AgriTaxon: Can Large Multimodal Models Identify What They See in Agriculture?

Xin Zeng, Benfeng Xu, Shancheng Fang, Huarui Wu

🌐 Project Page · 💻 GitHub · 📎 Supplementary Material

Overview

AgriTaxon is the first benchmark for open-ended taxonomic naming in agriculture. It tests whether Large Multimodal Models (LMMs) can correctly name a species from its image—without predefined options.

The benchmark spans four agricultural domains:

Track	Species	Source
🌿 Crops	1,971	FAO Ecocrop (via Wikidata P4753)
🐄 Livestock	178	FAO DAD-IS (via Wikidata P3380)
🐛 Pests	3,485	EPPO (via Wikidata P3031)
🌾 Weeds	1,798	EPPO (via Wikidata P3031)
Total	7,432

Every label is linked to authoritative FAO and EPPO databases via Wikidata, forming a traceable authority chain. A companion in-the-wild subset (wild/, real iNaturalist field photos with a multi-image setting) is described below.

Key Features

• 7,432 species across four agricultural domains with entity-level species/breed labels
• Authority-grounded labels: every entity linked to FAO/EPPO via Wikidata QIDs
• Two evaluation protocols: multiple-choice (with semantically hard negatives) and open-ended naming
• LLM-as-a-Judge scoring with 98% expert agreement for alias validation
• AgriTaxon-Hard: 1,052-sample diagnostic subset where ≤2 of 14 frontier models answer correctly
• Reveals a striking seeing-without-naming gap: best model reaches 83% on multiple-choice but drops to 33% on open-ended naming
• AgriTaxon-Wild (wild/): in-the-wild iNaturalist field photos with a multi-image (K∈{1,3,5}) setting

Dataset Structure

├── images/                     # curated: one Wikimedia Commons image per entity
│   ├── crop/      (1,971)   livestock/ (178)   pest/ (3,485)   weed/ (1,798)
├── annotations/
│   ├── {track}.jsonl           (open-ended: qid, label, image, track, source, wiki, KB_ID)
│   └── {track}_mc.jsonl        (multiple-choice: includes distractor options)
├── splits/
│   ├── hard.json               (AgriTaxon-Hard: 1,052 QIDs + selection criteria)
│   └── hard_model_accuracy.json
├── metadata.json
└── wild/                       # AgriTaxon-Wild: in-the-wild iNaturalist subset (see below)
    ├── images/{track}/...      (14,845 field photos, 5 per entity)
    ├── annotations/{track}.jsonl + {track}_mc.jsonl
    ├── attributions.jsonl      (per-photo license + credit; required)
    ├── metadata.json  LICENSES.md

Annotation Fields

Each {track}.jsonl entry contains:

Field	Type	Description
qid	string	Wikidata QID (e.g., Q12345)
label	string	Canonical species name
image	string	Relative path to image file
track	string	One of: crop, livestock, pest, weed
source	string	Wikidata property used for sourcing
enwiki	string	English Wikipedia article title (optional)
zhwiki	string	Chinese Wikipedia article title (optional)
ecocropID / faoID / eppoCode	string	Authority database identifier (track-dependent)

Each {track}_mc.jsonl additionally includes distractor options for multiple-choice evaluation.

Image Sources

Curated images are sourced from Wikimedia Commons under Creative Commons licenses (predominantly CC BY and CC BY-SA). Each image retains its original license. Images with a shorter dimension below 224px have been filtered out; images with max dimension above 4096px have been resized.

---

🌍 AgriTaxon-Wild (in-the-wild subset)

wild/ is an in-the-wild, multi-image extension. Where the main benchmark uses one curated Wikimedia Commons image per entity, AgriTaxon-Wild uses real iNaturalist research-grade field photographs — multiple independent observations per entity — to evaluate recognition under realistic field conditions.

Track	Entities	Coverage vs. main	Images
🌿 Crop	1,566	79.5%	7,830
🐄 Livestock	13	7.1%	65
🐛 Pest	374	10.7%	1,870
🌾 Weed	1,016	56.5%	5,080
Total	2,969	39.9%	14,845

• Strict alignment: Wikidata P3151 (iNaturalist taxon ID) only; exact observation taxon; quality_grade=research, verifiable, non-captive, licensed photos.
• Multi-image setting: each entity has up to 5 field photos; use the first K ∈ {1,3,5} for the multi-view evaluation. wild/annotations/{track}_mc.jsonl uses the same 4-option, text-embedding-hard-distractor protocol as the main benchmark.
• Shared qids with the curated set → supports controlled curated-vs-wild comparison.
• Livestock collapses to 13 entities under strict P3151 (most breeds lack an iNaturalist taxon).

import json
mc = [json.loads(l) for l in open("AgriTaxon/wild/annotations/weed_mc.jsonl")]
ex = mc[0]
print(ex["label"], ex["options"], ex["answer"])
print(ex["images"][:5])     # up to 5 field-image paths under wild/

⚠️ Wild licensing & attribution

Wild images are individual iNaturalist photos, each under its own CC license — not uniform, and predominantly NonCommercial. Treat the wild subset as non-commercial, and attribution is required (see wild/attributions.jsonl and wild/LICENSES.md).

License	Photos		License	Photos
CC BY-NC	12,400		CC BY-NC-SA	200
CC BY	1,493		CC BY-NC-ND	175
CC0	387		CC BY-SA	150
			CC BY-ND	40

215 photos are NoDerivatives (CC BY-NC-ND / CC BY-ND). Photos & observation metadata © their iNaturalist contributors; see wild/attributions.jsonl for per-photo credit.

---

Usage

pip install huggingface_hub
huggingface-cli download Xin1818/AgriTaxon --repo-type dataset --local-dir AgriTaxon

import json
samples = [json.loads(l) for l in open("AgriTaxon/annotations/crop.jsonl")]
print(f"Loaded {len(samples)} crop species"); print(samples[0])
hard = json.load(open("AgriTaxon/splits/hard.json"))
print(f"AgriTaxon-Hard: {len(hard)} samples")

Evaluation Protocols

Multiple-Choice

Each question presents 4 options: 1 correct answer + 3 semantically hard negatives generated from text embeddings of taxonomically close species.

Open-Ended Naming

Models must produce the species name from scratch without any options. Evaluated with:

• Exact Match (EM): after normalization (lowercasing, punctuation removal)
• Acc (LLM-as-a-Judge): GPT-5 Mini validates whether predictions that fail EM are valid aliases (common names, taxonomic synonyms, spelling variants), achieving 98% expert agreement

Main Results

Accuracy (%) on AgriTaxon, ranked by open-ended Acc.

Model	MC Mean	OE EM	OE Acc	Hard
gemini-3-pro-preview	82.5	44.0	51.2	9.0
doubao-seed-2-0-pro	79.4	44.1	48.6	6.0
gemini-3-flash-preview	83.0	33.4	48.1	8.7
doubao-seed-2-0-lite	77.6	37.6	44.0	5.6
kimi-k2.5	73.7	30.0	38.0	2.1
gpt-5	78.6	29.6	37.6	8.4
glm-4.6v	65.1	22.7	30.1	6.2
qwen3-vl-235b-a22b	68.4	21.7	27.5	2.2
gpt-5-mini	71.6	22.0	27.4	7.7
qwen3.5-397b-a17b	71.9	21.7	26.8	9.2
qwen3-vl-30b-a3b	59.9	16.0	22.5	2.2
glm-4.6v-flashx	60.3	15.2	19.6	5.8
qwen3.5-35b-a3b	67.8	11.3	17.4	4.2
claude-haiku-4-5	60.4	11.2	14.7	4.7

Licensing

• Benchmark metadata & annotations: CC BY 4.0
• Curated images (images/): Wikimedia Commons, predominantly CC BY and CC BY-SA; each retains its original license
• Wild images (wild/images/): iNaturalist, predominantly CC BY-NC (non-commercial); per-photo license + attribution in wild/attributions.jsonl (see wild/LICENSES.md)
• Source code & prompts: MIT License
• Authority identifiers: Wikidata QIDs under CC0; FAO and EPPO identifiers used for reference linking only

Because the wild subset adds NonCommercial images, the repository license is marked other (mixed CC BY / CC BY-NC). Commercial use is limited by the NC images in wild/.