UQE2E: Fine-tuned weights and evaluation data

Companion data for the manuscript

Uncertainty-Aware End-to-End AI Weather Forecasting: Disentangling Observation and Model Contributions Rodrigo Almeida, Noelia Otero, Jost Arndt, Simon Baur, Wojciech Samek, Jackie Ma Applied Machine Learning Group, Fraunhofer Heinrich-Hertz Institute, Berlin

Preprint: arxiv.org/abs/2608.30795 Code: https://gitlab.hhi.fraunhofer.de/ai-aml/uqe2e.

This repository holds:

1. the four fine-tuned model runs (selected checkpoint per run and per lead),
2. the normalisation factors as used by the runs,
3. a few small static auxiliary arrays the data loaders expect,
4. the evaluation caches behind every figure and number of the manuscript.

Everything else (raw observations, upstream deterministic weights, ERA5, climatology, WeatherBench 2 scores) is either in the upstream Aardvark release or regenerable with the code; see the code repository's DATA.md.

Layout

The tree mirrors the code repository, so a single download into the repository root puts every file where the code looks for it:

uv run hf download rodrigoalmeida1994/uqe2e --repo-type dataset --local-dir .

data/
  outputs/
    encoder_noise_cond_embed_ft_full_fixeddata00z/encoder_deviousiguana/            # prob. encoder, embedding noise (epoch 0)
    encoder_noise_cond_embed_ft_full_fixeddata00z_normmode/encoder_glaringsawfly/   # prob. encoder, norm noise (epoch 10)
    processor_mc_dropout_crps_fixeddata00z_7member/processor_malachitegoshawk/      # MC-dropout processor paired with deviousiguana
    processor_mc_dropout_crps_fixeddata00z_normmode_7member/processor_powerfulzebra/ # MC-dropout processor paired with glaringsawfly
  norm_factors/                    # mean_*/std_* per stream, hirs_means/hirs_stds
  era5/                            # elev_vars_1.npy, era5_orog_0p25deg_m.npz, era5_pressure_levels_4u*.npy
  lat_weights/weights_lat_1.npy
  era5_spatial_means.npy  era5_avdiff_means.npy  era5_avdiff_stds.npy  loss_weights.npy
  hadisd_station_info_v343_2025f.txt
results/                           # evaluation caches read by uq_e2e.paper_figures

Runs

Each run directory holds config.pkl, the selected checkpoint epoch_<k> and the per-epoch validation logs (losses_*.npy, rmse_*.npy, train_losses_*.npy, val_metrics_history_0.json) from which the checkpoint was selected (uq_e2e.checkpoint_selection.select_best_epoch). Processor runs have one forecast_<lead>/ sub-directory per daily rollout step (leads 1..10), each with its own selected checkpoint. Checkpoints of non-selected epochs and the push-forward intermediates (ic_*.mmap) are not included.

Run	Role	Selected epoch(s)
encoder_deviousiguana	probabilistic encoder, embedding noise mode, 7-member fair-CRPS fine-tune of the upstream encoder	0
encoder_glaringsawfly	probabilistic encoder, norm (conditional LayerNorm) noise mode; headline encoder	10
processor_malachitegoshawk	processor with trained-in MC dropout (p=0.05, fair CRPS), push-forward on deviousiguana's analyses	leads 1..10: 8, 1, 0, 0, 0, 1, 0, 0, 0, 0
processor_powerfulzebra	same, on glaringsawfly's analyses; headline pair	leads 1..10: 9, 1, 0, 0, 0, 1, 1, 0, 0, 0

The _normmode parent directories contain "embed" in their name for historical reasons; the runs inside are noise_mode="norm". The station decoders and the deterministic encoder/processor are the upstream release (https://huggingface.co/datasets/av555/aardvark-weather).

Evaluation caches (results/)

File	Content
e2e_gridded_leads_{powerfulzebra,malachitegoshawk}_entropy7x7.npz	7x7 nested ensemble vs ERA5, 2018, leads 1..10: skill, spread, ANOVA variance split, entropy/MI decomposition, per-init series
e2e_gridded_leads_upstream.npz	deterministic upstream baseline, same protocol
e2e_gridded_lead0_{glaringsawfly,deviousiguana,upstream}.npz	encoder analysis (lead 0) skill vs ERA5
e2e_nested_leads_{powerfulzebra,malachitegoshawk,upstream}_{tas,ws}.json	HadISD station verification, global and per region, per-init series
e2e_station_leads_powerfulzebra_7x7_{tas,ws}.npz	per-station metrics for the maps
ose_smoke_full_{norm,embed}_7x7.json	observation-denial (OSE) experiment, 9 streams
crossed_decomposition_powerfulzebra_crossed_fullyear.npz	crossed (frozen dropout-mask) decomposition

Provenance and licenses

The fine-tuned weights start from the Aardvark Weather release (Allen et al., Nature 2025) and were trained on the Aardvark observational dataset (CC BY-NC-ND 4.0) with ERA5 targets (Copernicus/ECMWF, via WeatherBench 2). Please cite those sources alongside this dataset and the manuscript (citation metadata in the code repository's CITATION.cff).