ahancock516/5gnr-pusch-iq-dmrs

5G NR PUSCH IQ DMRS Captures

Frequency-domain PUSCH IQ captures from an OAI 5G NR gNB (NI USRP X410 / USRP B210) with per-device IMSI labels embedded directly in each capture record. Collected over-the-air with commercial Quectel RM520N-GL modems and a software-defined USRP B210 UE.

Dataset summary

• Filtering: only captures with a strongly visible DMRS RE comb (active/quiet power ratio ≥ 1.5×) are accepted
• Format: v4 binary (.bin) — see nr_pusch_capture_oai for the reader
• Modulation: QPSK (transform-precoded DFT-s-OFDM)
• Band: n78 (3.5 GHz, 106 PRB, 40 MHz)
• Labeling: 100% labeled — IMSI resolved from AMF via socket and embedded per-capture

Files

Single-device datasets (200 captures each, 100% labeled)

File	Captures	IMSI	UE
pusch_dataset_20260618_200captures_b210_100pct_labeled.bin	200	…001	USRP B210 (OAI nr-uesoftmodem)
pusch_dataset_20260618_200captures_imsi006_100pct_labeled.bin	200	…006	Quectel RM520N-GL on Raspberry Pi 5
pusch_dataset_20260618_200captures_imsi007_100pct_labeled.bin	200	…007	Quectel RM520N-GL on Raspberry Pi 5
pusch_dataset_20260618_200captures_imsi008_100pct_labeled.bin	200	…008	Quectel RM520N-GL on Raspberry Pi 5
pusch_dataset_20260618_200captures_imsi010_100pct_labeled.bin	200	…010	Quectel RM520N-GL on Raspberry Pi 5

Multi-device mixed datasets (100% labeled)

File	Captures	Devices	UEs
pusch_dataset_20260619_1883captures_5ue_mixed_100pct_labeled.bin	1883	5	B210 + 4× Quectel RM520N-GL
pusch_dataset_20260619_1690captures_5ue_mixed_100pct_labeled.bin	1690	5	B210 + 4× Quectel RM520N-GL

Each dataset includes a matching label_map_*.json with full session metadata (RNTI→IMSI mapping, AMF UE NGAP ID, session timestamps).

Label distributions

Single-device datasets

Each single-device file contains 200 captures from one UE (100%).

Mixed 5-UE dataset — 1883 captures (June 19, 2026)

IMSI	Captures	Share	UE
001010000000001	388	20.6%	USRP B210 (OAI nr-uesoftmodem)
001010000000006	47	2.5%	Quectel RM520N-GL on Raspberry Pi 5
001010000000007	85	4.5%	Quectel RM520N-GL on Raspberry Pi 5
001010000000008	920	48.9%	Quectel RM520N-GL / Google Pixel 7
001010000000010	443	23.5%	Quectel RM520N-GL / Google Pixel 7

Mixed 5-UE dataset — 1690 captures (June 19, 2026)

IMSI	Captures	Share	UE
001010000000001	374	22.1%	USRP B210 (OAI nr-uesoftmodem)
001010000000006	207	12.2%	Quectel RM520N-GL on Raspberry Pi 5
001010000000007	246	14.6%	Quectel RM520N-GL on Raspberry Pi 5
001010000000008	459	27.2%	Quectel RM520N-GL / Google Pixel 7
001010000000010	404	23.9%	Quectel RM520N-GL / Google Pixel 7

File format

Each .bin file begins with a 64-byte file header followed by sequential capture records.

File header (64 bytes)

Field	Type	Description
magic	uint32	0x50555343 ("PUSC")
version	uint32	Format version (4)
max_captures	uint32	Capture limit configured at collection time
num_captures	uint32	Actual captures written

Capture record

Each record is a 144-byte header followed by interleaved int16 IQ samples.

Field	Type	Description
record_bytes	uint32	Total record size including header
frame, slot	int32	NR frame/slot timing
timestamp_ns	int64	CLOCK_MONOTONIC timestamp
rnti	uint16	Radio network temporary identifier
rb_size	int32	Allocated resource blocks
rb_start	int32	Starting resource block
num_symbols	int32	OFDM symbols in allocation
qam_mod_order	uint8	Modulation order (Qm)
transform_precoding	uint8	1 = DFT-s-OFDM, 0 = CP-OFDM
dmrs_config_type	uint8	DMRS type 1 or 2
num_dmrs_cdm_grps	uint8	CDM groups without data (1 or 2)
ul_dmrs_symb_pos	uint32	DMRS symbol position bitmask
nvar	uint32	Noise variance estimate
iq_bytes	int32	IQ payload size in bytes
imsi[16]	char	Null-terminated IMSI string

IQ payload: num_symbols × rb_size × 12 complex samples as interleaved int16 (real, imag).

Reading the dataset

import struct

FILE_HDR_SIZE = 64
CAP_HDR_SIZE  = 144

with open("pusch_dataset_20260619_1690captures_5ue_mixed_100pct_labeled.bin", "rb") as f:
    data = f.read()

magic, version, max_cap, num_cap = struct.unpack_from("<IIII", data, 0)
print(f"v{version}, {num_cap} captures")

off = FILE_HDR_SIZE
while off + CAP_HDR_SIZE <= len(data):
    record_bytes = struct.unpack_from("<I", data, off)[0]
    if record_bytes == 0 or off + record_bytes > len(data):
        break
    rb_size  = struct.unpack_from("<i", data, off + 40)[0]
    n_sym    = struct.unpack_from("<i", data, off + 48)[0]
    iq_bytes = struct.unpack_from("<i", data, off + 136)[0]
    imsi     = data[off+128:off+144].rstrip(b"\x00").decode("ascii", "replace")
    iq_data  = data[off + CAP_HDR_SIZE : off + CAP_HDR_SIZE + iq_bytes]
    # iq_data is n_sym × rb_size × 12 interleaved int16 (I, Q) samples
    off += record_bytes

Hardware

• gNB RF: NI USRP X410 (OAIBOX) running OAI nr-softmodem
• Core network: OAI CN5G (AMF/UPF/NRF)
• UEs:
  • Quectel RM520N-GL (5G SA modem) on Raspberry Pi 5 — IMSIs …006, …007, …010
  • Quectel RM520N-GL on Google Pixel 7 — IMSI …008
  • NI USRP B210 running OAI nr-uesoftmodem — IMSI …001
• Capture plugin: nr_pusch_capture_oai

Citation

If you use this dataset, please cite the OAI project and the C2A2 Lab at Florida Atlantic University.