agrimanager/wofost-weather-pool

WOFOST Weather Pool

This repository contains a reusable WOFOST-Gym weather-scenario pool for AgriManager crop-management benchmarks. Each parquet row is a scenario key, not a full rollout trajectory: the crop is encoded by the file name, and the year, latitude, and longitude columns identify the weather scenario consumed by WOFOST-Gym at runtime.

Intended Use

Use this dataset to reproduce WOFOST-Gym weather-scenario sampling for AgriManager training and evaluation. The pool is intended for crop-management reinforcement learning and policy evaluation experiments where train, validation, and held-out weather scenarios must be fixed and inspectable.

This dataset is not intended to be a standalone meteorological benchmark, a global climate product, or a source of real crop yield labels.

Provenance

• Weather source: NASA POWER meteorological data, retrieved through PCSE's NASAPowerWeatherDataProvider.
• Simulator/runtime source: WOFOST-Gym, PCSE, and WOFOST.
• Derived artifact: AgriManager generated, filtered, split, and packaged scenario rows and the accompanying PCSE weather cache archive.
• Ownership statement: this dataset does not claim ownership of NASA POWER, WOFOST, PCSE, or WOFOST-Gym.

Synthetic Data Status

croissant_rai.json sets rai:hasSyntheticData to true. The meteorological inputs are derived from NASA POWER weather data, but the released benchmark rows are generated and simulator-derived WOFOST-Gym scenario keys curated into fixed train, validation, and test splits by a deterministic AgriManager/WOFOST pipeline with fixed seeds and documented preprocessing scripts. This disclosure applies to the benchmark scenarios and splits, not to the underlying NASA POWER source data.

Crops

The pool contains 20 crop shards:

barley, chickpea, cotton, cowpea, fababean, groundnut, maize, millet, mungbean, pigeonpea, potato, rapeseed, rice, seed_onion, sorghum, soybean, sugarbeet, sunflower, sweetpotato, and wheat.

Crop-specific latitude and longitude windows are defined in integrations/wofost_gym/dataset_tools/weather_pool_configs/pool_crop_*.yaml in the AgriManager code release.

Files

Path pattern	Count	Rows per file	Total rows	Description
train/{crop}.parquet	20	3,200	64,000	Training weather scenarios.
val/{crop}.parquet	20	128	2,560	Validation weather scenarios.
test/{crop}.parquet	20	512	10,240	Held-out weather scenarios.
meteo_cache.tar.gz	1	n/a	n/a	Bundled PCSE/NASA POWER cache files required by the scenarios.
croissant_rai.json	1	n/a	n/a	Croissant JSON-LD metadata with core fields and Responsible AI metadata.

The total parquet scenario count is 76,800 rows.

Schema

All crop parquet files use the same schema:

Column	Type	Units	Description
year	int64	calendar year	Year passed into the WOFOST agromanagement/weather scenario.
latitude	float64	decimal degrees	Scenario latitude, rounded to 2 decimal places.
longitude	float64	decimal degrees	Scenario longitude, rounded to 2 decimal places.

The crop label is encoded by the parquet file name. The split label is encoded by the parent directory. The parquet files do not contain raw daily weather variables or rollout rewards.

Croissant Metadata

croissant_rai.json is the completed Croissant JSON-LD metadata file for this dataset. It contains the core Croissant metadata fields and the Responsible AI metadata fields used for NeurIPS 2026 Evaluations & Datasets review, including dataset limitations, data biases, personal or sensitive information, intended use cases, social impact, synthetic-data status, source datasets, and provenance activities.

Weather Cache

meteo_cache.tar.gz contains the PCSE NASAPowerWeatherDataProvider cache files needed by these rows. The AgriManager runtime helper agrimanager.env.wofost_gym.weather_pool.ensure_pool() downloads or locates the dataset and extracts this archive before WOFOST-Gym execution.

The cache is included so reviewers and users can:

• run the exact same weather scenarios without re-downloading from NASA POWER,
• avoid NASA POWER API failures, rate limits, or future data/version drift,
• reproduce WOFOST-Gym runs from the fixed scenario rows, and
• load the pool through AgriManager's runtime helper without manually preparing PCSE weather files.

The packaging script includes cache files for each scenario year plus a one-year padding window on both sides (year_padding=1) so WOFOST-Gym has the weather context required around each season.

Generation and Filtering

The experiment configs consume this materialized Hugging Face weather pool; they do not recompute the source-level filtering logic at training time.

The pool was generated with the AgriManager weather-pool builder:

python integrations/wofost_gym/dataset_tools/build_weather_pool.py \
  pool_crop_<crop> \
  --output-dir <weather_pool_dir> \
  --num-workers 32

python integrations/wofost_gym/dataset_tools/package_weather_pool.py \
  --pool-dir <weather_pool_dir> \
  --clean \
  --year-padding 1

Generation settings visible in the codebase:

• generation_seed: 42
• year_range: [1984, 2019]
• num_train_samples: 3200 per crop
• num_val_samples: 128 per crop
• num_test_samples: 512 per crop
• latitude and longitude are sampled uniformly from crop-specific windows and rounded to 2 decimals.

Weather validation runs a no-op WOFOST-Gym episode for each sampled scenario and keeps only scenarios where:

• weather data can be loaded completely through PCSE/NASA POWER,
• the crop reaches at least min_dvs_threshold: 1.5, and
• the crop produces non-zero storage organ biomass (max WSO > 0).

Equivalently, each candidate scenario s = (crop, year, latitude, longitude) is kept only if:

weather_complete(s) = rollout has no missing-weather/cache errors
phenology_ok(s)     = max_DVS(s) >= 1.5
yield_ok(s)         = max_WSO(s) > 0
keep(s)             = weather_complete(s) and phenology_ok(s) and yield_ok(s)

Failed scenarios are retried and replaced. Scenario keys are deduplicated by (crop, year, latitude, longitude).

Split Construction and Leakage Prevention

The generation code uses deterministic split-specific RNG streams derived from generation_seed: 42:

• train uses seed 42,
• validation uses seed 43,
• test uses seed 44.

Each later split inherits the previously used scenario keys, so the builder avoids overlap across train, validation, and test by (crop, year, latitude, longitude).

Limitations

• The pool is simulator-oriented: it contains scenario keys and cache files for WOFOST-Gym, not observed management decisions or measured crop yields.
• The geographic windows are crop-specific benchmark design choices, not claims about complete crop suitability regions.
• The WOFOST viability filter removes scenarios that fail simulator completeness or maturity/yield checks, so the pool is not an unbiased sample of all NASA POWER grid cells.
• Reproduction requires compatible versions of AgriManager, WOFOST-Gym, PCSE, and their crop configuration files.

Size and Inspection Sample

The current hosted dataset artifact includes a Xet-backed meteo_cache.tar.gz of 17,563,723,039 bytes. Because this hosted artifact is over 4 GB, a small inspection sample is provided for NeurIPS reviewers.

The reviewer inspection sample is provided under sample/ and contains:

sample/
  README_sample.md
  sample_scenarios.parquet
  sample_scenarios.csv
  sample_manifest.json
  meteo_cache_sample.tar.gz

Selection rule: first 2 cache-available rows for maize, potato, and wheat from each of train, val, and test.

Sample contents:

• 18 scenario rows.
• 12 unique trimmed PCSE/NASA POWER cache files.
• meteo_cache_sample.tar.gz size: 794,556 bytes.

sample_scenarios.parquet and sample_scenarios.csv contain the same scenario table; the parquet file matches the main dataset format and renders as a typed table in the Hugging Face viewer, while the CSV is a plain-text fallback. sample_manifest.json is separate metadata describing the sample-selection rule, archive size, and cache-file inventory. meteo_cache_sample.tar.gz is the runnable cache artifact and is not expected to be previewable in the browser.

The sample connects scenario rows to the cache explicitly: sample_scenarios.parquet includes split, crop, year, latitude, longitude, cache_file, and cache_years. Cache filenames are computed with the same convention as the full AgriManager packaging code, NASAPowerWeatherDataProvider_LAT{int(latitude*10):05d}_LON{int(longitude*10):05d}.cache, and each cache file is trimmed to year - 1, year, and year + 1.

License

This dataset is released under the CC-BY-4.0 license.

Users may share and adapt the dataset with appropriate attribution.

Citation

If you use this dataset, cite the accompanying AgriManager benchmark paper or code release, NASA POWER, and the upstream WOFOST-Gym/PCSE/WOFOST software used to execute the scenarios.

Release, Maintenance, and Private Artifacts

This dataset is publicly available through this Hugging Face dataset repository. A frozen version corresponding to the paper will be maintained for at least two years after publication. Future updates, if any, will be published as new Hugging Face revisions rather than silently replacing the documented release.

No essential component required to reproduce the benchmark dataset will remain private. The released artifacts include the scenario parquet files, dataset card, Croissant/RAI metadata, runtime weather cache, and inspection sample where required. Large training logs, intermediate checkpoints, W&B run histories, local scratch files, and machine-specific cluster outputs are not part of the dataset release artifact.