---

synthbench generates synthetic datasets for benchmarking. You choose how complicated the signal is, layer noise or missing data on top, and get back a dataset that carries a record of how it was made. One integer seed reproduces the whole thing.

The point is knowing the answer in advance. On real data you can measure that your model scored 0.83 and have no idea whether 0.85 was available or whether you already hit the ceiling. Here the ceiling is in the metadata.

Installation

pip install synthbench

Parquet serialization needs pip install "synthbench[io]", and RandomNeuralDGP needs pip install "synthbench[neural]" for PyTorch. Everything else works from the base install.

Basic usage

from synthbench import BenchPipeline, LinearDGP, MissingDataCorruptor

pipeline = BenchPipeline(
    LinearDGP(complexity="medium", task_type="classification"),
    corruptors=[MissingDataCorruptor(proportion=0.1, mechanism="mar")],
)
result = pipeline.run(n_samples=500, n_features=10, random_state=42)

print(result.X.shape)  # (500, 10)
print(result.metadata["bayes_error_analytic"])  # exact error floor
print(result.metadata["effective_rank"])  # feature space dimensionality

Data-generating processes

Each one takes a complexity parameter and records which features actually carry signal.

DGP	Signal
LinearDGP	Linear combination, sparsity and noise set by complexity
PolynomialDGP	Polynomial terms and interactions
TreeDGP	Axis-aligned splits, depth set by complexity
FriedmanDGP	The Friedman 1/2/3 benchmark functions
AdditiveDGP	Sum of univariate functions, GAM-style
SparseDGP	Explicit number of informative features via k
GeometricDGP	Moons, circles, spirals
RandomNeuralDGP	A randomly initialised MLP (needs [neural])

Corruptors

MeasurementNoise, Outlier, MissingData, Collinearity, and Categorical transform the feature matrix; LabelNoiseCorruptor goes in the separate label_corruptors= argument and touches only y. Feature corruptors always run in a fixed order regardless of how you list them, so two pipelines with the same components produce the same data. Each one records how much of each feature's information it destroyed, in metadata["effective_feature_importances"].

Severity is a preset: "low", "medium", or "high". Pass the underlying parameter (proportion, noise_level, n_bins) when you want a specific value instead.

More than two classes

Pass n_classes to any DGP except FriedmanDGP, with class_weight="balanced" or an explicit list of priors.

from synthbench import BenchPipeline, LinearDGP

result = BenchPipeline(
    LinearDGP(
        task_type="classification", n_classes=4, class_weight=[0.4, 0.3, 0.2, 0.1]
    )
).run(n_samples=1000, n_features=10, random_state=0)

How the classes are produced differs by DGP, and metadata["label_mechanism"] says which you got. LinearDGP, SparseDGP, TreeDGP, and RandomNeuralDGP emit one latent score per class and draw through a softmax, so the classes are unordered. PolynomialDGP and AdditiveDGP are built on a fixed term basis with only one signal to work with, so theirs is cut into ordered bins — fine for genuinely ordinal targets, misleading as a general multiclass benchmark. FriedmanDGP refuses: its functions are specific published formulas, and a multi-output variant would not be the thing anyone cites.

How hard is the dataset?

Two numbers, and the difference between them matters:

• bayes_error_analytic is exact. The DGP draws each label from a known probability, so the irreducible error is computable rather than estimated. No model can do better. It reflects the labels only, so feature corruption does not change it.
• bayes_error is an empirical 1-NN leave-one-out error on the corrupted features. It moves when you corrupt things, which makes it useful for comparing severity levels against each other, but it is biased upwards in more than a few dimensions. It is not a bound.

For a balanced LinearDGP classification task the analytic floor is about 0.325, while the 1-NN estimate reads about 0.456 at n=500, p=10. Use the first when you need a floor and the second when you need a relative signal.

Ablations

severity_sweep and difficulty_sweep vary one axis and return a dict keyed by level. experiment_grid crosses sample size, complexity, and severity. Seeds come from a nested SeedSequence hierarchy, so cells are independent of each other but reproducible across runs.

from synthbench import LinearDGP, OutlierCorruptor, experiment_grid

grid = experiment_grid(
    LinearDGP,
    OutlierCorruptor,
    n_samples_list=[200, 500, 1000],
    complexities=["low", "medium", "high"],
    severities=["low", "medium", "high"],
    n_features=10,
    random_state=0,
    task_type="classification",
)

print(grid[(500, "high", "medium")].metadata["bayes_error_analytic"])

Suites and serialization

BenchSuite("easy-classification").run() generates a curated collection in one call and returns it keyed by label — handy as a shared baseline, since the suite name alone pins the data. BenchSuite.from_dict and BenchSuite.from_json take your own specs.

to_parquet / to_csv round-trip the data and the full metadata, and BenchPipeline.from_metadata rebuilds the pipeline from that metadata so a saved dataset can be regenerated rather than shipped.

Docs

Full reference and runnable notebooks at JanTeichertKluge.github.io/synth-bench.