TailWeather label data

The label data itself (NetCDF files, ~106 GB total across all hazards and compounds) is not stored in this code repository. It is hosted in two places that serve different purposes:

• Primary, practical access — Hugging Face Hub: https://huggingface.co/datasets/zhisong-liu/TailWeather. This is the full archive (all hazards, all years, regional subsets), where anyone training or evaluating an AI weather model on TailWeather should get it from — git-lfs/Xet-backed, no practical size limit, versioned, with HF's built-in dataset viewer. The same repo also has a small demo/ subfolder (see below) for browsing without a full download. The YAML front matter above is meant to be used directly as that repo's README.md (its dataset card).
• Companion citation-of-record — Zenodo: [DOI to be added]. A small, permanently archived record containing only this documentation (this file plus docs/LABELING_METHODOLOGY.md) and an explicit link to the Hugging Face repo — no bundled data files, since the illustrative demo now lives on HF where it can actually be browsed interactively. Cite this DOI as the archival reference for the dataset; use the HF repo to obtain the data itself.

This file documents the directory layout, file naming, and every variable in the archived data, so it is usable standalone in either location without needing the labeling code in this repository (though that code is what produced it, and can regenerate it from ERA5 — see the top-level README.md of the code repository).

Illustrative demo subset (demo/ on the HF repo)

A ~275 MB hand-picked cross-section, meant to be browsed via HF's dataset viewer or a quick download before committing to the full 106 GB archive:

Path	Size	What it shows
demo/events/	13 MB	50 individually-cropped, named case-study events (Germany/Nordics, 2020–2023) with events_catalogue.csv indexing hazard/region/date — e.g. the well-known summer 2022 European heatwave
demo/regional/	71 MB	Full heatwave/precip time series for two small regional grids (Nordics, Germany)
demo/global_sample/	220 MB	One complete global year (2003 — the historic European heatwave) at full 721×1440 resolution, illustrating the real archive's format at a manageable size

Staged locally at /scratch/project_462001157/Extreme_weather/hf_demo_subset/ (with its own README.md), ready to upload as a subfolder of the main repo:

hf upload zhisong-liu/TailWeather \
    /scratch/project_462001157/Extreme_weather/hf_demo_subset demo \
    --repo-type dataset

Uploading / downloading the full archive via Hugging Face

# Upload (once, from wherever the full labels/ directory currently lives) --
# excludes the ~450 GB of disposable intermediate CDF caches, see below
hf upload zhisong-liu/TailWeather \
    /scratch/project_462001157/Extreme_weather/labels . \
    --repo-type dataset \
    --exclude "*_cdf_memmap.dat" "*_precip_cdf.dat" "*_precip_spell_cdf.dat"

# Download (for a downstream user)
hf download zhisong-liu/TailWeather --repo-type dataset --local-dir ./labels

Keep the directory layout below unchanged inside the HF repo, so paths in any code (including this repository's own analysis scripts, which take --label-glob/--label-dir arguments) work unmodified after download.

What is (and isn't) archived

Included — the actual label products, per hazard:

Hazard	Files	Approx. size
Heatwave	heatwave_<year>.nc (1981–2023) + heatwave_threshold.nc + run_log.csv	~9.4 GB
Cold wave	coldwave_<year>.nc + coldwave_threshold.nc + run_log.csv	~10 GB
Heavy precipitation	precip_<year>.nc + precip_p95_wetday.nc + precip_p75_spell.nc + run_log.csv	~19 GB
Drought (SPI-3)	drought_<year>.nc (monthly) + drought_daily_<year>.nc (broadcast)	~22 GB
Extreme wind	wind_<year>.nc + wind_threshold.nc	~29 GB
Compound: hot-dry	hotdry_<year>.nc	~3.8 GB
Compound: hot-dry-windy	hotdry_windy_<year>.nc	~3.6 GB
Compound: cold-wet	cold_wet_<year>.nc	~4.6 GB
Compound: wind-wet	wind_wet_<year>.nc	~4.6 GB
Regional subsets (Nordics, Germany)	per-hazard, per-year	~71 MB
Case-study extracts	small JSON/NetCDF extracts used in the paper's case-study figure	~13 MB

See "Illustrative demo subset" above for the ~275 MB browsable cross-section on Hugging Face — the Zenodo record intentionally stays documentation-only.

Excluded everywhere — intermediate caches, regenerable from the code, not useful downstream:

• _cdf_memmap.dat (heatwave, coldwave, wind — 150 GB each) and _precip_cdf.dat / _precip_spell_cdf.dat (411 MB each) are raw memmap caches of the empirical CDF used only during label computation — ~450 GB combined, by far the largest thing on disk, carrying no information not already in the label files. Regenerate via --stage clim if you need to recompute intensity scores at custom percentiles.

Not included and not redistributable — obtain independently:

• ERA5 input data (WeatherBench2 / Copernicus CDS, own terms).
• EM-DAT, GDIS, NOAA Storm Events impact-catalogue records used only for validating the labels (code/analysis/concordance/), not for producing them — see that directory's README for sources and terms.

Directory layout

labels/
├── heatwave/
│   ├── heatwave_threshold.nc        P90 climatology threshold (Stage 1 output)
│   ├── heatwave_1981.nc             per-year global labels
│   ├── ...
│   ├── heatwave_2023.nc
│   └── run_log.csv                  per-year summary statistics
├── coldwave/            (mirrors heatwave/, P10 lower-tail)
├── precip/
│   ├── precip_p95_wetday.nc         R95p wet-day threshold (Stage 1 output)
│   ├── precip_p75_spell.nc          P75 spell-length threshold (Stage 1 output)
│   ├── precip_1981.nc ... precip_2023.nc
│   └── run_log.csv
├── drought/
│   ├── drought_1981.nc ... drought_2023.nc     (monthly, native)
│   └── drought_daily_1981.nc ... drought_daily_2023.nc   (daily-broadcast)
├── wind/
│   ├── wind_threshold.nc            P98 climatology threshold
│   └── wind_1981.nc ... wind_2023.nc
├── hotdry/, hotdry_windy/, cold_wet/, wind_wet/   (compound events, same per-year file pattern)
└── regional/
    ├── nordics/{heatwave,precip}/<hazard>_nordics_<year>.nc
    └── germany/{heatwave,precip}/<hazard>_germany_<year>.nc

Coordinates (all files)

Coordinate	Description
time	Daily timestamps (UTC calendar day), datetime64. Monthly for the native drought file.
latitude	Degrees north, ascending, 0.25° spacing, −90 to 90
longitude	Degrees east, 0.25° spacing, 0 to 359.75

Note the dimension names are latitude/longitude, not lat/lon.

Per-hazard variables

heatwave_<year>.nc / coldwave_<year>.nc

Variable	Type	Dims	Units	Description
binary_label	int8	time, lat, lon	0/1	1 = event day
intensity_score	float32	time, lat, lon	[0,1]	percentile rank in the 1991–2020 climatology CDF; 1.0 = historical extreme; NaN off-event/ocean. For cold wave the rank is inverted (colder = higher score).
hw_duration / cw_duration	int16	time, lat, lon	days	running consecutive-day streak length
hw_severity / cw_severity	int8	time, lat, lon	0–3	ordinal tier (added by add_severity.py)

heatwave_threshold.nc / coldwave_threshold.nc: T90/T10 per (dayofyear, lat, lon) in °C.

precip_<year>.nc

Variable	Type	Dims	Units	Description
wet_day	int8	time, lat, lon	0/1	daily precip ≥ 1 mm
wet_spell_id	int16	time, lat, lon	—	sequential wet-spell index within the year (0 = dry/isolated)
wet_spell_len	int16	time, lat, lon	days	length of the wet spell this day belongs to
heavy_precip	int8	time, lat, lon	0/1	wet day AND precip > local R95p (wet-day-only percentile)
heavy_long_spell	int8	time, lat, lon	0/1	heavy_precip AND spell length ≥ local P75 (Zolina et al. 2010 compound diagnostic)
intensity_score	float32	time, lat, lon	[0,1]	wet-day percentile rank; NaN on dry days/ocean
precip_mm	float32	time, lat, lon	mm/day	daily total precipitation
precip_severity	int8	time, lat, lon	0–3	ordinal tier

precip_p95_wetday.nc: R95p threshold (mm/day) per cell. precip_p75_spell.nc: long-spell threshold (days) per cell.

drought_<year>.nc (SPI-3, McKee et al. 1993)

Variable	Type	Dims	Units	Description
spi3	float32	time(month), lat, lon	—	SPI-3; negative = drier; ≈N(0,1) over the reference period
binary_label	int8	time, lat, lon	0/1	1 = SPI-3 ≤ −1.0
intensity_score	float32	time, lat, lon	[0,1]	drought depth, clip(−SPI/3, 0, 1); NaN off-drought/ocean
drought_severity	int8	time, lat, lon	0–3	ordinal tier (McKee thresholds −1.0/−1.5/−2.0)

drought_daily_<year>.nc broadcasts the monthly label to every day of that month — a modelling convenience, not genuine daily resolution.

wind_<year>.nc

Variable	Type	Dims	Units	Description
binary_label	int8	time, lat, lon	0/1	daily-max wind > local P98
intensity_score	float32	time, lat, lon	[0,1]	percentile rank; NaN off-event/ocean
wind_speed	float32	time, lat, lon	m/s	daily maximum 10 m wind (gust if available, else sqrt(u10²+v10²))
wind_severity	int8	time, lat, lon	0–3	WMO Beaufort-band tier, computed in-pipeline

wind_threshold.nc: P98 threshold (m/s) per (dayofyear, lat, lon).

Compound events (hotdry_<year>.nc, hotdry_windy_<year>.nc, cold_wet_<year>.nc, wind_wet_<year>.nc)

Same grid/time convention; each is an intersection of two constituent single-hazard labels (e.g. hot-dry = heatwave AND drought at the same cell/day) with a compound severity tier and intensity score derived from the constituents — see docs/LABELING_METHODOLOGY.md for the exact combination rule used for each compound.

Regional files

Direct spatial subsets of the global files (Nordics: 52–73°N, 0–43°E, 85×173 cells; Germany: 46–56°N, 4–17°E, 41×53 cells). Same variables, coordinates and attributes as the global files, plus region, region_bbox, grid_size global attributes.

Key methodological notes

• In-sample labeling: years within the 1991–2020 reference period are labelled against that same period's climatology (standard in the literature, e.g. Perkins & Alexander 2013) — ~10% of land cell-days within the reference period exceed the heatwave P90 threshold by construction.

• Precipitation de-accumulation: ERA5 accumulated precipitation is correctly de-accumulated via WeatherBench2's total_precipitation_6hr, avoiding the 00/12 UTC reset-boundary error from naive differencing.

• Full definitions, formulas, and references for every hazard are in ../docs/LABELING_METHODOLOGY.md.

  To cite this data, please use the following bibtex

    @misc{liu2026tailweathertailextremesglobal,
        title={TailWeather: from tail to extremes, a global climatological dataset for machine-learning weather forecasting}, 
        author={Zhi-Song Liu and Michael Boy and Risto Makkonen},
        year={2026},
        eprint={2609.12496},
        archivePrefix={arXiv},
        primaryClass={cs.CE},
        url={https://arxiv.org/abs/2609.12496}, 
    }