and collaborators at BORN Ontario/CHEO-RI using the ExaTune framework. The underlying tabular datasets are produced by their respective original authors (see citations above); ExaTune evaluates ML models trained on those datasets but does not redistribute the raw data.

Annotations

No manual annotations. All performance metrics are computed programmatically by scikit-learn's cross-validation routines.

Personal and Sensitive Information

The dataset contains no personal or sensitive information. The underlying source datasets (Titanic, UCI Adult, Maternal Risk, Gestational Diabetes) contain demographic and clinical variables in their original form, but only aggregate cross-validated model performance metrics derived from those datasets are recorded here — no individual-level records are included.

Bias, Risks, and Limitations

Grid discretisation: All hyperparameter axes are discretised; the true continuous optimum may lie between grid points. Landscape metrics are valid only on the defined discrete grid graph.
Dataset-specificity: Landscape topology (ruggedness, multimodality) is specific to the combination of dataset and model family. Results do not generalise to other datasets without re-running exhaustive evaluation.
Single random seed: Cross-validation splits were fixed by the scikit-learn default random state. Performance estimates may differ under different split seeds, particularly for small datasets (Gestational Diabetes, Maternal Risk).
Narrow performance range: The Diabetes landscape (R² ∈ [0.252, 0.279]) has an extremely compressed range; importance and interaction analyses on this landscape should be interpreted cautiously.
Fixed hyperparameters: Several XGBoost landscapes hold colsample_bytree, gamma, and min_child_weight constant. The reported landscapes are projections of a higher-dimensional space, not full joint landscapes.

Recommendations

Users should verify that the hyperparameter grid resolution is appropriate for their analysis before drawing conclusions about landscape smoothness or importance. Subsampling analyses (see companion figure scripts) can quantify sensitivity to grid coverage.

Citation

BibTeX:

@inproceedings{dick2025exatune,
  title     = {{ExaTune}: Exhaustive Hyperparameter Landscape Exploration at Scale},
  author    = {Dick, Kevin and others},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026},
  note      = {Under review}
}

APA:

Dick, K., et al. (2026). ExaTune: Exhaustive Hyperparameter Landscape Exploration at Scale. Advances in Neural Information Processing Systems (NeurIPS). Under review.

Glossary

Landscape / fitness landscape: The mapping from hyperparameter configuration λ ∈ Λ_G to scalar performance metric f(λ), treated as a function on the grid graph G = (Λ_G, E).
Grid graph G: The Cartesian product of discretised hyperparameter axes; two configurations are adjacent if they differ in exactly one parameter by one rank step.
Node-hours: Total computation measured as K × Σᵢ(fit_time_mean_i + score_time_mean_i) / 3600, where K is the number of CV folds.
Moran's I: Spatial autocorrelation statistic on G; I > E[I] indicates smooth landscapes.
Ruggedness R: Ratio of mean neighbourhood variance to global variance; R → 1 indicates a fully random landscape.
FDC (fitness-distance correlation): Pearson correlation between performance and grid distance to the global optimum; FDC > 0.15 indicates a "big valley" structure.
η² (eta-squared): One-way ANOVA effect size quantifying the fraction of total performance variance explained by a single hyperparameter.

Dataset Card Authors

Kevin Dick — CHEO-RI / BORN Ontario

Dataset Card Contact

kdick@bornontario.ca

Downloads last month: 74

Total file size:

249 MB