Bayesian Bechmarks (bbench)

Thin wrapper around HuggingFace Datasets for the Bayesian Benchmarks collection.

All data lives on HuggingFace Hub. This library adds shuffled train/test splitting and optional z-score standardization on top. If you need more control, use datasets directly (see below).

Installation

With uv (recommended):

uv pip install -e /path/to/bbench

Or in a uv-managed project, add it as a dependency:

uv add /path/to/bbench

With pip:

pip install -e /path/to/bbench

Quick start

import bbench

# Browse available datasets by category
bbench.regression_datasets()       # ['3droad', 'airfoil', 'autompg', ...]
bbench.classification_datasets()   # ['abalone', 'acute-inflammation', ...]
bbench.reinforcement_datasets()    # ['Ant-v2', 'HalfCheetah-v2', ...]

# Load a dataset (seed=0, 90/10 train/test, standardized)
data = bbench.load_dataset("boston")
data.input_train.shape   # (455, 13)
data.target_train.shape  # (455, 1)

# Custom seed
data = bbench.load_dataset("boston", seed=42)

# Pass a numpy Generator directly
import numpy as np
rng = np.random.Generator(np.random.PCG64(42))
data = bbench.load_dataset("boston", seed=rng)

# Change train/test fraction
data = bbench.load_dataset("boston", train_fraction=0.8)

# Disable standardization to get raw data
data = bbench.load_dataset("boston", standardize=False)

Standardization

Enabled by default (standardize=True). Features are always z-score standardized. Targets are standardized for regression and reinforcement datasets but left as raw class labels for classification. Stats are computed on the training set and applied to the test set.

To recover original-scale regression predictions:

y_original = data.target_std * y_predicted + data.target_mean

Pass standardize=False to skip all normalization and get raw arrays.

Using datasets directly

from datasets import load_dataset
import numpy as np

ds = load_dataset("OccaMLab/bayesian-benchmarks", "boston", split="train")
X = np.array(ds["features"])
Y = np.array(ds["target"])

Each HF config has a single "train" split with columns features (list of floats) and target (list of floats).