Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('parquet', {}), NamedSplit('test'): (None, {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Protostar Climate Challenge Submission — OlmoEarth Provenance Audit + Cutoff-Safe Forecast

Negative result: no (forecast core is positive; the EO branch result is reported in the Method section, whichever way its gate resolves)

Forecasts of yearly weather for 210 U.S. cities, 2019–2025, made as of the end of 2018 (no data that reveals 2019–2025 weather), for the Protostar Challenge: Climate Risk Intelligence.

Headline finding on reuse: the released OlmoEarth foundation models (v1, v1.1, v1.2, v1.3), the OlmoEarth Platform, and the precomputed OlmoEarth v1.3 embeddings cannot legally be used as forecast inputs in this challenge — every generation was pretrained on satellite windows sampled uniformly from January 2016 to December 2024, plus 2021 land-cover/crop map targets, i.e. they learned from the 2019–2025 test period. This submission therefore (a) builds its forecast on a cutoff-safe tabular climate model, (b) attempts reuse of the OlmoEarth architecture retrained from random initialisation exclusively on ≤2018-12-31 satellite data ("OlmoEarth-2018", gated by backtest), and (c) uses released OlmoEarth artifacts only for analysis/policy, firewalled from the forecast code. Full audit: OLMOEARTH_PLAN.md.

Team

Name Hugging Face username University, Program
Joseph Pollack @Tonic Team Tonic

Method

  • Model (final, v2): per-metric models fitted on valid years of the challenge dataset (≤10 missing days for the values each metric uses). annual_avg_temp: median of three pooled-trend models — 15-year level + national 40-year trend, 15-year level + regional (32-nearest-neighbour) 40-year trend, and the national variant on station-step-adjusted series (39 neighbour-difference step adjustments). cdd: 50/50 blend of the delta method (each of the last 30 years' daily temperatures warmed by the forecast trend, CDD recomputed from a −3…+6 °F shift table that reproduces the official CDD column exactly at shift 0, median over years) and a linear 15-year-level + national-trend CDD model. total_precip_in, hottest_day_temp: 30-year mean — every contested alternative (wet-day decomposition: skill +0.000; GEV median: −0.070; second-hottest-day: −0.81) was tested and rejected.
  • Features: each city's own history up to 2018 (annual indices + daily Tmax/Tmin for CDD), station segments from coverage_and_stations, lat/lon for neighbour pooling. No external forecast data. The EO branch (below) did not pass its gate and contributes nothing to the forecasts.
  • How we checked it: rolling-origin backtests on pre-2019 data only — 22 origins (1990–2011), forecast +1…+7 years, official scoring replicated locally (reproduces the organisers' practice numbers, CDD MAE 212.5 vs 212 and 175.8 vs 175). Candidates were pre-declared (code/candidates.yaml), selected on tune-era origins (1990–2000), and the assembled v2 package was confirmed once on check-era origins (2001–2011): package Score (mean skill ×100) tune 2.39 → 4.39, check 4.96 → 6.55 vs the v1 package (v2−v1 skill: temperature +0.036 ± 0.037 tune / +0.038 ± 0.013 check; CDD +0.045 ± 0.021 / +0.024 ± 0.022, year-block bootstrap). Full detail: outputs/v2/, decision in outputs/v2/v2_decision.json. The superseded v1 model (shared/shrunk slopes on a 30-year level) is retained in code/models.py/code/forecast.py, its forecast in outputs/submission_v1.csv.
  • V3 sweep (final iteration, negative result): a pre-registered third round (V3_PLAN.md, grid frozen in code/candidates_v3.yaml, 21 of a ≤40-config budget) tested damped-trend/state-space dynamics, AR-anomaly persistence, median/quantile pooling, a hottest-day delta method, bootstrap-median and asymmetric-warming (Tmin>Tmax) CDD deltas, an Evanguo-style recent-level×sensitivity CDD model, seasonal precipitation, and confounder adjustments (volcanic-year exclusion, city sensitivity to the national anomaly, an ONI year-1 adjustment, relaxed ≤30-missing-day validity, exponentially-weighted slopes) on a fresh selection era (origins 1990–2004, gate: beat v2 by >1 year-block bootstrap SE). All 21 failed the gate (closest: AR-anomaly +0.005 vs SE 0.007); several failed informatively — asymmetric warming double-counts (CDD bias +13…+21 °F-days), warming historical annual maxima by the mean trend overshoots (+0.33 °F), and relaxed validity imports the missing-summer bias (−57 °F-days). The 2005–2011 confirmation era was deliberately not spent on a package identical to v2, and submission.csv remains the v2 forecast byte-for-byte. Full detail: outputs/v3/, decision in outputs/v3/v3_decision.json, tests in code/test_v3.py. After organiser approval, a second frozen amendment grid (code/candidates_v3_amendment.yaml) also tested CMIP6 DCPP decadal hindcasts (MPI-ESM1-2-HR/MIROC6, city-extracted in code/dcpp_fetch.py, origin-initialised hindcasts only, drift climatology from starts ≤ origin), an urbanisation slope modifier, EO-embedding similarity pooling, and a walk-forward ML stacker — all 8 failed the same gate (DCPP temperature pattern −0.004…−0.028, precipitation −0.040; decision in outputs/v3/amendment_decision.json, tests in code/test_amend.py), so the submission remains the v2 forecast.
  • Plausibility audit: code/audit_forecasts.py checks every forecast against the city's own ≤2018 history (outputs/audit_forecast_plausibility.csv). Findings: only one city-metric (Huntsville temperature, +0.13 °F) exceeds its historical record by 2025; no hottest-day forecast exceeds its city's record; the largest CDD exceedance is Eureka, CA (47 vs record 15 °F-days — a near-zero-CDD coastal city starting to accumulate cooling demand, trivial in absolute terms). CDD values are cumulative °F-days, not day counts (Beckley, WV forecast 696–732 vs its observed 893 in 2018; Phoenix's normal is ~4,300). No forecast required correction.
  • OlmoEarth branch (attempted reuse): code/eo_fetch.py downloads Sentinel-2 L2A quarterly least-cloudy windows (64×64 px @10 m on each station, scene IDs logged) for 2017–2018 only, hard-capped at 2018-12-31. code/eo_model.py trains an OlmoEarth-architecture Nano-scale encoder (patch tokenisation over space/time/band-sets, ViT encoder, pooled embedding) from random initialisation (seed logged) on those scenes; released weights, Ai2 normalisation statistics (computed over 2016–2024), and all post-2018 map layers are excluded and actively tested against (code/test_leakage.py). The resulting city embedding must pass code/eo_gate.py (≥3% leave-cities-out improvement over lat/lon for temperature or CDD trend slopes) before it may modify any forecast. Gate outcome: FAIL — the EO branch was dropped and submission.csv is the tabular core alone. The embedding is genuinely informative about where a city sits climatically: in 5-fold leave-cities-out CV it beats a lat/lon-polynomial baseline for 1989–2018 normals by +28.2% MAE (precipitation), +8.8% (CDD) and +4.1% (average temperature). But it adds nothing to trend slopes — the quantity a 2019–2025 forecast actually needs beyond a city's own history (−1.0% for temperature slope, +1.3% for CDD slope, vs the +3% gate). Because the gate was pre-registered, we did not move it after seeing the result. Details: outputs/eo/gate_decision.json. This is the submission's documented negative result for Earth-observation reuse: a 2017–2018 satellite snapshot can describe a city's climate normal, but two years of imagery cannot measure its multi-decadal warming trend — and no legally usable pretrained EO model exists for this challenge, because all released OlmoEarth generations were trained through 2024.

AI Tools and Agents

  • Tools and models: Cursor (Muse Wonder) AI agent; no AI model produced any forecast value.
  • What they did: researched the OlmoEarth provenance question, wrote and ran all code (data loading, scorer replication, backtests, EO fetch/training, policy analysis), drafted this README and OLMOEARTH_PLAN.md.
  • What a person reviewed: Joseph Pollack reviewed every download and external-data source (challenge training data, Sentinel-2 scenes, NOAA ONI, CMIP6/DCPP hindcasts), the data-rule and OlmoEarth provenance interpretation, the leakage tests, the model and pipeline code, all pre-registered gate decisions (v2 adoption, EO gate, v3 and amendment sweeps), and the final submission numbers before upload.
  • Agent instructions file: code/AGENTS.md (data rule + OlmoEarth-specific prohibitions).

Policy Idea

  1. What risk did we notice? Cooling demand grows most in absolute terms along the Gulf Coast and in the desert Southwest (Galveston, Houston, Baton Rouge, Tampa, Phoenix: +210…+240 °F-days vs their 1989–2018 normals in our v2 forecasts; outputs/policy_city_risk.csv), and fastest relative to local norms in the historically cool Northwest and Great Lakes (Seattle +41%, Sault Ste Marie +42%, Duluth +37%), where housing stock is least air-conditioned. Reno, NV combines both risks: +212 °F-days (+23%) with a Sentinel-2 green-vegetation fraction of 0.03 around the station — among the lowest natural shade in the country, shared with Elko, Phoenix and Spokane (≤0.04).
  2. What would we do about it, and who would do it? A cool-corridor programme for Reno, replicable to the high-growth/low-canopy group: the City of Reno and its electric utility jointly fund street-tree planting along transit and pedestrian corridors, cool-roof standards for public and low-income housing, and shaded bus stops, prioritised by forecast CDD growth × current green-cover deficit (the citywide extension can be mapped at 10 m from OlmoEarth v1.3 embeddings — analysis use only, never a forecast input). The utility co-pays because every avoided CDD is avoided peak generation capacity it would otherwise have to build. Gulf-Coast cities with the largest absolute growth are the parallel priority for grid capacity planning, using the same forecast table.
  3. Why is it sustainable? Shade and reflective surfaces reduce cooling demand itself rather than meeting it with more energy: avoided air-conditioning load cuts peak emissions, household bills and heat exposure simultaneously, trees appreciate as they mature, and the programme targets the demand growth our forecasts show is coming rather than reacting after peak capacity is built.

Feedback to the Ecosystem

  • To Ai2 (OlmoEarth): publish a machine-readable training-data interval per checkpoint (per modality and per map layer) on the Hugging Face model card; consider releasing cutoff-filtered checkpoints (e.g. trained only on windows ending ≤2018) so temporal-holdout challenges can reuse the model legally; expose per-sample date ranges in the pretraining dataset metadata so users can filter by cutoff without downloading ~10 TB.
  • To the challenge organisers: a public ruling/registry for common pretrained models (TimesFM, Chronos, OlmoEarth, GraphCast…) with their training cutoffs would prevent good-faith violations (a peer team documented nearly submitting TimesFM 2.5, trained through 2024); publishing the scorer and a backtest helper, and adding per-metric valid flags to annual_indices, would remove re-implementation risk (we verified our scorer only by matching the 212/175 practice numbers).
  • To Hugging Face: a standard training_data_end_date model-card field that challenge Spaces can check automatically.

Files

File Content
submission.csv Our forecasts: city_id,year,metric,predicted_value (5,880 rows)
code/ All pipeline code: dataset.py, scorer.py, models.py, backtest.py, forecast.py, check_submission.py, EO branch (eo_fetch.py, eo_model.py, eo_gate.py), leakage tests, AGENTS.md
data/ Curated data described in data/README.md; large EO rasters are regenerated by code/eo_fetch.py (scene manifest in data/eo/manifest.csv)
analysis_policy/ Firewalled analysis/policy code — never imported by code/ forecast paths
outputs/ Backtest detail, frozen method decision, EO gate decision, policy ranking
OLMOEARTH_PLAN.md The full OlmoEarth provenance audit and implementation plan

Metric codes: annual_avg_temp (°F), total_precip_in (inches), hottest_day_temp (°F), cdd (°F-days). City IDs: the station ID from the city_info table of the training dataset (e.g., Pittsburgh = USW00094823).

How to Reproduce

  1. python -m venv .venv && .venv/Scripts/python -m pip install -r code/requirements.txt (plus torch pystac-client planetary-computer rasterio for the EO branch)
  2. .venv/Scripts/python code/cdd_delta.py --build → builds data/cdd_shift_table.npz from the daily subset (validates exactly against the official CDD column at shift 0)
  3. .venv/Scripts/python code/forecast_v2.py → writes the final submission.csv (v1 remains reproducible via code/forecast.py after code/backtest.py; the v3 sweep that left v2 unchanged is reproducible via code/run_v3_eval.py all — selection era only — plus code/test_v3.py; the approved-items amendment via code/dcpp_fetch.py → code/run_v3_amend.py all plus code/test_amend.py)
  4. .venv/Scripts/python code/check_submission.py submission.csv → must print OK (5,880 forecasts); rerunning step 3 is byte-identical (SHA-256 E84391C0…608)
  5. EO branch (optional, gated): .venv/Scripts/python code/eo_fetch.py → .venv/Scripts/python code/eo_model.py → .venv/Scripts/python code/eo_gate.py
  6. .venv/Scripts/python code/test_leakage.py → all leakage tests pass

Data

  • Challenge data (forecast input): CMU-OSPO/protostar-climate-train-to-2018, pinned revision bbbdebdab2a18e00f97ef87bd3f35a71e222180b, subsets annual_indices, city_info, coverage_and_stations, 1851–2018, CC0-1.0.
  • Sentinel-2 L2A via Microsoft Planetary Computer (forecast input, EO branch): scenes dated 2017-01-01–2018-12-31 only, scene IDs in data/eo/manifest.csv, ESA Copernicus terms.
  • Released OlmoEarth v1.3 embeddings (analysis only, not used by any forecast code): documented for the policy extension, https://data.source.coop/olmoearth/olmoearth-v1-3-embeddings/README.md
  • No other external data. No 2019–2025 weather-responsive data was used as a forecast input in any form.

Training data citations:

Downloads last month
235

Space using Tonic/protostar-climate-2026 1