"""Download NOAA GHCN-Daily rows for the 210 challenge stations, SCORING ONLY. Post-2018 observations are used only to score a frozen forecast, never as a forecast input (user-approved 2026-11-07 because the private answer key is not accessible). Keeps 2018-2021 TMAX/TMIN/PRCP (tenths of degC / tenths of mm) per station in one CSV.""" import io, sys, urllib.request, concurrent.futures as cf from pathlib import Path import pandas as pd from huggingface_hub import hf_hub_download URL = "https://www.ncei.noaa.gov/data/global-historical-climatology-network-daily/access/{}.csv" ids = sorted(pd.read_parquet(hf_hub_download("CMU-OSPO/protostar-climate-train-to-2018", "annual_indices/train-00000-of-00001.parquet", repo_type="dataset"), columns=["ID"]).ID.unique()) def rng(url, a): return urllib.request.urlopen(urllib.request.Request(url, headers={"Range": f"bytes={a}"}), timeout=60).read().decode() def get(i): u = URL.format(i) header = rng(u, "0-8000").split("\n")[0] tail = 2_500_000 while True: body = rng(u, f"-{tail}") body = body[body.index("\n") + 1:] df = pd.read_csv(io.StringIO(header + "\n" + body), dtype=str, usecols=lambda c: c in ("DATE", "TMAX", "TMIN", "PRCP")) if df.DATE.min() <= "2018-01-01" or tail > 40_000_000: break tail *= 2 df = df[(df.DATE >= "2018-01-01") & (df.DATE <= "2021-12-31")] for c in ("TMAX", "TMIN", "PRCP"): df[c] = pd.to_numeric(df[c], errors="coerce") if c in df else float("nan") df.insert(0, "ID", i) return df with cf.ThreadPoolExecutor(8) as ex: parts = list(ex.map(get, ids)) out = pd.concat(parts, ignore_index=True) Path(sys.argv[1]).parent.mkdir(parents=True, exist_ok=True) out.to_csv(sys.argv[1], index=False) print(out.shape, out.ID.nunique(), out.DATE.min(), out.DATE.max())