ilyakolosov/WIFI-CSI-Dataset

---

WiFi CSI Router Reception Dataset tags: - wifi - csi - esp32 - signal-processing - time-series size_categories: - 100K<n<1M

WiFi CSI Router Reception Dataset

This repository contains a small, structured dataset of WiFi Channel State Information (CSI) captures recorded as text logs (.data). The data is organized as repeated trials (test_XX) for 4 class labels (label_00..label_03) and 4 subjects/persons (id_person_01..id_person_04), recorded simultaneously by 3 receiver devices (dev1..dev3).

The “Hugging Face–ready” dataset folder is wifi_data_set/.

Dataset summary

• Modality: WiFi CSI (I/Q samples) + per-frame metadata
• Format: plain text files; 1 CSI frame per line
• CSI vector: 128 signed integers per frame (typically interpreted as 64 complex subcarriers as interleaved I/Q pairs)
• People (subjects): 4 (id_person_01..04)
• Labels (classes): 4 (label_00..03) (see “Label meanings” below)
• Trials: 100 per (person, label)
• Devices per trial: 3 (dev1, dev2, dev3)
• Frames per file: 100 lines (see “Known issues” for malformed first lines in some files)
• Total files: 4 persons × 4 labels × 100 tests × 3 devices = 4,800 .data files
• Total frames: 4,800 files × 100 lines = 480,000 logged frames
• Collection time range: 2026-04-18 20:17:59.599 +03:00 … 2026-04-18 23:21:01.561 +03:00
• Typical sampling: ~0.039–0.044 s between frames (≈ 23–28 Hz); ~4 s per trial

Dataset structure

Primary structured dataset:

wifi_data_set/
  id_person_01/
    label_00/
      test_01/
        test1__dev1_64_E8_33_57_AA_F4.data
        test1__dev2_64_E8_33_58_9E_28.data
        test1__dev3_64_E8_33_58_9C_CC.data
      ...
    label_01/ ...
    label_02/ ...
    label_03/ ...
  id_person_02/ ...
  id_person_03/ ...
  id_person_04/ ...

Notes:

• Each test_XX/ folder represents one trial.
• Each trial contains 3 .data files (one per receiver device).
• The receiver “device id” and device MAC-like identifier are embedded in the filename.

Raw/original layout (also present in this repo):

<label><suffix-of-dashes>/
  testN/
    testN__devK_<device_id>.data

Here, the number of trailing - characters encodes the person/subject. The scripts restructure_dataset.py / restructure_dataset.ps1 convert the raw layout into wifi_data_set/.

Label meanings

label_00..label_03 are the 4 class labels present in the dataset.

This repository does not contain a textual mapping from label IDs to human-readable class names. If you publish this dataset, consider filling in the table below:

Label directory	Class name/description
label_00	TODO
label_01	TODO
label_02	TODO
label_03	TODO

File format

Each .data file is a newline-delimited text log. Each line contains:

1. A timestamp prefix (with timezone offset), then
2. A CSV record that starts with the literal tag CSI_DATA, then
3. A final field which is a JSON-like list of integers (the CSI vector), usually quoted.

Example line (wrapped for readability):

18.04.2026 22:44:41.143 +03:00 CSI_DATA,
  328252,50:ff:20:e5:a8:f3,-43,11,1,7,0,1,1,1,1,0,0,-90,0,6,0,669572834,0,83,0,128,0,
  "[0,0,-39,16,-37,13,...]"

Parsed schema (standard CSI_DATA layout)

After splitting at " CSI_DATA,", the remaining CSV fields are typically:

Index	Name	Type	Notes
0	seq	int	Frame/packet sequence counter (monotonic within a file)
1	mac	str	Transmitter/AP MAC in colon form (observed constant: 50:ff:20:e5:a8:f3)
2	rssi	int	Received signal strength (dBm), e.g. -43
3	rate	int	PHY rate field as logged by the capture tool
4	sig_mode	int	Signal mode (often 0/1 in ESP32 CSI logs)
5	mcs	int	Modulation and Coding Scheme index (0..7 observed)
6	bandwidth	int	Bandwidth flag (0 observed in valid lines)
7	smoothing	int	0/1 flag
8	not_sounding	int	0/1 flag
9	aggregation	int	0/1 flag
10	stbc	int	0/1 flag
11	fec_coding	int	0/1 flag (0 observed)
12	sgi	int	Short guard interval flag (0/1 observed)
13	noise_floor	int	Noise floor (dBm), values like -90..-98
14	ampdu_cnt	int	AMPDU counter (0 observed in valid lines)
15	channel	int	WiFi channel (6 observed in valid lines)
16	secondary_channel	int	Secondary channel (0 observed)
17	local_timestamp	int	Device local timestamp (may wrap into negative due to 32-bit overflow)
18	ant	int	Antenna index (0 observed)
19	sig_len	int	Signal length (24/83/118 observed)
20	rx_state	int	RX state (0 observed)
21	len	int	CSI vector length (128 observed in valid lines)
22	first_word	int	First word / reserved (0 observed)
23	csi	list[int]	CSI vector as 128 signed integers

CSI vector

• Length: 128 integers
• Value range (observed): [-92, 87]
• Common interpretation: 64 complex values as interleaved I/Q pairs: (I0,Q0,I1,Q1,...).

Known issues / data quality

Most lines parse cleanly, but there are a small number of malformed first lines:

• Valid frames (parse to 24 fields + 128-length CSI vector): 479,402 / 480,000 (99.875%)
• Malformed frames: 598 / 480,000 (0.125%)
• Affected files: 598 / 4,800 (12.46%)
• Pattern: the malformed record is always line 1 of the affected file; the remaining 99 lines are well-formed.

Typical corruption patterns include missing commas/quotes around the CSI vector, truncated first lines, or concatenated CSI_DATA tokens.

If you are building a loader, the simplest robust strategy is:

1. Parse every line, and
2. Keep only records where:
   • CSV field count is exactly 24, and
   • the csi field parses as a list of length 128.

Loading example (Python)

import csv
import json
from dataclasses import dataclass
from datetime import datetime

TS_FMT = "%d.%m.%Y %H:%M:%S.%f %z"

@dataclass(frozen=True)
class CSIFrame:
    ts: datetime
    meta: dict
    csi: list[int]  # length 128

def parse_csi_line(line: str) -> CSIFrame | None:
    if " CSI_DATA," not in line:
        return None
    ts_str, rest = line.split(" CSI_DATA,", 1)
    ts = datetime.strptime(ts_str, TS_FMT)

    row = next(csv.reader([rest], delimiter=",", quotechar='"'))
    if len(row) != 24:
        return None

    try:
        csi = json.loads(row[-1])
    except json.JSONDecodeError:
        return None
    if not isinstance(csi, list) or len(csi) != 128:
        return None

    keys = [
        "seq","mac","rssi","rate","sig_mode","mcs","bandwidth","smoothing","not_sounding",
        "aggregation","stbc","fec_coding","sgi","noise_floor","ampdu_cnt","channel",
        "secondary_channel","local_timestamp","ant","sig_len","rx_state","len","first_word",
    ]
    meta = {k: (int(v) if k != "mac" else v) for k, v in zip(keys, row[:-1])}
    return CSIFrame(ts=ts, meta=meta, csi=[int(x) for x in csi])

Suggested evaluation setups

There is no predefined train/validation/test split. Common choices for this type of dataset:

• Within-subject split: random trials into train/test per person.
• Cross-subject split: hold out id_person_0X as test (leave-one-subject-out).
• Device generalization: train on dev1+dev2, evaluate on dev3 (or vice versa).

License

No license file or explicit license statement is included in this repository. If you plan to publish/share the dataset, add a license and update the YAML header at the top of this README.

How wifi_data_set/ was produced

The scripts in the repository root reorganize the raw folder structure:

• restructure_dataset.py (Python)
• restructure_dataset.ps1 (PowerShell)

They map source directories named like 0, 0-, 0--, 0---, 1, 1-, ... into:

wifi_data_set/id_person_XX/label_XX/test_XX/*.data

Citation

If you use this dataset in academic work, consider citing it as:

@dataset{wifi_csi_router_reception_2026,
  title        = {WiFi CSI Router Reception Dataset},
  year         = {2026},
  note         = {Version as of 2026-04-18},
}