SysNetS/otel-data

POWDER and local-testbed cellular network data for OTel 2.5

Operational cellular-network records collected by the UT Dallas / Purdue team for the OTel 2.5 training effort: the data classes the OTel 2.0 model card lists as absent from its training mix (event data, network-performance KPIs, RF/signal-quality measurements, 5G core control-plane and inter-NF signalling).

Three collections, each a versioned tarball with a per-file MANIFEST.csv (sha256), plus browsable run records, summaries and documentation:

Collection	Content	Provenance
powder-lte-v1	63 LTE runs on the POWDER testbed (srsRAN_4G), benign vs. fake base station: S1AP/NAS/RRC events, MAC byte counts, EPC measurement reports (RSRP/RSRQ/SINR), eNB/UE log events, frame-level malicious labels, detector outputs (kept separate), run records, derived summaries	measured, real radios
powder-5g-core-sim-v1	7 standalone-5G runs on POWDER (Open5GS 2.8.0 + UERANSIM): NGAP, 5G NAS, SBI, PFCP, GTP-U counters, NF logs, RADIUS, injected NF restarts and link impairment with ground-truth timelines	real core software, simulated radio
zmq-5g-core-sim-v1	87 local runs (Open5GS 2.7.7 + ocudu gNB + srsUE over ZMQ): NGAP, 5G NAS, NR RRC, SBI, PFCP, GTP-U/MAC-NR counters, gNB scheduler metrics, per-second UE radio metrics (RSRP, SNR, MCS, BLER, bitrates), iperf/ping/HTTP results, NF restarts/outages, scheduler MCS caps, TCP/UDP load, ground-truth timelines	real core/gNB/UE software, simulated radio

Record format

One JSON object per line (.ndjson). Every record carries run_id, scenario_label (run-level ground truth), capture_point, provenance (measured / derived / simulated), parser, source_file, and a UTC ts where the source clock is trustworthy. Subscriber identities are keyed pseudonyms (ue-…); no raw IMSI/SUPI/IMEI. Schema: docs/event.schema.json in each collection. Field meanings: docs/data_dictionary.md.

Rules the data follows

• Provenance label on every record; simulated-radio values are never presented as over-the-air measurements.
• Injected conditions (ground truth) live in each run record's scenario.timeline, separate from any detector output.
• Train/test splits must keep whole runs together (train_test_split.single-ue.json in the LTE collection).
• Raw packet captures are not included (45 GB for the local runs, 10.7 GB for the LTE archive); they are catalogued with sha256 in the run records and available from the owner on request.
• Third-party material (operator captures, drive tests) is excluded.

Layout

powder-lte-v1/              README.md, MANIFEST.csv, TERMS.md, run_records/, summaries/, docs/, catalog/
powder-lte-v1.tar.gz        full collection incl. processed/ NDJSON
powder-5g-core-sim-v1/      README.md, MANIFEST.csv, TERMS.md, run_records/, summaries/, docs/, runs/ (timelines, app results, configs)
powder-5g-core-sim-v1.tar.gz
zmq-5g-core-sim-v1/         same, plus runs/<run>/ran/metrics (UE metrics CSV)
zmq-5g-core-sim-v1.tar.gz

Owner: Mashroor Hasan Bhuiyan (UT Dallas). Collection and tooling: SysNetS Lab; pipeline at github.com/mahabhu/otel-data-collection.