Madelon

libsvm:madelon@1 · Registry 1.4.0. Default version in this registry.

Artificial binary classification data from the NIPS 2003 feature selection challenge, distributed by LIBSVM with 500 features. The predefined test artifact contains the original challenge validation set, not its unlabeled test set.

Dimensions and tasks

Task	Type	Split	Samples	Predictors
default	classification	test	600	500
default	classification	train	2,000	500

Predictor counts follow each task's feature selection and exclude targets and identifiers. Sample counts are reported separately for each split.

Retrieve

These examples pin both the dataset version and the registry's exact index. An uncached dataset requires network access to its source. The CLI retrieves verified artifacts. The language clients also decode them.

Python

from datamonger import Registry, fetch_data

registry = Registry(
    release="1.4.0",
    index_sha256="ef83679266af9fcbb9ea194acdfb9bb6c380b440a574fc2ee7292899c5102909",
    index_url="https://github.com/jolars/datamonger/releases/download/registry-v1.4.0/index.json",
)
data = fetch_data("libsvm:madelon@1", registry=registry)

R

library(datamonger)

registry <- registry_selector(
  release = "1.4.0",
  index_sha256 = "ef83679266af9fcbb9ea194acdfb9bb6c380b440a574fc2ee7292899c5102909",
  index_url = "https://github.com/jolars/datamonger/releases/download/registry-v1.4.0/index.json"
)
data <- fetch_data("libsvm:madelon@1", registry = registry)

Julia

using Datamonger

registry = registry_selector(
    "1.4.0",
    "ef83679266af9fcbb9ea194acdfb9bb6c380b440a574fc2ee7292899c5102909",
    "https://github.com/jolars/datamonger/releases/download/registry-v1.4.0/index.json",
)
data = fetch_data("libsvm:madelon@1"; registry=registry)

CLI

cat > datamonger-registry.json <<'JSON'
{
  "index_sha256": "ef83679266af9fcbb9ea194acdfb9bb6c380b440a574fc2ee7292899c5102909",
  "index_url": "https://github.com/jolars/datamonger/releases/download/registry-v1.4.0/index.json",
  "release": "1.4.0",
  "schema_version": 1
}
JSON
datamonger --selector datamonger-registry.json fetch 'libsvm:madelon@1'

Representation and variables

Tabular data with a sparse representation. Decoder: libsvm-split version 2.

Component	Kind	Type	Shape
train_features	sparse_matrix	float64	2,000 × 500
train_response	vector	int64	2,000
test_features	sparse_matrix	float64	600 × 500
test_response	vector	int64	600

Preprocessing

Upstream

Status: documented.

The exact numerical transformations used to prepare LIBSVM's feature values have not been established.

synthetic-generation. Generate clusters on a five-dimensional hypercube, add 15 linear combinations to the five informative features and 480 distractor features, and randomize feature and observation order.

split-selection. LIBSVM uses the original 600-row validation set as its test set because labels for the original 1800-row test set are unavailable.

Evidence: https://archive.ics.uci.edu/dataset/171/madelon, https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary.html#madelon

Datamonger

Status: none.

LIBSVM v2 omits explicitly stored zeros without changing feature values. Datamonger applies no statistical preprocessing.

Evidence: spec/decoders/libsvm-v2.md

Provenance and citation

Provider: LIBSVM. Upstream name: madelon. Retrieved: 2026-10-26.

Original source

Authors: Isabelle Guyon.

Guyon, I. (2004). Madelon [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C5602H.

Original dataset and creator as cited by UCI.

Isabelle Guyon, Steve Gunn, Asa Ben Hur, and Gideon Dror. Result analysis of the NIPS 2003 feature selection challenge. Advances in Neural Information Processing Systems, volume 17, 2005.

Reference IG05a associated with Madelon by LIBSVM: https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/ref.html#IG05a.

License

Unknown

UCI licenses the original Madelon dataset under CC BY 4.0: https://archive.ics.uci.edu/dataset/171/madelon. LIBSVM identifies that source and documents the split selection but states no license terms for its converted artifacts: https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary.html#madelon. The applicability of the original license to these distributed artifacts has not been established.

Artifacts

Total download: 9.7 MiB. Sizes describe the distributed files, including any compression.

Artifact	Format	Compression	Size	Distribution	Preservation
train	libsvm	none	7.4 MiB	upstream-only	unknown
test	libsvm	none	2.2 MiB	upstream-only	unknown

train: upstream

test: upstream

Versions in this registry

1 (default)

Registry record

Download the complete registry record, registry selector, or full registry index.

Index SHA-256: ef83679266af9fcbb9ea194acdfb9bb6c380b440a574fc2ee7292899c5102909.