Datasets:

SysNetS
/

Twin6G

license: mit
pretty_name: Twin6G — Over-the-Air 5G NR PHY Dataset
tags:
  - 5g
  - 5g-nr
  - physical-layer
  - rf
  - iq
  - sigmf
  - srsran
  - open5gs
  - ran-telemetry
  - wireless
  - digital-twin
  - fake-base-station
size_categories:
  - 1K<n<10K
configs:
  - config_name: gnb_metrics
    data_files: '**/tier2_metrics/gnb_sched_ue.csv'
  - config_name: ue_nr
    data_files: '**/tier2_metrics/ue_nr_adb.csv'

Twin6G — Over-the-Air 5G NR PHY Dataset

Real over-the-air 5G NR measurements for the Twin6G project: raw radio-frequency samples and physical-layer telemetry of a live standalone 5G cell serving a commercial phone. Unlike simulator traces (for example UERANSIM), every recording here is a genuine radio observation, so it carries true physical-layer (Layer 1) information.

Each run is self-contained in a timestamped directory and provides two paired tiers captured in the same session:

Tier 1 — raw IQ: the rawest possible form, complex baseband samples of the cell sniffed off the air.
Tier 2 — PHY telemetry: per-UE radio metrics from the base station and radio measurements reported by the phone, plus the control-plane and MAC packet captures.

Provenance

All data is measured, over the air. It is collected on a software-defined radio testbed, not simulated. Each run directory carries a run_info.yaml and a timeline.ndjson with exact timing, hardware serials, configuration, and ground-truth events (attach, each IQ burst, teardown). The IQ and the base-station and phone metrics are simultaneous observations of the same live cell.

Runs

run_id	duration	band / carrier	UE	label
`2026-10-15T1621Z_poco-n41-phy`	20 min (2026-10-15 16:21:52–16:42:00 UTC)	n41 TDD, ARFCN 518670 = 2593.35 MHz, 20 MHz, 30 kHz SCS	Xiaomi Poco (COTS phone)	benign single-cell baseline

Testbed for this run

Core: Open5GS (standalone 5G core), PLMN 99970.
gNB: srsRAN Project (release 25.10), PCI 1, on a USRP B210.
UE: a real Xiaomi Poco phone, SIM on PLMN 99970, attached over the air and kept RRC-connected with light background traffic.
Sniffer: a second USRP B210 tuned to the same 2593.35 MHz carrier, recording raw IQ while the cell ran.

Directory layout (per run)

<run_id>/
  tier1_iq/        raw IQ bursts (.sc16) + SigMF sidecars (.sigmf-meta)
  tier2_metrics/   gnb_sched_ue.csv, ue_nr_adb.csv, ue_adb.ndjson
  pcap/            gnb_mac.pcap (MAC-NR), gnb_ngap.pcap (NGAP + 5G NAS)
  logs/            srsRAN console + srslog, per-burst sniffer logs, phone radio logcat, keepalive ping
  config/          exact Open5GS core and srsRAN gNB configs used
  meta/            final size report
  run_info.yaml    run metadata
  timeline.ndjson  ground-truth event timeline (UTC)

Tier 1 — raw IQ (`tier1_iq/`)

Duty-cycled over-the-air capture of the live cell: a 2-second burst every 20 seconds across the whole run, so samples span the full session at a practical size. For the run above: 46 bursts, ~7.9 GB.

iq_<utc>_c<freq>_r<rate>.sc16 — interleaved signed 16-bit complex samples (I, Q, I, Q, …), little-endian. One 2-second burst = 46,080,000 complex samples = 184.32 MB. Sample rate 23.04 Msps, center frequency 2593.35 MHz.
iq_<utc>_c<freq>_r<rate>.sigmf-meta — a SigMF metadata sidecar (core:datatype: ci16_le, sample rate, center frequency, hardware serial, UTC datetime, label).

Read one burst in Python:

import numpy as np
x = np.fromfile("iq_….sc16", dtype=np.int16).astype(np.float32).view(np.complex64)
fs, fc = 23.04e6, 2593.35e6   # sample rate, center frequency

Typical uses: RF fingerprinting (carrier-frequency offset, IQ imbalance, phase noise), synchronization-signal (SSB / PSS-SSS) analysis, spectrograms, and physical-layer inputs to a digital-twin or anomaly model.

Tier 2 — PHY telemetry (`tier2_metrics/`, `pcap/`)

gnb_sched_ue.csv — per-UE physical-layer metrics from the base-station scheduler:

column	meaning
`pci`, `rnti`	cell identity, UE radio identifier
`dl_cqi`, `dl_ri`, `dl_mcs`, `dl_brate`	downlink channel quality, rank, modulation-and-coding, bitrate
`dl_ok`, `dl_nok`, `dl_bler_pct`	downlink HARQ ACK / NACK and block error rate
`ul_pusch_snr`, `ul_rsrp`	uplink SNR and received power at the base station
`ul_ri`, `ul_mcs`, `ul_brate`, `ul_ok`, `ul_nok`, `ul_bler_pct`	uplink rank, MCS, bitrate, HARQ, BLER
`ul_bsr`, `ul_ta`, `ul_phr`	buffer status, timing advance, power headroom

ue_nr_adb.csv — the phone's own NR radio measurements, polled every 2 seconds over adb:

column	meaning
`ts`, `reg`	timestamp, registration state
`ss_rsrp`, `ss_rsrq`, `ss_sinr`	synchronization-signal received power, quality, SINR (the UE's downlink view)
`csi_rsrp`, `csi_rsrq`, `csi_sinr`	CSI-RS equivalents when reported

ue_adb.ndjson is the raw per-poll source for that CSV. The tier2_metrics/*.csv are derived from the raw logs and can be regenerated.

Packet captures (pcap/): gnb_mac.pcap is MAC-NR (per-transport-block MCS, grants, HARQ); gnb_ngap.pcap is NGAP control plane and contains the 5G NAS (5GMM) signaling, including the full registration sequence (Registration Request, Authentication Request/Response, Identity Request/Response, Security Mode Command). Read with Wireshark or tshark -r gnb_ngap.pcap -Y nas-5gs.

Loading in `datasets`

The IQ files are binary and are referenced by their SigMF sidecars; load them with NumPy as shown above. The CSV metrics are exposed as dataset-viewer configs:

from datasets import load_dataset
gnb = load_dataset("<your-username>/Twin6G", "gnb_metrics")  # per-UE PHY metrics
ue  = load_dataset("<your-username>/Twin6G", "ue_nr")        # phone-side NR measurements

Intended uses

Physical-layer feature extraction and RF fingerprinting of 5G base stations.
Training and validating digital-twin / anomaly models on real radio telemetry.
Fusing physical-layer features with higher-layer (NAS / RRC) detectors, for example to reduce false positives in fake-base-station detection.

Limitations and honest notes

Benign only. This run is a single legitimate cell. It is not a fake-base-station capture; that requires a separate two-transmitter benign-versus-fake handover scenario.
IQ is duty-cycled, 2 seconds every 20 seconds. It spans the full run but totals about 92 seconds of samples per 20-minute run, not continuous. This bounds storage and avoids radio contention with the live cell.
A minority of IQ bursts logged a transient hardware overflow (a few dropped samples at the start of the burst). Files are full-size and usable; the per-burst sniffer logs in logs/ record which bursts.
Identifiers are test values on a private PLMN (99970). Subscriber keys are not included; the UE config in config/ has credentials stripped.

Citation

Part of the Twin6G project (digital twin for 5G/6G). If you use this dataset, please cite the project and link this repository.