AgMMU
Harbor: pendingAgMMU is a challenging real‑world benchmark for evaluating and advancing vision-language models (VLMs) in the knowledge‑intensive domain of agriculture.
- Samples
- N/A
- Released
- 2025
- Scores
- 0
AgroGround
AgroGround: Multi-Granularity Grounded Recognition in Agriculture
A benchmark for grounded agricultural recognition measuring joint identity and localization correctness plus abstention on 1,480 human-verified images.
How to run
ltb · agroground
$ npm i -g lettokenburn-cli$ ltb bench pull agroground --n 50$ ltb bench run ./agroground -a claude-code -m anthropic/claude-opus-4-8$ ltb bench push
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
5 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| AgroGround-8B | 78.2 | +existence 12% | Independent | medium | Oct 3, 2026 | arxiv.org |
| AgroGround-2B | 74.7 | +existence 12% | Independent | medium | Oct 3, 2026 | arxiv.org |
| Qwen3-VL-8B-Instruct | 64.4 | zero-shot | Independent | medium | Oct 3, 2026 | arxiv.org |
| Qwen3-VL-2B | 51.8 | zero-shot | Independent | medium | Oct 3, 2026 | arxiv.org |
| GroundingDINO | 27.6 | best candidate | Independent | medium | Oct 3, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
AgroGround converts labels from eight agricultural VQA datasets into spatial annotations, creating a training set of 794,850 examples. The evaluation benchmark consists of 1,480 human-verified images disjoint from training data.
It scores predicted identities, regions, joint correctness, and healthy-image abstention. Combining recognition and localization instructions raises joint accuracy to 43.3%, while healthy negatives improve abstention to 95.0%.