techynilesh/streamarena

A Living Benchmark for Machine Learning on Streaming Data

---

🏆 Live leaderboard: techynilesh.github.io/StreamArena · Dataset catalog

StreamArena aggregates datasets for stream learning — classification, regression, clustering, and anomaly detection under concept drift — into one consistently organized, task-first collection. It plays the same role for streaming/online ML that TabArena plays for tabular ML: a single place to find curated, ready-to-use datasets instead of hunting through individual paper repos.

See the GitHub repo for loaders, examples, and a download.py helper. Datasets were consolidated from several independent research codebases, deduplicated where the same dataset appeared in multiple sources, and reorganized by task. Every dataset is stored as a single unified format — CSV — chosen because it's what the streaming-ML ecosystem (River's stream.iter_csv, MOA, scikit-multiflow's FileStream) actually consumes row-by-row, unlike batch/columnar formats.

Dataset structure

classification/
├── real/               # real-world streams (electricity, forest cover, airlines, ...)
└── synth/               # synthetic drift generators (SEA, RBF, Hyperplane, Agrawal, Madelon, ...)
regression/
├── real/               # housing, wages, sensor/physical measurements, ...
└── synth/               # Friedman & Hyperplane synthetic generators
clustering/
├── real/               # real-world streams reused from classification
└── synth/               # synthetic drift streams + blobs
anomaly_detection/        # ODDS/ADBench-style outlier detection sets (all real-world)

See DATASETS.md for the full per-dataset table — exact instance/feature/class counts computed directly from each file, plus a best-effort source attribution (UCI, OpenML, DELVE, MOA/River generators, ODDS/ADBench, etc.) for every dataset.

All files are .csv. Anomaly-detection files hold feature columns plus a trailing label column; everything else follows the same feature-columns-plus-target convention. Every task except anomaly detection (which is entirely real-world benchmark data) is split into real/ and synth/.

Task	Count	Notes
Classification	42 files (22 real + 20 synthetic)	real/: electricity, forest cover, airlines, poker, weather, KDD-99, insects, Nomao, MNIST, Usenet, Gisette, Dota, Spambase, HAR, etc. synth/: classic drift generators (SEA, RBF, Hyperplane, Agrawal, Madelon)
Regression	30 files (25 real + 5 synthetic)	real/: housing (king's county, california, miami, brazilian), wages, sensor/physical (sarcos, naval propulsion, superconductivity, kin8nm), and more. synth/: Friedman & Hyperplane generators
Clustering	13 files (6 real + 7 synthetic)	Streaming clustering benchmarks — reuses classification drift streams plus a dedicated synthetic blobs set
Anomaly Detection	51 files	ODDS/ADBench-style outlier detection collection (annthyroid, mnist, shuttle, satellite, mammography, etc.) — all real-world, no real//synth/ split

Usage

pip install huggingface_hub

from huggingface_hub import snapshot_download

path = snapshot_download(repo_id="techynilesh/streamarena", repo_type="dataset")

Or download just one task:

from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id="techynilesh/streamarena",
    repo_type="dataset",
    allow_patterns=["classification/**"],
)

Then load files directly — it's always just a CSV:

import pandas as pd
df = pd.read_csv(f"{path}/classification/real/electricity.csv")

Using it with River or CapyMOA

Since every dataset is plain CSV, it plugs directly into the two most common Python streaming-ML libraries — no conversion needed.

# River
import pandas as pd
from river import metrics, stream, tree

path = "classification/real/electricity.csv"
sample = pd.read_csv(path, nrows=100)
target = sample.columns[-1]
# Convert only numeric feature columns to float; categorical/string columns
# (e.g. in adult.csv) pass through as-is — River trees handle them natively.
converters = {
    c: float for c in sample.columns[:-1] if pd.api.types.is_numeric_dtype(sample[c])
}

dataset = stream.iter_csv(path, target=target, converters=converters)
model = tree.HoeffdingTreeClassifier()
metric = metrics.Accuracy()

for x, y in dataset:
    y_pred = model.predict_one(x)
    model.learn_one(x, y)
    metric.update(y, y_pred)

print(metric)

# CapyMOA (requires a working JVM — Java 11+)
from capymoa.classifier import HoeffdingTree
from capymoa.evaluation import prequential_evaluation
from capymoa.stream import stream_from_file

stream = stream_from_file(
    "classification/real/electricity.csv",
    dataset_name="Electricity",
    class_index=-1,       # StreamArena's convention: label is the trailing column
    target_type="categorical",
)
learner = HoeffdingTree(schema=stream.get_schema())
results = prequential_evaluation(stream, learner)
print("accuracy:", results.cumulative.accuracy())

See examples/river_usage.py and examples/capymoa_usage.py on GitHub for the full runnable scripts.

License

MIT for the aggregation/curation. Individual datasets retain their original licenses/terms from their respective sources — check before redistribution.

Citation

If you use StreamArena in your research, please cite it as below:

@misc{verma2026streamarena,
  title  = {StreamArena: A Living Benchmark for Machine Learning on Streaming Data},
  author = {Verma, Nilesh},
  year   = {2026},
  url    = {https://github.com/TechyNilesh/StreamArena}
}

Please also cite the original dataset sources where applicable.