TyphoonUQ-Bench

TyphoonUQ-Bench v1 — derived data release. This deposit ships the derived parquets that drive the TyphoonUQ-Bench benchmark: an aligned multi-agency manifest, the canonical chronological split, an auxiliary leave-one-agency-out split, and pre-built per-sample tabular caches for three tasks. The bundle is ~38 MB and contains no raw imagery, no best-track text files, and no model checkpoints. Users who only need to score an external UQ model on the canonical splits can use this bundle alone — no Digital Typhoon download is required for the metadata-only baseline path. What's inside benchmark/ (8.3 MB) — benchmark_manifest.parquet (180,363 rows × 229 columns), three match tables linking Digital Typhoon ↔ JMA ↔ IBTrACS storm IDs, a per-agency coverage report, and a small CSV of manually reviewed cases. splits/ (3.5 MB) — forward_main.parquet, the canonical year-based train/val/test split (train 1988–2015, val 2016–2019, test 2020–2023), plus leave_one_agency_out.parquet (4× expansion holding out JMA / JTWC / CMA / HKO in turn). sample_cache/ (~26 MB) — per-task sample caches with a 6-step history window: analysis-0h (58,813 samples), forecast-6h (57,010 samples), forecast-12h (55,204 samples). Each row carries 48 tabular history features (8 features × 6 time steps) plus consensus target, agency metadata, and split label. Quick start import pandas as pd manifest = pd.read_parquet("benchmark/benchmark_manifest.parquet") splits = pd.read_parquet("splits/forward_main.parquet") samples = pd.read_parquet("sample_cache/forecast-12h_png_h6_samples.parquet") print(manifest.shape, splits['split'].value_counts()) To use the bundle with the GitHub repository, unzip and drop the three folders into data/processed/benchmark/, outputs/splits/, and outputs/splits/sample_cache/ respectively. After that, training and evaluation scripts run without re-executing the alignment or split-construction stages. Upstream sources The manifest is built against snapshots dated 2024-12-31 of: Digital Typhoon (DIAS, CC-BY-4.0), JMA RSMC best-track bst_all.txt, and IBTrACS v04r01 (NOAA/NCEI). Attribution to each upstream source is required on redistribution — see LICENSES.md in the bundle for the per-source citation block. What this bundle does NOT contain Raw Digital Typhoon PNG / HDF5 imagery (~120 GB, from DIAS) Raw IBTrACS CSV and raw JMA best-track text Model checkpoints and per-storm prediction files (regenerated locally by training the baselines from the GitHub repo) License Derived parquets: CC-BY-4.0. Cite TyphoonUQ-Bench and the three upstream providers when redistributing.