Penguin-N's picture
Final de-anonymized metadata: authors and citation
6df6443 verified
|
Raw History Blame Contribute Delete
6.16 kB
metadata
license: cc-by-4.0
pretty_name: AgriManager per-run results (t32-diffmodels)
language:
  - en
task_categories:
  - reinforcement-learning
tags:
  - agriculture
  - crop-management
  - benchmark
  - evaluation-results
  - croissant
configs:
  - config_name: index
    data_files:
      - split: test
        path: index.csv

AgriManager per-run results: schema-results-t32-diffmodels

Per-scenario evaluation records of the AgriManager benchmark (NeurIPS 2026, Evaluations and Datasets track) for T3.2 unified policy: every checkpoint listed below scored once, with greedy decoding, on the frozen held-out test suite AgriManager-Framework/agrimanager-test-suite.

Contents

8 scored runs, 72 (run, test split) results, 9,216 episodes.

Scoring-queue run Path in this repository Role
t32-diffmodels-qwen25-3b-no-think-grpo t32-diffmodels-qwen25-3b-no-think-grpo/step-800 terminal checkpoint
t32-diffmodels-qwen25-3b-think-grpo t32-diffmodels-qwen25-3b-think-grpo/step-800 terminal checkpoint
t32-diffmodels-qwen3-0p6b-think-grpo t32-diffmodels-qwen3-0p6b-think-grpo/step-800 terminal checkpoint
t32-diffmodels-qwen3-4b-no-think-drgrpo t32-diffmodels-qwen3-4b-no-think-drgrpo/step-800 terminal checkpoint
t32-diffmodels-qwen3-4b-no-think-ppo t32-diffmodels-qwen3-4b-no-think-ppo/step-800 terminal checkpoint
t32-diffmodels-qwen3-4b-tool-drgrpo t32-diffmodels-qwen3-4b-tool-drgrpo/step-800 terminal checkpoint
t32-diffmodels-qwen3-4b-tool-ppo t32-diffmodels-qwen3-4b-tool-ppo/step-800 terminal checkpoint
t32-diffmodels-qwen3-8b-think-grpo t32-diffmodels-qwen3-8b-think-grpo/step-800 terminal checkpoint

Layout, one directory per (run, revision):

<run_name>/<revision>/eval_manifest.json        what was scored: checkpoint repo and revision, training
                                                  manifest, suite files with SHA-256, protocol, per-split metrics
<run_name>/<revision>/<suite group>/<split>/
    results.json             one record per episode: env_id, env_config, and every turn (prompt,
                             observation, raw model response, executed action, reward, metrics)
    metrics.json             split-level metrics (primary metric mean and std, episode reward, invalid-action rate)
    validation_metrics.json  the same metrics under the trainer's validation-metric names
    eval_env_config.json     response-format settings asserted and applied at scoring time
index.csv                    one row per (run, split): checkpoint, seed, suite file and SHA-256,
                             primary metric mean/std, number of episodes, scoring GPU, path

Untrained-model rows (baseline-*) score Qwen/Qwen3-4B-Instruct-2507 in the think and no-think response formats; their revision is a label (base-<tier>).

How the records were produced

  • Scored on 2026-07-26 (manifest timestamps) with entrypoints/eval/e5_evaluate_checkpoint.sh from the code repository, one pass per checkpoint at its terminal revision, greedy decoding, no tuning on the test scenarios.
  • The scoring GPU of these runs is not recorded (column scoring_gpu). Greedy decoding is not bit-stable across GPU types and drivers: re-scoring a checkpoint on different hardware changes a few episodes per split.
  • suite_sha256 of every split matches the SHA-256 of the file with the same path in AgriManager-Framework/agrimanager-test-suite (column suite_sha256_in_test_suite). The manifests name the suite source AgriManager-Framework/schema-shift-confirmatory-eval-v1 at revision b4f68ec (owner name rewritten like the others): the internal working dataset of the same files, which is not released.
  • Hub owner names inside eval_manifest.json and eval_env_config.json were rewritten to the release organization (AgriManager-Framework/<repository>); the revisions and file checksums are those of the scored checkpoints. No other field was changed. Worker logs are not included.

Scope. These are the July 2026 test-suite scores of the T3.2 backbone and optimizer comparison runs (one training seed each). They were scored before the three Cycles crop-growth suites were rebuilt, so their Cycles cells use the superseded first-build suites (e4_t32_cycles_yield, e4_t32_cycles_profit). The paper treats these runs as exploratory and uses this dataset only for the Qwen3-8B diagnostic of its backbone appendix; t32-diffmodels-qwen3-0p6b-think-grpo is not reported in the paper.

Relation to the paper's numbers

The aggregate tables in the code repository (results/tables/*.csv) are the source of every number in the paper. Where a table cell was computed from an earlier scoring pass of the same checkpoint, the split means here can differ slightly from it because of the GPU nondeterminism noted above.

Intended use

Inspecting, re-aggregating and auditing the paper's test-suite results at the level of single episodes and turns. Model responses are outputs of research checkpoints in simulation; they are not agronomic advice.

License

CC BY 4.0 for the records. The model outputs were produced by checkpoints whose own licenses are stated on their model cards.

Citation

@inproceedings{agrimanager2026,
  title     = {AgriManager: A Framework and Benchmark for LLM-RL Generalization in Agricultural Management},
  author    = {Gui, Chi and Zheng, Junwen and Nigam, Ritwik and Yang, Qianlan and Lang, Meagan and Dongre, Vardhan and Barr, Jeremiah and Wang, Yu-Xiong and Adve, Vikram S.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026},
  url       = {https://agrimanager-framework.github.io/}
}