t32-diffmodels-qwen3-4b-tool-ppo

A crop-management policy from the AgriManager benchmark (NeurIPS 2026, Evaluations and Datasets track): Qwen/Qwen3-4B-Instruct-2507 fine-tuned with PPO on the T3.2 unified prompt-conditioned policy, backbone/optimizer comparison setting, think response format, training seed 42.

• Paper: TODO(release-url)
• Code: https://github.com/AgriManager-Framework/AgriManager
• Held-out test suite: AgriManager-Framework/agrimanager-test-suite

Every checkpoint lives on its own branch (step-N); main holds no weights. Load a checkpoint with an explicit revision:

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("AgriManager-Framework/t32-diffmodels-qwen3-4b-tool-ppo", revision="step-800")
tokenizer = AutoTokenizer.from_pretrained("AgriManager-Framework/t32-diffmodels-qwen3-4b-tool-ppo", revision="step-800")

What this checkpoint is

Field	Value
Repository	AgriManager-Framework/t32-diffmodels-qwen3-4b-tool-ppo
Base model	Qwen/Qwen3-4B-Instruct-2507
Algorithm	PPO (VeRL 0.7.0)
Setting	T3.2 unified prompt-conditioned policy, backbone/optimizer comparison; Tier 3
Simulator(s)	WOFOST, DSSAT, Cycles
Response format	think: reasoning inside <tool_call>…</tool_call>, then the action inside <answer>…</answer>
Training seed	42 (data seed 42, data-loader seed 42)
Revisions	step-800
Terminal revision	step-800
Early stop	none (trained for the planned length)
Selection rule	terminal or last-valid checkpoint, never best-validation
Run script	not recorded in the run manifest
Paper	the backbone-comparison appendix
Evidence status	exploratory: single training seed

Each revision carries a run_manifest.json with the base model, seeds, training schedule, response format and a SHA-256 for every file.

Results

Not part of the three-seed test-suite results. The paper reports this run only in the backbone-comparison appendix (exploratory).

How to use it

The model expects the AgriManager prompt of its simulator and replies with one management action in the format above. Use it through the AgriManager environment loop; it is not a general-purpose agronomy assistant. To score it on the held-out test suite with the code repository:

bash entrypoints/eval/e5_evaluate_checkpoint.sh \
    --checkpoint-repo AgriManager-Framework/t32-diffmodels-qwen3-4b-tool-ppo --revision step-800 \
    --suite-repo AgriManager-Framework/agrimanager-test-suite --tier t32 --no-upload \
    --suite-group e4_t32_wofost_seen --suite-group e4_t32_wofost_heldout \
    --suite-group e4_t32_dssat_yield --suite-group e4_t32_dssat_profit \
    --suite-group e4_t32_cycles_yield_v2 --suite-group e4_t32_cycles_profit_v2

The evaluator reads the response format from run_manifest.json and refuses to score the checkpoint in a different format.

Limitations

• Trained and evaluated in simulation only (WOFOST, DSSAT, Cycles). Simulator scores are not field outcomes; do not apply its recommendations to a real farm.
• The scripted agronomic reference used in the paper is a conservative codified schedule, not calibrated against experts or optimal control.
• Scores use greedy decoding; re-scoring on other GPU types changes a few episodes per split.

License

The weights are released under the Apache License 2.0. The base model Qwen/Qwen3-4B-Instruct-2507 is released by its authors under Apache 2.0.

Citation

@inproceedings{agrimanager2026,
  title     = {AgriManager},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026},
  note      = {TODO(release-url): final title, authors and URL}
}