Arko007/weathergpt-d1-mos-dataset

WeatherGPT D1 — multi-model NWP forecasts vs. ERA5-Land truth (India)

The training corpus for M2 (distributional bias correction) and M4 (precipitation calibration) in the WeatherGPT project (SIH 2026). One row per (location, valid time): four NWP models' forecasts at that point in time, and what ERA5-Land actually observed there.

M5 (trust ranker) trained a day earlier on a slightly smaller snapshot of this same pipeline (9,507,456 rows vs. this upload's 9,582,912) that was overwritten by a later rebuild before this dataset was captured -- so this is the exact corpus M2 and M4 trained on, and is extremely similar to but not byte-identical to what produced M5. Disclosed here rather than glossed over.

What's in it

• 9,582,912 rows, one parquet file per location, 127 locations across India (real geocoded district headquarters), 127 shards total.
• Source: Open-Meteo's historical-forecast API for the 4 NWP models, and its ERA5-Land archive for truth. Real API calls, real forecast/truth pairs -- not synthetic.

column	meaning
loc_id, lat, lon, elevation_m, admin1	the location
valid_time, hour_utc, doy, month	when this forecast/observation is for
lead_hours, lead_age_days	how far ahead the forecast was made
fc_<variable>_<model>	forecast value, for variable in {temperature_2m, precipitation, wind_speed_10m, relative_humidity_2m} and model in {gfs_seamless, ecmwf_ifs025, icon_seamless, gem_seamless} (16 columns)
truth_<variable>	ERA5-Land observed value for that variable (4 columns)
chunk_misses	how many API chunks failed for this location during the build (counted, not silently dropped)

How M2/M4 actually used it

Loaded whole (modal_jobs/features_d1.py::load_d1), then per target variable (temperature_2m, precipitation, wind_speed_10m):

• Split: chronological 70% cutoff (2026-05-03 02:00:00+00:00) for train/val, AND 20% of locations (25 of 127) held out entirely for a genuinely unseen spatial test set -- so the reported skill isn't just "interpolating a location it already trained on with different timestamps."
• Per-target usable rows (a row is usable only if that variable's truth and all 4 forecasts are present): train 5,388,048, val 2,308,464, test (spatially held out) 565,800 -- identical across all 3 target variables since they're reshaped from the same row set.
• 30 input features per row, built by weathergpt_models/features.py::assemble_features -- the same function training and inference both call, so there's no train/serve skew: the target variable's 4 raw forecasts, 7 ensemble summary stats (mean/sd/min/max/median/spread ratio/wet fraction), mean+sd of the 3 other variables as cross-predictors, 9 time/location features (lead hours, lead age, cyclic hour-of-day and day-of-year, elevation in km, lat, lon), and 4 per-model missingness flags.
• Target: the corresponding truth_<variable>, with precipitation fit in cube-root space (exact quantile-preserving transform, not an approximation) because hourly Indian rainfall is a spike at zero with a long monsoon tail.

Quickstart

import pandas as pd
from huggingface_hub import snapshot_download

path = snapshot_download(repo_id="Arko007/weathergpt-d1-mos-dataset", repo_type="dataset")
import glob
frame = pd.concat(pd.read_parquet(f) for f in glob.glob(f"{path}/d1_mos/*.parquet"))
print(frame.shape, frame.columns.tolist())

Full context

Model cards, held-out metrics and baselines for M2/M4 (trained on this data) are at Arko007/weathergpt-models. Integration and real-world-usability notes are in docs/MODEL_REGISTRY_INTEGRATION.md and docs/REAL_WORLD_READINESS.md.