juliensimon/gw-candidate-alerts

Gravitational-Wave Candidate Alerts

Credit: NASA/CXC/A. Hobart

Part of a dataset collection on Hugging Face.

Dataset description

Low-latency gravitational-wave candidate alerts from GraceDB, the LIGO/Virgo/KAGRA candidate event database.

When the detector network is observing, a set of online search pipelines analyses the strain data continuously and uploads anything that looks like a signal within seconds to minutes. Related uploads are grouped into a "superevent" and given a public alert. This dataset is the complete public record of those superevents across the O3 and O4 observing runs.

Most of these candidates are not real. That is the point of the dataset: the confirmed detection catalogs describe only the events that survived offline analysis, while this record describes everything the detection system actually flagged in real time, including the noise. It makes the selection function visible.

Nine pipeline labels appear in the record. gstlal, pycbc, MBTA and spiir are matched-filter searches that correlate the data against modelled compact-binary waveforms; MBTAOnline is an older label for MBTA seen only in O3. CWB and oLIB are unmodelled burst searches that look for coherent excess power across detectors without assuming a waveform, oLIB having been retired after O3. MLy and aframe are machine-learning searches, and their arrival in O4 makes this dataset a record of when neural-network pipelines entered production gravitational-wave astronomy.

The false alarm rate, in hertz, is the central quantity. A candidate at 1e-7 Hz corresponds to roughly one false alarm every four months of observing; candidates near 1e-5 Hz are expected several times a day from noise alone. The reporting latency column records how long each pipeline took to upload its candidate, which is what determines whether a telescope can be pointed at a merger while its electromagnetic counterpart is still bright.

Coverage note: the O4 run ended on 2025-11-18 and no new superevents are expected until O5 begins. The pipeline continues to run weekly and will resume collecting automatically when the network returns to observing.

This dataset is suitable for tabular classification, tabular regression tasks.

Schema

Column	Type	Description	Sample	Null %
superevent_id	str	GraceDB superevent identifier, e.g. 'S251118cm'; 'S' prefix plus the UTC date and an alphabetic suffix ordering candidates within that day	S190405ar	0.0%
created	datetime64[us, UTC]	UTC timestamp when the superevent record was created in GraceDB; differs from the candidate time by the alert latency	2019-04-05 16:01:56+00:00	0.0%
event_time_utc	datetime64[us, UTC]	UTC time of the candidate signal, converted from gps_time assuming the constant 18 s GPS-UTC offset in force since 2017	2019-04-05 16:01:30.863646+00:00	0.0%
gps_time	float64	GPS time of the candidate signal in seconds (API field 't_0'); join key against the 'gps' column of juliensimon/gravitational-wave-events	1238515308.863646	0.0%
gps_start	float64	GPS time of the start of the superevent time window (API field 't_start')	1238515307.863646	0.0%
gps_end	float64	GPS time of the end of the superevent time window (API field 't_end')	1238515309.863646	0.0%
far_hz	float64	False alarm rate of the superevent in Hz (inverse seconds); 1e-7 Hz is roughly one false alarm per four months; lower means more significant	0.00021409731905668	0.0%
pipeline	str	Online search pipeline that produced the preferred event: gstlal, pycbc, MBTA and spiir are matched-filter CBC searches, CWB and oLIB are unmodelled burst searches, MLy and aframe are machine-learning searches; 'MBTAOnline' is an older label for MBTA and appears only in O3	gstlal	0.0%
search	str	Search configuration within the pipeline, e.g. 'AllSky', 'EarlyWarning', 'SSM'	AllSky	0.0%
analysis_group	str	Analysis group of the preferred event: 'CBC' (compact binary coalescence) or 'Burst'	CBC	0.0%
instruments	str	Comma-separated detectors contributing to the preferred event: H1 (LIGO Hanford), L1 (LIGO Livingston), V1 (Virgo), K1 (KAGRA)	H1,L1,V1	0.0%
n_detectors	int64	Number of detectors in 'instruments'; derived	3	0.0%
labels	str	Comma-separated GraceDB processing labels, e.g. SKYMAP_READY, EMBRIGHT_READY, PASTRO_READY, DQOK, EM_COINC, PE_READY	ADVNO,SKYMAP_READY,EMBRIGHT_READY,PAS...	0.0%
is_low_significance	bool	True when the superevent carries a LOW_SIGNIF_* label, marking it as below the public significant-alert threshold; derived from labels	False	0.0%
has_skymap	bool	True when a sky localization was produced (SKYMAP_READY label); derived from labels	True	0.0%
is_em_ready	bool	True when the alert was released for electromagnetic follow-up (EM_READY label); derived from labels	False	0.0%
preferred_event_id	str	GraceDB identifier of the single-pipeline event chosen to represent the superevent (API field 'graceid')	G328997	0.0%
preferred_event_far_hz	float64	False alarm rate of the preferred event alone in Hz; may differ from the superevent FAR after trials factors are applied	0.00021409731905668	0.0%
reporting_latency_s	float64	Seconds between the candidate signal and the pipeline uploading it to GraceDB; a measure of low-latency alert performance	23.136354	0.0%
n_events	Int64	Number of individual pipeline events grouped into this superevent	2	39.1%
is_offline	bool	True when the preferred event came from an offline (archival) analysis rather than the low-latency online search	False	0.0%
likelihood	float64	Pipeline likelihood or ranking statistic of the preferred event; scale is pipeline-dependent and not comparable across pipelines	1.926275146981	6.4%
submitter	str	GraceDB account that created the superevent record; typically an automated follow-up robot	emfollow	0.0%
em_type	str	Identifier of an associated electromagnetic or neutrino counterpart candidate, when one was recorded	E412926	71.0%
time_coinc_far_hz	float64	False alarm rate in Hz for a time coincidence with an external trigger (e.g. a gamma-ray burst); null when no external coincidence was evaluated	1.1859867769933503e-09	71.0%
space_coinc_far_hz	float64	False alarm rate in Hz for a joint time-and-sky-position coincidence with an external trigger; null when no external coincidence was evaluated	6.477091351157536e-10	71.3%

Quick stats

• 5,546 candidate alerts (2019-04-05 to 2025-11-18)
• 9 online search pipelines, most prolific: CWB
• 80 without a low-significance label
• Median reporting latency: 57.5 s

Usage

from datasets import load_dataset

ds = load_dataset("juliensimon/gw-candidate-alerts", split="train")
df = ds.to_pandas()

import pandas as pd
from datasets import load_dataset

ds = load_dataset("juliensimon/gw-candidate-alerts", split="train")
df = ds.to_pandas()

# Alert latency by pipeline: how fast is each search?
print(df.groupby("pipeline")["reporting_latency_s"].describe()[["count", "50%", "max"]])

# Which candidates became confirmed detections? Join on GPS time (within 1 s).
confirmed = load_dataset("juliensimon/gravitational-wave-events", split="train").to_pandas()
matched = pd.merge_asof(
    df.sort_values("gps_time"),
    confirmed[["name", "gps"]].sort_values("gps"),
    left_on="gps_time", right_on="gps", tolerance=1, direction="nearest",
)
print(matched["name"].notna().sum(), "of", len(df), "candidates match a confirmed event")

# False alarm rate distribution
import matplotlib.pyplot as plt
import numpy as np
fig, ax = plt.subplots(figsize=(10, 5))
ax.hist(np.log10(df["far_hz"].dropna()), bins=60)
ax.set_xlabel("log10(false alarm rate / Hz)")
ax.set_ylabel("Candidates")
ax.set_title("GW candidate significance, O3-O4")
plt.show()

Data source

https://gracedb.ligo.org/

Update schedule

Weekly

Related datasets

• juliensimon/gravitational-wave-events

• juliensimon/black-hole-catalog

• juliensimon/nicer

• juliensimon/maxi

• juliensimon/auger

• juliensimon/crdb

If you find this dataset useful, please consider giving it a like on Hugging Face. It helps others discover it.

About the author

Created by Julien Simon — AI Operating Partner at Fortino Capital. Part of the Space Datasets collection.

Citation

@dataset{gw_candidate_alerts,
  title = {Gravitational-Wave Candidate Alerts},
  author = {Simon, Julien},
  year = {2026},
  url = {https://huggingface.co/datasets/juliensimon/gw-candidate-alerts},
  note = {Derived from GraceDB, LIGO Scientific Collaboration, Virgo Collaboration and KAGRA Collaboration, https://gracedb.ligo.org/},
  publisher = {Hugging Face}
}

License

CC-BY-4.0