VertexAIco/nws-weather-plus-eas-30k

NWS Weather + EAS Alerts 30k

30,000 real emergency alerts: 15,000 weather alerts and 15,000 non-weather EAS alerts, kept word-for-word as the issuing authority wrote them. The non-weather half covers AMBER / child abduction, missing and endangered persons, civil emergency and local area emergency messages, evacuation and shelter-in-place orders, law enforcement warnings, fire warnings, hazardous materials, 911 telephone outages, geologic events and more, which is what makes the collection resemble a real EAS feed instead of a weather-only one. Built for text-generation and text-to-speech-oriented training on authentic alert language.

Nothing here was synthesized, paraphrased, improved, summarized or rewritten, and no LLM touched any text.

Versions

version	rows	weather	non-weather	repo
weather-only small	10,000	10,000	0	VertexAIco/nws-weather-only-10k
weather-only large	30,000	30,000	0	VertexAIco/nws-weather-only-30k
EAS 10k	10,000	5,000	5,000	VertexAIco/nws-weather-plus-eas-10k
EAS 30k	30,000	15,000	15,000	VertexAIco/nws-weather-plus-eas-30k
EAS 60k	60,000	35,885	24,115	VertexAIco/nws-weather-plus-eas-60k

Every EAS version is nested: EAS 10k is a subset of 30k, which is a subset of 60k. The weather half of 10k and 30k is a stratified subset of the weather-only large version; the 60k's weather half is that weather-only large version plus 5,885 more weather alerts chosen the same way.

About the sizes. The original plan was 30,000 weather + 30,000 non-weather. After collecting everything available (11,063 NWS-relayed civil products from IEM and 82,222 non-NWS alerts from FEMA IPAWS) and applying the same quality and de-duplication standard as the weather half, only 24,115 unique real non-weather alerts exist, so 30,000 was not reachable. Padding with duplicates or synthetic text would defeat the purpose. So:

• EAS 60k = all 24,115 unique non-weather alerts + 35,885 weather alerts, to reach 60,000 rows (40% non-weather).
• EAS 10k / 30k are exact 50/50 splits.

Sensitive content and responsible use

The non-weather half includes person-identifying alerts, on purpose, so that models can learn the format and language of these messages. Real AMBER, SAFE, Silver, missing-person, blue-alert and law-enforcement messages contain real names, ages, physical descriptions, clothing, vehicle descriptions and license plates, addresses and public phone numbers of missing people, suspected abductors and suspects. These were broadcast publicly by the authorities, but they are still about real people, some of them children and some of them not convicted of anything.

• Treat the data with care; do not use it to identify, locate or contact the people mentioned.
• Do not use models trained on it to fabricate or spoof emergency alerts, or to impersonate an alerting authority.
• Sender email addresses from IPAWS were never collected. Phone numbers and web addresses that appear inside the alert text (for example a police tip line) are kept, because they are part of the message.
• If you need a version without person-identifying alerts, filter on category in (child_abduction_amber, missing_person, law_enforcement).

Sources

Weather half: the Iowa Environmental Mesonet (IEM) NWS Text Product Archive (Iowa State University), exactly as in the weather-only versions (see their cards for full collection, cleaning and de-duplication details).

Non-weather half, two real sources:

1. IEM NWS-relayed civil products: Civil Emergency Message, Civil Danger Warning, Local Area Emergency, Evacuation Immediate, Shelter In Place Warning, Law Enforcement Warning, Child Abduction Emergency, Blue Alert, Hazardous Materials Warning, Fire Warning, Avalanche Watch/Warning, 911 Telephone Outage Emergency, and similar. Issued by NWS offices on behalf of local authorities.
2. FEMA IPAWS Archived Alerts via the public OpenFEMA API: real alerts sent through IPAWS by state and local authorities, filtered server-side to status Actual, message type Alert / Update / Cancel, and not the NWS CAP feed (which is the weather alerts already covered above).

Text products issued by the National Weather Service are works of the US Government. IPAWS alert content is published by FEMA as open government data; check FEMA's terms for your use. Please credit the Iowa Environmental Mesonet (Iowa State University) and FEMA OpenFEMA.

This dataset uses 2008-01-01 to 2026-09-29 (issuance time, UTC). Non-weather alerts are heavily concentrated in recent years (see the year table).

What is in a row

field	description
id	stable id: nws- (IEM rows) or ipaws- (IPAWS rows) + first 12 hex chars of SHA-1 of source_product_id
text	clean message text
raw_text	for IEM rows, the untouched archived product; for IPAWS rows, the alert's event/headline/description/instruction fields as issued, as JSON
alert_class	weather or non_weather_eas
event	weather rows: event name from the VTEC code; IEM civil rows: product name; IPAWS rows: the event string exactly as the sender typed it (inconsistent casing/spelling is real)
product_type	AFOS product category for IEM rows (TOR, WSW, CEM, LAE, ...); IPAWS for IPAWS rows
phenomena, significance, vtec_action, vtec_actions, vtec_event_id	VTEC fields for weather rows; null for non-weather rows
category	weather rows: hazard group (tornado, winter, ...); non-weather rows: alert type (child_abduction_amber, civil_emergency, ...)
cap_category	IPAWS rows: the CAP category (Safety, Other, Fire, Geo, Transport, ...); null otherwise
wfo, cccc	issuing NWS office and WMO id (null for IPAWS rows)
issued	UTC ISO-8601 issue time
year, n_segments, char_count	year, number of $$ segments, length of text
is_followup	weather: VTEC action other than NEW; non-weather: IPAWS Update or Cancel
source_product_id	IEM product id (resolves via https://mesonet.agron.iastate.edu/api/1/nwstext/<id>?nolimit=1) or the IPAWS record id
source_url	permalink to the record at its source
source	IEM NWS Text Product Archive or FEMA IPAWS Archived Alerts (OpenFEMA)

Non-weather methodology

IEM civil rows are cleaned with the same removal-only rules as the weather text (transmission header, UGC/VTEC lines, &&/LAT...LON trailers, $$, forecaster sign-offs), plus removal of trailing alert-vendor tracking-id lines and name sign-offs.

IPAWS rows: text is the alert's own headline, description and instruction fields verbatim, joined by blank lines (a field that repeats another is dropped). When those are empty, the longest free-text CAP parameter is used, because that is where WEA messages keep their text. Nothing else is altered besides removing control characters, trailing whitespace and a trailing vendor tracking-id line.

Excluded (never replaced with synthetic text): tests, demos, drills and exercises; records with no English info block; alerts whose CAP category is meteorological; alerts that are really weather (event name or text mentions two or more distinct weather phenomena, a heuristic that also catches weather relayed as a civil message); messages under 60 characters or 10 words; messages over 6,000 characters; text that is mostly non-alphabetic.

De-duplication: exact text, then near-duplicates (normalized word 5-gram MinHash, cutoff 0.93, the same as the weather half), then cross-source duplicates where an IPAWS message already appears inside an IEM relay of the same alert (IEM copy kept).

Alert-type labels (category) are assigned from the event name (then the opening text) by keyword rules, so a few borderline messages are labeled other_non_weather or a neighbouring type.

Selection: the EAS 60k version contains every unique non-weather alert. Smaller versions are nested, stratified subsets in which smaller alert types are deliberately over-represented relative to their share, so the 10k and 30k versions cover all alert types rather than mostly the largest few.

Non-weather pipeline counts

• IEM civil products collected: 11,063; IPAWS records scanned: 82,222
• candidates after quality filters: 30,072
• removed as exact duplicates: 3,043; near duplicates: 2,391; cross-source duplicates: 523
• unique non-weather alerts available: 24,115; in this version: 15,000

non-weather rejection reason	records
ipaws_test_or_demo	50100
ipaws_too_short_or_empty	5349
ipaws_no_info	3720
iem_weather_content	2910
ipaws_meteorological_category	481
ipaws_weather_content	318
iem_test_or_demo	131
ipaws_clutter	126
iem_truncated_no_terminator	28
ipaws_mostly_non_text	26
iem_clutter	16
iem_too_short_or_empty	3
iem_too_long	3
iem_mostly_non_text	2

Near-duplicate check on the 15,000 non-weather rows in this version:

similarity threshold	non-weather rows with a neighbour at/above it
0.80	1262
0.85	901
0.90	338
0.92	143
0.93	0
0.95	0

Statistics

Rows: 30,000 (15,000 weather + 15,000 non-weather) | years: 19 | follow-up rows: 5,846. Non-weather text length (chars): median 345, mean 474, max 4489. Weather text median 1377.

Non-weather alert types

non-weather alert type	examples
civil_emergency	3837
child_abduction_amber	3156
evacuation_shelter	1758
fire	1620
law_enforcement	1501
administrative_message	812
telephone_911_outage	795
missing_person	665
earthquake_geologic	452
hazmat_radiological	151
transportation_infrastructure	118
other_non_weather	102
public_health	33

Non-weather sources

source	examples
FEMA IPAWS Archived Alerts (OpenFEMA)	11183
IEM NWS Text Product Archive	3817

Non-weather events as issued

event name as issued (top 25)	examples
Child Abduction Emergency	3066
Local Area Emergency	2402
Fire Warning	1484
Law Enforcement Warning	1386
Civil Emergency Message	1116
Evacuation Immediate	1113
Administrative Message	761
911 Telephone Outage Emergency	710
Avalanche Warning	336
Shelter in Place Warning	326
Missing and Endangered Persons	266
SAFE Alert Activation	154
Hazardous Materials Warning	124
Local Area Emergency Warning	110
Avalanche Watch	103
Civil Danger Warning	91
Silver Alert Activation	87
Shelter In Place Warning	76
Evacuation Immediate Warning	74
child abduction emergency	46
Administrative Message/Follow up Statement	44
Immediate Evacuation	42
LAE	33
missing and endangered persons	33
local area emergency	29

IPAWS CAP categories

CAP category (IPAWS rows)	examples
Safety	5569
(IEM relay, no CAP category)	3817
Other	2193
Fire	1387
Transport	663
Geo	409
Rescue	304
Security	196
Health	188
Infra	138
Env	77
CBRNE	32
Safety	Security
Safety	Health
Safety	Fire
Health	Env
Fire	Safety
Health	Infra
Safety	Infra
Safety	CBRNE
Safety	Geo
Other	Safety
Safety	Infra
Safety	Env
Safety	Infra
Safety	Rescue

Non-weather examples per year

year	non-weather examples
2008	230
2009	190
2010	189
2011	302
2012	241
2013	305
2014	329
2015	322
2016	383
2017	463
2018	421
2019	476
2020	882
2021	867
2022	932
2023	1555
2024	2087
2025	2368
2026	2458

Weather hazard categories

weather hazard category	examples
winter	2212
extreme_temperature	1714
marine	1571
flood	1406
coastal_surf	1323
severe_thunderstorm	1123
tornado	1084
flash_flood	1017
special_weather_statement	936
dust_visibility	787
wind	667
tropical	606
fire_weather	478
other_rare	76

Full counts (per event, per year, per office, per VTEC combination) are in stats/stats.json.

Known limitations

• Recency skew. Most non-weather alerts are from 2023-2026, because IPAWS use and archiving grew and older IEM civil volume is small.
• Template-heavy categories. Local Area Emergency and law-enforcement messages from a few agencies reuse phrasing; near-duplicates were removed but repeated style remains, as it does in reality.
• Local Area Emergency is used loosely by senders (road closures, incidents, boil notices); category labels are keyword-based and imperfect.
• The weather-content filter is a heuristic; a few real non-weather alerts that mention weather may have been dropped, and a few weather-adjacent ones may remain.
• IPAWS event strings are as typed by senders (LAE, Local Area Emergency Warning, misspellings). Cancellations and updates are kept as follow-up rows (is_followup).
• Person-identifying content is present by design (see above), and other details such as names may also appear in ordinary local-emergency text.
• Weather-half limitations (style change over time, hard line wraps, county name lines, multi-segment products, etc.) are described on the weather-only dataset cards.

Loading

from datasets import load_dataset
ds = load_dataset("VertexAIco/nws-weather-plus-eas-30k", split="train")
eas = ds.filter(lambda r: r["alert_class"] == "non_weather_eas")

Files

• data/train.parquet - the dataset (30,000 rows)
• jsonl/nws_eas_alerts_30k.jsonl - the same rows as JSON Lines
• stats/stats.json - full statistics and validation results
• build_scripts/ - the collection, cleaning, de-duplication, selection and validation code

Citation and acknowledgement

Iowa Environmental Mesonet, Iowa State University (NWS Text Product Archive); FEMA OpenFEMA, IPAWS Archived Alerts; original text by the US National Weather Service and the issuing state and local authorities.