VertexAIco/nws-weather-only-30k

NWS Weather-Only Alerts 30k (large)

30,000 unique, real National Weather Service alert products (warnings, watches, advisories and statements), cleaned of machine-only formatting but otherwise word-for-word as the NWS wrote them. Built to preserve authentic NWS writing style for text-generation and text-to-speech-oriented training.

This is the large version (30,000 rows). It contains every row of the small version (VertexAIco/nws-weather-only-10k, 10,000 rows) plus about 20,000 more, with much heavier coverage of rare and uncommon alert types. Versions that add non-weather EAS alerts (AMBER, civil emergency, evacuation, law enforcement, ...) are nws-weather-plus-eas-10k, nws-weather-plus-eas-30k and nws-weather-plus-eas-60k. The weather-only versions here contain no non-weather alerts.

Nothing here was synthesized, paraphrased, improved, modernized, summarized or rewritten, and no LLM touched any text. Every text value is an archived NWS product with routing/encoding lines removed; every raw_text value is the untouched archived product.

Source

All records come from the Iowa Environmental Mesonet (IEM) NWS Text Product Archive (Iowa State University), retrieved through its public web services. This is a processed subset of that archive, not the whole of it.

• VTEC archive documentation: https://mesonet.agron.iastate.edu/info/datasets/vtec.html
• IEM web services: https://mesonet.agron.iastate.edu/api/
• IEM datasets index: https://mesonet.agron.iastate.edu/info/datasets/

Text products issued by the US National Weather Service are works of the US Government and are not subject to US copyright. Please credit the Iowa Environmental Mesonet, Iowa State University, as the archive source.

Note: the IEM documentation describes the period from 2008 onward as having very good archive coverage. This dataset uses 2008-01-01 to 2026-09-29 (issuance time, UTC). Records for 2026 are partial-year.

What is in a row

field	description
id	stable id: nws- + first 12 hex chars of SHA-1 of source_product_id
text	clean product text (see cleaning below)
raw_text	original archived product, exactly as IEM returned it (includes header/routing lines)
event	human-readable event name from the primary VTEC code (e.g. "Tornado Warning"); "Special Weather Statement" for SPS
product_type	3-letter AFOS product category (TOR, SVR, SVS, FFW, WSW, NPW, SPS, WCN, MWW, ...)
phenomena	VTEC phenomena code (e.g. TO, WS); null for products without VTEC (SPS)
significance	VTEC significance (W warning, A watch, Y advisory, S statement); null for SPS
vtec_action	VTEC action of the product's first operational VTEC line (NEW, CON, EXT, EXA, EXB, UPG, CAN, EXP, COR)
vtec_actions	all distinct operational VTEC actions in the product, comma-joined
vtec_event_id	lifecycle key WFO.PHENOMENA.SIG.ETN.YEAR of the primary VTEC event
category	hazard group used for sampling (tornado, winter, marine, ...)
wfo	issuing NWS office (3-letter id; almost all are Weather Forecast Offices, plus a few products from national centers such as NHC and the tsunami warning centers)
cccc	4-letter WMO source id
issued	product entered time, UTC ISO-8601
year	issuance year
n_segments	number of $$-delimited segments in the product
char_count	length of text
is_followup	true if the primary VTEC action is anything other than NEW
source_product_id	IEM product id (YYYYMMDDHHMM-CCCC-TTAAII-PIL); resolves via https://mesonet.agron.iastate.edu/api/1/nwstext/<id>
source_url	permalink to the product on IEM
source	"IEM NWS Text Product Archive"

Cleaning (removal only)

The text field is produced by deleting machine-oriented content. No wording is ever changed, reflowed or rephrased.

Removed:

• the transmission char-count line, WMO header line and AWIPS/PIL routing line
• UGC zone/county code blocks (e.g. IAZ005>007-015-010415-)
• P-VTEC and H-VTEC lines
• the machine trailer of each segment: everything from the && end-of-text marker (or, when a product has none, the first LAT...LON line) up to the segment's $$. This holds LAT...LON, TIME...MOT...LOC and impact-tag lines (MAX HAIL SIZE... etc.), which are not part of the text meant to be read out
• $$ delimiters, NNNN, FORECASTER: NAME lines, and the short forecaster sign-off after the final $$
• trailing spaces and runs of blank lines

Preserved exactly: bulletin/headline wording, overview text, county and place wording, issuance and expiration wording, hazard/source/impact text, preparedness and call-to-action text (e.g. TAKE COVER NOW!), punctuation, original capitalization, and the products' hard line wraps.

Sampling methodology

1. Collection. IEM's bulk AFOS endpoint was queried over randomly chosen date windows (seasonally weighted, balanced across 2008-2026) for 22 product families, plus the targeted full-history crawl of rare-hazard families described below, giving 349,181 unique raw products. A large pool was needed to keep enough distinct storm days, years and offices after the lifecycle and duplicate filters.
2. Cleaning and quality filters (see rejection table) left 335,171 candidates.
3. Deduplication left 270,261 (details below).
4. Nested, stratified selection. The small version was selected first (quotas per hazard category x VTEC significance, with per-office, per-year and per-day caps). The large version then adds a rare-first block and a second round of quotas (3x the small quotas) on top, under proportionally larger caps, and any shortfall in a stratum is redistributed dynamically to the other strata. Nothing was fabricated to fill a quota.
5. Lifecycle policy. NEW (initial issuance) products are preferred; 31% of the dataset (9,168 rows) is deliberately follow-up products (statements, continuations, extensions, cancellations) because follow-up wording is itself a distinct NWS style. Each VTEC event contributes at most one NEW product plus at most one follow-up.

Watch coverage comes from WFO-issued VTEC products (including Watch County Notifications for tornado/severe thunderstorm watches).

Rare-alert coverage (what makes the large version different)

A random sample of NWS alerts is dominated by a few very common products (severe thunderstorm warnings, special weather statements, flash flood warnings...). To give the large version real depth on uncommon hazards:

• Targeted full-history crawl. For the product families where uncommon alerts live (non-precipitation, winter, marine, coastal, tropical, tsunami, dust, snow squall, extreme wind...), every window from 2008 onward was fetched and scanned, and uncommon alert types were kept (hash-sampled so that no single type contributes more than about 3,000 candidates).
• Rare-first allocation. After the small version's rows were selected, every uncommon (VTEC phenomena, significance) combination was given a target of 220 rows (or all it has, if it has fewer) before the general quotas filled the remainder.
• The dataset covers 114 distinct (phenomena, significance) combinations, and 17,204 rows (57%) come from combinations that are uncommon in a random sample. Counts for every combination are in stats/stats.json (by_phenomena_significance).
• Some alert types simply have very few real examples (for example tsunami warnings, hurricane warnings, storm surge products); those are included in full rather than padded.

Deduplication

Applied in this order:

1. Duplicate PROD_IDs removed: 14,002.
2. Exact duplicates: cleaned text normalized (whitespace collapsed, case-folded) and hashed: 2 removed.
3. Warning-lifecycle duplicates (more than one NEW, or more than one follow-up, for the same VTEC event): 58,495 removed.
4. Near-duplicates: 6,413 removed. Each cleaned text was normalized for comparison only (times, dates, weekday/month names and numbers masked; county-name lines dropped; place names masked using corpus statistics), split into word 5-gram shingles, and compared with MinHash (128 permutations) + LSH, with each candidate hit confirmed by an exact MinHash Jaccard estimate. Cutoff: 0.93. Inspecting flagged pairs showed this removes the same alert re-issued or corrected while keeping different alerts that merely share NWS boilerplate (for example two Tulsa tornado warnings in different months score about 0.88 and are both kept).

Near-duplicate check on the final 30,000 (nearest-neighbour similarity, normalized text):

similarity threshold	final-set docs with a neighbour at/above it
0.80	1234
0.85	806
0.90	314
0.92	167
0.93	0
0.95	0

Statistics

Rows: 30,000 | offices: 128 | years: 19 | follow-up products: 9,168 | text length (chars): min 259, median 1376, mean 1922, max 8995.

Per hazard category

hazard category	examples
winter	4426
extreme_temperature	3426
marine	3139
flood	2813
coastal_surf	2649
severe_thunderstorm	2246
tornado	2167
flash_flood	2035
special_weather_statement	1871
dust_visibility	1573
wind	1335
tropical	1213
fire_weather	956
other_rare	151

Per VTEC significance

VTEC significance	examples
W	14071
Y	7323
A	5738
none (SPS)	1871
S	997

Per primary VTEC action

primary VTEC action	examples
NEW	18961
CON	4257
CAN	2105
none (SPS)	1871
EXP	1130
UPG	800
EXT	599
EXA	123
EXB	71
ROU	56
COR	27

Per product type

AFOS product type	examples
NPW	5231
WSW	4095
MWW	3765
CFW	2649
SPS	1871
SVR	1476
TOR	1441
FFW	1379
FLS	1187
FLW	1174
RFW	956
SVS	821
FFA	790
WCN	674
HLS	548
SQW	426
DSW	420
SMW	387
FFS	318
TCV	215
TSU	114
EWW	63

Per event

event (top 60 of 114)	examples
Severe Thunderstorm Warning	1891
Special Weather Statement	1871
Tornado Warning	1848
Flash Flood Warning	1697
Flood Warning	864
Areal Flood Advisory	658
Red Flag Warning	582
Areal Flood Warning	548
High Wind Warning	433
Snow Squall Warning	426
Winter Weather Advisory	414
Special Marine Warning	387
Winter Storm Watch	384
Red Flag Watch	374
Winter Storm Warning	374
Tropical Storm Warning	365
Severe Thunderstorm Watch	355
Flash Flood Watch	338
Tornado Watch	319
Frost Advisory	280
High Wind Watch	276
Marine Dense Fog Advisory	254
Small Craft Advisory	247
Ice Storm Warning	246
Wind Advisory	244
Blizzard Warning	243
Gale Warning	242
Flood Advisory	232
Areal Flood Watch	232
Tropical Storm Watch	230
Lake Wind Advisory	228
Freeze Warning	227
Freeze Watch	226
Small Craft for Hazardous Seas Advisory	226
Storm Warning	226
Dust Storm Warning	224
Low Water Advisory	224
Coastal Flood Warning	223
Gale Watch	223
Small Craft for Winds Advisory	223
Ice Accretion Warning	222
Cold Weather Advisory	222
Small Craft for Rough Bar Advisory	222
High Surf Advisory	221
Coastal Flood Advisory	221
Coastal Flood Statement	221
Freezing Rain Advisory	221
Hazardous Seas Warning	221
Rip Current Statement	221
Lakeshore Flood Advisory	221
Beach Hazard Statement	221
Excessive Heat Warning	221
Hard Freeze Watch	221
Lakeshore Flood Statement	220
Blowing Dust Advisory	220
Extreme Heat Warning	220
Extreme Cold Warning	220
Snow Advisory	220
Storm Watch	220
Heavy Snow Warning	220

Per year

year	examples
2008	2287
2009	1567
2010	1614
2011	1624
2012	1356
2013	1451
2014	1276
2015	1410
2016	1209
2017	1415
2018	1400
2019	1836
2020	1583
2021	1255
2022	1472
2023	1538
2024	1726
2025	1902
2026	2079

Per office

office (10 most represented)	examples
DTX	501
LOT	451
BUF	441
PQR	436
MFR	433
MQT	418
IWX	417
BOX	414
CLE	396
GYX	385

Full per-office, per-event, per-year and per-(phenomena, significance) counts are in stats/stats.json.

Pipeline counts

• unique raw products collected: 349,181
• duplicate PROD_IDs dropped: 14,002
• rejected during cleaning/quality filtering: 14,010
• candidates after cleaning: 335,171
• removed as exact duplicates: 2
• removed as lifecycle duplicates: 58,495
• removed as near duplicates: 6,413
• deduplicated pool: 270,261
• final: 30,000

Cleaning/quality rejections (records dropped, never replaced with synthetic text):

cleaning rejection reason	products
no_operational_vtec_and_not_sps	6380
too_long	5560
too_short_or_no_narrative	1192
mostly_non_text	406
test_product	228
truncated_no_terminator	208
not_alert_significance	23
clutter_survived_cleaning	13

Quality filters: test products; no operational VTEC (other than SPS); VTEC significance other than warning/watch/advisory/statement (outlooks, forecasts, synopses); no $$ terminator (truncated/corrupt); fewer than 250 characters or 40 words after cleaning; more than 9,000 characters; text that still contained machine clutter after cleaning; mostly non-alphabetic content.

Known limitations

• Not a random sample of all NWS alerts. It is stratified for breadth, so it over-represents rare hazards relative to their real-world frequency (the large version does so far more deliberately).
• Style changes over time. Products before roughly 2011 are mostly ALL CAPS; newer ones are mixed case. Formats (bullet-style "WHAT/WHERE/WHEN/IMPACTS" messages, impact-based warnings) also evolved, and both are represented.
• Hard line wraps are kept (about 66 columns) since re-wrapping would alter the archived text. Unwrap them yourself if your use needs single-line paragraphs.
• County/zone name lines (for example Emmet-Kossuth-Winnebago-) and "Including the cities of..." lines are kept because they are human-readable location wording.
• Watch County Notification (WCN) rows are largely county lists rather than prose; tornado and severe-thunderstorm watch rows come from these. SPC-issued watch text (SEL) and NHC advisories (TCP/TCD) are not included.
• Multi-segment products (a single product covering many zones, e.g. winter storm products) are kept whole as a single row up to 9,000 characters; longer ones were dropped, which slightly under-represents very large multi-zone products.
• VTEC fields describe the first operational VTEC line; a multi-segment product can contain several different events and actions (see vtec_actions).
• Forecaster sign-offs (names/initials after the final $$) were removed from text but remain in raw_text.
• Scarce categories (tropical, tsunami, and some extreme-weather and rare-phenomena products) have fewer unique records than their target quota; the shortfall went to other categories (see below).
• Non-NWS civil EAS messages (AMBER alerts, local civil emergency messages relayed by authorities, etc.) are generally outside the scope of this corpus.
• Rarely, IEM holds two different products under one product id (for example an original and a CCA correction issued in the same minute). Each id appears once here, and IEM's one-shot nwstext/<id> call may return the other version; add ?nolimit=1 to see both.
• The IEM archive can contain gaps or re-stored products; source_product_id lets you check any row against the live archive.
• Quota shortfalls redistributed during selection: {"other_rare|W": 130, "other_rare|Y": 58, "tropical|W": 100, "tropical|A": 69}.

Loading

from datasets import load_dataset
ds = load_dataset("VertexAIco/nws-weather-only-30k", split="train")
print(ds[0]["text"])

Files

• data/train.parquet - the dataset (30,000 rows)
• jsonl/nws_alerts_30k.jsonl - the same rows as JSON Lines
• stats/stats.json - full statistics and validation results
• build_scripts/ - the collection, cleaning, dedup, selection and validation code used to build it

Citation and acknowledgement

Source archive: Iowa Environmental Mesonet, Iowa State University, NWS Text Product Archive (https://mesonet.agron.iastate.edu/). Original text: US National Weather Service.