LEVEL Running Dataset

Open, raw running recordings from body-worn LEVEL motion sensors (feet, lower back, wrist, arm), time-synchronized with a Polar H10 chest strap (heart rate, RR intervals, ECG), phone GPS, barometer and step counter, and additional ground truth where available. Collected with the LEVEL Collector platform, which runs on Android phones and Windows PCs, across deliberately varied sensor setups, and tracked run over run for open research on running performance, fatigue and injury prevention.

In 2025, my doctor gave me "the talk." You know the one. The one where I need to exercise more and eat better. So I picked up running, and somewhere along the way I thought: I'm a biomechanics engineer, and I work for a company that makes IMU motion sensors. Why not take some measurements?

Two purposes

1. LEVEL Collector, working anywhere. An open record of the collection platform itself: different sensors, placements, settings and devices, in and out of the research lab. LEVEL Inez is a small wearable 6-axis motion sensor (accelerometer ±16 g, gyroscope) that clips or straps onto the body -- here both feet, the lower back and a wrist or arm -- and streams for about 36 hours (a day and a half) at 100 Hz on a charge. LEVEL Collector is the research app that records several of them at once, over the phone's Bluetooth or the LEVEL Hub USB receiver, with the sensor clocks kept in step, alongside the phone's GPS, barometer and step counter and add-ons such as a Polar H10 chest strap, all on one clock, as plain documented CSV files (the format spec).

2. Running research. A longitudinal, within-participant record of real running: pace, heart rate, cadence and foot mechanics across runs, conditions and fatigue. Almost all running-injury research compares runners with each other, cross-sectionally or in prospective cohorts with a single baseline measurement, so following a runner's own mechanics over time is still rare; and the public running-IMU datasets are nearly all lab- or treadmill-bound and rarely bilateral at the foot. This one is within-subject and longitudinal, outdoors, with both feet and every stream on one clock, built to ask whether your own mechanics drift predicts your breakdown, with chest-strap, phone-GPS and other reference data to check the sensors against.

One 61-minute run at a glance: movement from each LEVEL sensor, heart rate, speed, cadence and relative elevation, all on one clock. Gaps are real dropouts.

The sessions

9 sessions, 1 participant, 4.9 h of recording, 25 km outdoors, 2026-09-07 to 2026-10-03.

Date	Session	Conditions	Duration (min)	Distance (km)	LEVEL sensors	Reference data	Packet loss (%)	Eff. rate (Hz)	Max gap (s)
2026-09-07	R01	outdoor pavement	61	7.21	4: Arm, Back, Left Foot, Right Foot	watch	0.6	99.3	0.31
2026-09-12	R02	outdoor pavement, parkrun	44	5.29	4: Back, Left Foot, Left Wrist, Right Foot	chest strap, watch	10.4	89.5	0.66
2026-09-14	R03	outdoor pavement	32	3.47	4: Back, Left Foot, Left Wrist, Right Foot	chest strap, watch	4.0	95.9	0.51
2026-09-14	R04	outdoor pavement	28	2.91	4: Back, Left Foot, Left Wrist, Right Foot	watch	4.1	95.8	1.51
2026-09-17	R05	indoor track	8		4: Back, Left Foot, Left Wrist, Right Foot	chest strap	0.3	99.6	0.31
2026-09-19	R06	outdoor pavement, race	15	1.94	4: Back, Left Foot, Left Wrist, Right Foot	chest strap	6.6	93.3	0.71
2026-09-28	R07	outdoor pavement	37	4.24	4: Back, Left Foot, Left Wrist, Right Foot		3.6	96.3	0.41
2026-09-30	R08	outdoor pavement	45		4: Back, Left Foot, Right Foot, Wrist		0.7	99.2	0.31
2026-10-03	R09	outdoor pavement, parkrun	20		4: Back, Left Foot, Right Foot, Wrist		0.0	99.9	0.01

9 sessions. Click a session for its overview figure; every column (setup, versions, heart rate, capture quality) is in sessions.csv.

Every run's figures sit in figures/<subject>/<session>/: overview.png (start there), data-*.png for each published source as recorded (imu, phone, h10, fit3), derived-*.png for anything computed from them (spectrum), and analysis-*.png from our running analysis: gait (cadence, ground contact time, vertical oscillation over the run), foottilt (a 3 s window of foot swing), groundtruth (IMU cadence against the watch, heart rate and GPS pace, where recorded) and pace (pace and elevation, where GPS was on). The analysis figures are provisional: the algorithms are not yet validated against a lab reference, and make_figures.py does not regenerate them. sessions.csv has one row per session with every column, in groups that read left to right: conditions, the setup that was varied, what the run measured, then the capture analysis (packet loss, clock alignment, phone battery); the column definitions are in the format spec.

Failures stay in: a chest strap that dropped out, a capture cut short, sensor clocks that did not align. They are flagged in sessions.csv (the capture columns and each session's notes) and in the figures.

The data

data/<subject>/R01/level/  one folder per session: LEVEL sensor CSVs, chest strap, phone streams, metadata
data/<subject>/R01/fit3/   watch data for the run (where worn)
data/context/              the only non-anonymous folder: per-session date, conditions, event; participants
data/data_dictionary.md    every file, column and unit, and the de-identification policy
figures/                   per session: overview.png, data-*.png (as recorded), derived-*.png (computed),
                           analysis-*.png (running analysis, provisional)
scripts/                   lvl_running (iterate / load / plot) + make_figures.py, make_catalog.py

Quick start

# pip install -r scripts/requirements.txt    (run from the dataset folder)
import sys; sys.path.insert(0, "scripts")
from lvl_running import iter_sessions, load_session, plot_session

for run_dir in iter_sessions("data"):          # every run, all participants and sessions
    run = load_session(run_dir)                # LEVEL sensors by placement + phone streams + metadata
    print(run.name, sorted(run.imus), f"{run.duration_s / 60:.0f} min")

plot_session("data/S01/R01/level", "my_figures")   # the same figures as figures/S01/R01/

python scripts/make_figures.py regenerates the overview, data-* and derived-* figures from data/; python scripts/make_catalog.py regenerates sessions.csv and the session table above.

Limitations

• One participant so far, the dataset's author. More participants will join; until then nothing here generalizes beyond one runner.
• Setups vary between runs on purpose (see sessions.csv); compare runs on the sensors they share.
• Consumer references: the watch and phone GPS are convenient, not gold standards; the Polar H10 is the heart-rate reference.
• Accelerometer range: foot impacts can reach the ±16 g limit, and 100 Hz under-samples the impact transient (see derived-spectrum.png); timing-based measures are unaffected.

License and citation

The data and figures are CC BY-NC 4.0 (see LICENSE): free for research and other non-commercial use with attribution; commercial use reserved. The code in scripts/ is MIT (see scripts/LICENSE). Copyright Level Health Innovations Corp.

Cite as: Lin, J. (2026). LEVEL Running Dataset. Level Health Innovations Corp. https://huggingface.co/datasets/lvlmotion/running

Built at LEVEL. Jonathan Lin designed, collected and analysed this dataset; the experiment and reporting were built with AI assistance from Claude (Anthropic). Questions and collaboration: jlin@lvlmotion.com