Datasets

Datasets

30 real and synthetic streams ship with SAMLB — 15 classification, 15 regression. They are downloaded from GitHub on first use and cached locally, so the first run of a dataset needs network access and later runs do not.

from samlb.datasets import list_datasets, load, stream

list_datasets("classification")
list_datasets("regression")

Classification

Dataset	Notes
electricity	Australian electricity market, the standard drift benchmark
covertype	Forest cover type, 7 classes
insects	Optical insect classification, gradual and abrupt drift
adult	Census income
credit_card	Fraud detection, heavily imbalanced
nomao	Place-data deduplication
poker_hand	Poker hands, many classes
shuttle	NASA shuttle telemetry, imbalanced
vehicle_sensIT	Vehicle acoustic/seismic sensors
new_airlines	Flight delay
movingRBF, moving_squares	Synthetic, continuous drift
synth_RandomRBFDrift, synth_agrawal	Synthetic generators
sea_high_abrupt_drift	Synthetic, abrupt concept changes

Regression

Dataset	Notes
california_housing, House8L, kings_county	Housing prices
bike, MetroTraffic	Demand and traffic counts, strong seasonality
ailerons, elevators	Flight control surfaces
superconductivity	Critical-temperature prediction
diamonds, fifa, cps88wages	Tabular real-world regression
wave_energy	Wave farm power output
fried, FriedmanGra, hyperA	Synthetic, including drifting variants

Streaming

stream() yields (x_dict, y) pairs in the dataset's natural order — the order carries the drift, so it is never shuffled.

for x, y in stream("electricity", task="classification", max_samples=5000):
    ...   # x: {"feature": float, ...}, y: int (cls) or float (reg)

Argument	Meaning
name	dataset name
task	"classification" (default) or "regression"
max_samples	stop after N instances
normalize	min-max scale features online, from the range seen so far

normalize=True scales using statistics from instances seen so far only — never the whole file — so it leaks no future information. It is what BenchmarkSuite(normalize=True) passes through, and is recommended for gradient-based regressors.

Loading in bulk

When you want arrays rather than a stream:

X, y, meta = load("electricity", task="classification", max_samples=None)

meta["feature_names"]   # list[str]
meta["target_name"]
meta["n_samples"], meta["n_features"]
meta["n_classes"]       # classification only

X is float32 of shape (n_samples, n_features); y is int32 for classification and float32 for regression.

Adding your own dataset

Write an .npz with four arrays and drop it in the right directory:

import numpy as np

np.savez_compressed(
    "samlb/datasets/classification/my_stream.npz",
    X=X.astype("float32"),                 # (n_samples, n_features)
    y=y.astype("int32"),                   # int32 cls / float32 reg
    feature_names=np.array(feature_names), # (n_features,)
    target_name=np.array("target"),
)

list_datasets() and load() pick it up automatically — no registry to edit.

Keep the rows in their real temporal order. A shuffled stream measures something else entirely.

Using a stream SAMLB does not ship

Nothing in the benchmark requires samlb.datasets; the evaluator only needs an iterable of (x_dict, y). Drive a model directly over your own iterator:

def my_stream():
    for row in my_source:
        yield {"f1": row.a, "f2": row.b}, row.label

hits = n = 0
for x, y in my_stream():
    pred = model.predict_one(x)
    if pred is not None:
        hits += pred == y
        n += 1
    model.learn_one(x, y)

To get BenchmarkSuite's tables and JSON output for such a stream, add it as an .npz as above — that is the only path the suite knows.