NetFlow V3

Parquet mirror of the four NetFlow V3 datasets published by the University of Queensland. The files contain 53 flow and temporal features plus the binary Label and multiclass Attack targets. Together they contain 66,935,021 completed bidirectional flow records.

Contents

Configuration	Rows	Attack rows	Attack values	Observed flow period (UTC)	Official record
NF-UNSW-NB15-v3	2,365,424	127,693	10	2015-01-22 to 2015-02-18	10.48610/6E0EDA1
NF-BoT-IoT-v3	16,933,808	16,881,819	5	2018-05-15 to 2018-06-19	10.48610/73C4EBC
NF-ToN-IoT-v3	27,520,260	10,728,046	10	2019-04-23 to 2019-04-29	10.48610/44D7C5E
NF-CSE-CIC-IDS2018-v3	20,115,529	2,600,903	15	2018-02-14 to 2018-03-02	10.48610/ECE9B83

The source archive names the final CSV NF-CICIDS2018-v3.csv; the Hub configuration retains the official NF-CSE-CIC-IDS2018-v3 dataset name. metadata/NetFlow_v3_Features.csv is the feature dictionary supplied by UQ.

Loading

from datasets import load_dataset

flows = load_dataset(
    "keys-i/netFlow",
    "NF-UNSW-NB15-v3",
    split="full",
)

full is a storage split only. UQ does not provide one common leakage-safe train/validation/test split for these four files. Do not interpret it as a training partition.

Conversion and checks

The source packages were downloaded from the four UQ records above. Their BagIt payload SHA-1 checks passed before conversion. Each CSV was converted with Zstandard compression using tools/scripts/convert.py:

pixi run python tools/scripts/convert.py \
  --input <source.csv> \
  --output <dataset.parquet> \
  --overwrite

No filtering, relabelling, deduplication, or feature engineering is performed by that converter. Verification on 2026-08-16 established that every Parquet row count equals its source CSV line count minus the header, all files have the same 55 column names in the same order, and every file opens as Zstandard Parquet.

Three inferred numeric types differ across configurations because the source values differ:

Field	UNSW / BoT-IoT	ToN-IoT / CSE-CIC-IDS2018
L7_PROTO	float64	int64
SRC_TO_DST_SECOND_BYTES	float64	int64
DST_TO_SRC_SECOND_BYTES	float64	int64

Cast these fields explicitly before concatenating configurations. Do not rely on an implicit cross-dataset coercion.

Research-use limits

• These are converted benchmark lineages, not four independent operational networks collected under one protocol.
• The files include endpoint addresses, absolute timestamps, labels, and other fields that can leak capture or scenario identity. A random row split does not demonstrate temporal, endpoint, attack-family, or cross-network generalisation.
• Label and Attack must be hidden from self-supervised pretraining. Dataset taxonomies are configuration-specific and must not be merged by fuzzy name matching.
• Duplicate and near-duplicate flows must be audited before constructing any evaluation split.
• Results on these four files alone are benchmark evidence, not proof of a general-purpose network foundation model or operational intrusion detector.

Rights and attribution

Each UQ catalogue record reports open access and lists its licence as “Other.” UQ currently defines open access research data as available for access, use, reuse, and redistribution. This mirror therefore uses Hugging Face's other licence marker, links to that UQ guidance, and does not relicense the source data. Review the applicable UQ record and terms before reuse. The previous dataset card's CC-BY-NC-SA-4.0 label was not supported by the UQ records and has been removed.

Please cite the applicable dataset DOI from the contents table and the V3 feature paper:

@article{luay2026time,
  author  = {Majed Luay and Siamak Layeghy and Niloufar Noorbin and
             Mohanad Sarhan and Gayan Kulatilleke and Nour Moustafa and
             Marius Portmann},
  title   = {Time Matters: Temporal NetFlow Features for ML-Based Network
             Intrusion Detection},
  journal = {IEEE Access},
  volume  = {14},
  pages   = {66899--66913},
  year    = {2026},
  doi     = {10.1109/ACCESS.2026.3688204}
}