AgriManager / Docs
Website GitHub

Models and data

The released datasets, test suite and checkpoints, and where each one lives.

All released artifacts live on Hugging Face under the AgriManager-Framework organization. The code is at AgriManager-Framework/AgriManager.

By setting#

Each benchmark setting with its Hugging Face collection, W&B training runs and test-suite folder.

TierSettingModelsTraining curvesTest set
Tier 1 · Within SchemaT1.1 Weather regimeModelsTraining curvesTest set
T1.2 Crop varietyModelsTraining curvesTest set
T1.3 Price regimeModelsTraining curvesTest set
Tier 2 · Single-Axis SchemaT2.1 Observation schemaModelsTraining curvesTest set
T2.2 Action menuModelsTraining curvesTest set
T2.3 ObjectiveModelsTraining curvesTest set
Tier 3 · Compositional SchemaT3.1 Cross-simulator transferModelsTraining curvesTest set
T3.2 Unified policyModelsTraining curvesTest set
ExploratoryOther backbones and RL algorithmsModelsTraining curves—

Data#

ArtifactRepositoryContent
20-crop WOFOST weather poolAgriManager-Framework/wofost-weather-pool(crop, year, latitude, longitude) scenarios for 20 crops with train / val / test splits, plus the weather cache; CC BY 4.0
WOFOST weather-regime poolAgriManager-Framework/wofost-weather-regime-poolchickpea and potato scenarios with weather-regime groups; CC BY 4.0
WOFOST maize poolAgriManager-Framework/weather_pool_maizemaize scenarios with train / val splits; CC BY 4.0
Held-out test suiteAgriManager-Framework/agrimanager-test-suiteevery frozen test split with manifests, checksums and audits (Held-out test suite); CC BY 4.0
Per-run result bundlesAgriManager-Framework/schema-results-*per-scenario outputs and eval_manifest.json for every scored checkpoint; CC BY 4.0
Aggregate tablesthe code repository, results/tables/the CSV tables behind every paper number (Reproducing the paper); CC BY 4.0

Cycles and DSSAT scenarios are generated from configs and seeds; their only static datasets are the test-suite splits.

Language-model checkpoints#

Each checkpoint is a full Transformers model fine-tuned from Qwen/Qwen3-4B-Instruct-2507 with GRPO: one repository per training run, one branch per saved step (step-N). The terminal checkpoint, the one the paper scores, is the highest step-N. Each repository carries a run_manifest.json with the base model, seed, response format, setting, step and file checksums, which the evaluator reads. results/scoring_manifest.tsv lists every revision scored for the paper.

SettingRepositories (AgriManager-Framework/…)
weather-regimeschema-t11-{chickpea,potato}-{think,no-think}-r{2,3,4}
crop-varietyschema-t12-crosscrop-{4id,8id,16id}-…
price-regimeschema-t13-price-… (r1–r3)
observation-schemaschema-t21-observation-…
action-menuschema-t22-family-{a,b}-…
reward-formulationschema-t23-{objectives,yield-only}-…
cross-simulatorschema-t31-{wofost,dssat,cycles}-…
unified policyschema-t32-qwen-{think,no-think}-r{2,3,4}

Exploratory checkpoints, one seed each: Llama-3.2-3B-AgriManager-t32-unified-{think,no-think} (the unified policy on Llama-3.2-3B-Instruct) and t32-diffmodels-* (the unified policy with other backbones and algorithms).

The terminal agent.zip of every NN baseline run is kept with its config and will be uploaded under AgriManager-Framework/.

Licenses#

ComponentLicense
AgriManager codeApache-2.0
Datasets we publish (weather pools, test suite, result bundles, tables)CC BY 4.0
Checkpoints fine-tuned from Qwen3 modelsApache-2.0
Llama-3.2-3B-AgriManager-t32-unified-* (base Llama-3.2-3B-Instruct)Llama 3.2 Community License
t32-diffmodels-qwen25-3b-* (base Qwen2.5-3B-Instruct)Qwen Research License

Simulators and wrappers are not vendored and keep their own licenses (THIRD_PARTY_NOTICES.md). The native Cycles binary that CycleGym downloads at install time is distributed under CC BY-NC-ND 4.0.