AgriTaxon: Knowledge-grounded benchmarking of open-ended agricultural entity naming with large multimodal models

Xin Zeng, Benfeng Xu, Qian Chen, Jialin Kuai, Wentao Zhang, Liguo Lang, Shancheng Fang, Huarui Wu

🌐 Project Page · 💻 GitHub · 📎 Supplementary Material

Overview

Open-ended taxonomic naming is a foundational capability for intelligent agricultural decision-making. It requires a model to generate a standardized name for an organism or breed in an image without relying on a predefined label set, yet most agricultural vision benchmarks assume such a set.

AgriTaxon is a knowledge-grounded and AI-ready benchmark suite for evaluating this capability in large multimodal models (LMMs). Each core sample pairs an image with machine-readable identifiers, a canonical name, collected aliases, a domain label, and an executable evaluation protocol. Wikidata links the samples to FAO Ecocrop, FAO DAD-IS, or the EPPO Global Database, providing a traceable path from each image to an authoritative record.

Dataset	Purpose	Entities	Images
AgriTaxon Core	Broad agricultural entity naming	7,432	7,432
AgriTaxon-Hard	Performance-conditioned diagnostic stress test	1,052	1,052
AgriTaxon-Wild	In-the-wild multi-image evaluation	2,969	14,845

The core benchmark spans four agricultural domains:

Track	Entities	Authority source
Crops	1,971	FAO Ecocrop (via Wikidata P4753)
Livestock	178	FAO DAD-IS (via Wikidata P3380)
Pests	3,485	EPPO Global Database (via Wikidata P3031)
Weeds	1,798	EPPO Global Database (via Wikidata P3031)
Total	7,432

Here, taxonomic naming covers the standardized entity name required by the corresponding authority source. For Crop, Pest, and Weed, this is normally an organism name; for Livestock, the required entity can be a domesticated species or a breed recorded by FAO DAD-IS.

Key Features

• Knowledge-grounded and AI-ready: machine-readable Wikidata, FAO, and EPPO identifiers; canonical names and aliases; domain labels; and executable evaluation files.
• Two evaluation settings: four-option recognition with text-semantically similar distractors and open-ended naming without candidate options.
• Alias-aware evaluation: strict exact match (OE-EM) is complemented by GPT-5 Mini alias validation (OE-Acc), with 98% agreement on 100 EM-error cases annotated by a domain expert.
• Performance-conditioned diagnosis: AgriTaxon-Hard isolates 1,052 samples answered correctly by at most two of the 14 evaluated models under the default multiple-choice protocol.
• Multi-image field evaluation: AgriTaxon-Wild provides five iNaturalist field photographs per aligned entity and supports K ∈ {1, 3, 5} views.

Dataset Structure

├── images/                         # Core: one Wikimedia Commons image per entity
│   ├── crop/                       # 1,971 images
│   ├── livestock/                  # 178 images
│   ├── pest/                       # 3,485 images
│   └── weed/                       # 1,798 images
├── annotations/
│   ├── {track}.jsonl               # Open-ended annotations
│   └── {track}_mc.jsonl            # Multiple-choice annotations and distractors
├── splits/
│   ├── hard.json                   # 1,052 AgriTaxon-Hard QIDs and criteria
│   └── hard_model_accuracy.json
├── metadata.json
└── wild/
    ├── images/{track}/...          # 14,845 field photographs
    ├── annotations/{track}.jsonl
    ├── annotations/{track}_mc.jsonl
    ├── attributions.jsonl          # Required per-photo license and credit
    ├── metadata.json
    └── LICENSES.md

Annotation Fields

Each core {track}.jsonl entry contains:

Field	Type	Description
qid	string	Wikidata QID
label	string	Canonical entity name
image	string	Relative image path
track	string	crop, livestock, pest, or weed
source	string	Wikidata property used for authority linking
enwiki	string	English Wikipedia URL, when available
zhwiki	string	Chinese Wikipedia URL, when available
ecocropID / faoID / eppoCode	string	Track-dependent authority identifier

Each {track}_mc.jsonl file additionally includes the four candidate options and the correct answer. Wild annotation files contain multiple image paths for each aligned entity.

Image Sources

Core images are sourced from Wikimedia Commons under their original licenses. Images with a shorter edge below 224 px were removed, and images with a longer edge above 4,096 px were resized with Lanczos resampling. A livestock-specific content check removed six invalid image–entity pairs, leaving 178 Livestock entities for both evaluation settings.

AgriTaxon-Hard

AgriTaxon-Hard contains core samples answered correctly by at most two of the 14 evaluated models under the default multiple-choice protocol. Because membership is defined using model performance, it is a performance-conditioned diagnostic stress test. It is intended for failure analysis and tool-assisted recovery, not as an estimate of average benchmark performance or as an independently sampled difficulty distribution.

AgriTaxon-Wild

AgriTaxon-Wild aligns core entities to iNaturalist taxa strictly through Wikidata P3151, without fuzzy name matching. It retains five research-grade, verifiable, non-captive field photographs per entity.

Track	Entities	Core coverage	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

• Candidate names and evaluation procedures remain fixed between the curated and Wild settings.
• The first K ∈ {1, 3, 5} images are used for multi-image evaluation.
• Livestock results should be interpreted cautiously because strict P3151 alignment leaves only 13 entities; most breeds do not have an iNaturalist taxon.

import json

with open("AgriTaxon/wild/annotations/weed_mc.jsonl", "r") as f:
    sample = json.loads(next(f))

print(sample["label"], sample["options"], sample["answer"])
print(sample["images"][:5])

Usage

Download the complete repository rather than relying on the Dataset Viewer preview:

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

import json

with open("AgriTaxon/annotations/crop.jsonl", "r") as f:
    crop_samples = [json.loads(line) for line in f]

with open("AgriTaxon/annotations/crop_mc.jsonl", "r") as f:
    crop_mc_samples = [json.loads(line) for line in f]

with open("AgriTaxon/splits/hard.json", "r") as f:
    hard = json.load(f)

print(f"Crop: {len(crop_samples)} entities")
print(f"AgriTaxon-Hard: {len(hard)} samples")

Evaluation Protocols

Multiple-Choice Recognition

Each sample presents one correct entity name and three text-semantically similar distractors. This setting measures recognition when candidate names are available.

Open-Ended Naming

The model must produce the entity name without candidate options. Outputs are evaluated using:

• OE-EM: strict exact match after normalization.
• OE-Acc: alias-aware accuracy. GPT-5 Mini validates whether an EM-error prediction is a valid alias, common name, taxonomic synonym, or spelling variant.

For AgriTaxon Core, MC accuracy, OE-EM, and OE-Acc summary scores are unweighted macro-averages across the four tracks. Each track contributes equally regardless of its number of samples. Track-specific open-ended results report EM.

Main Results on AgriTaxon Core

Accuracy (%). Models are grouped by availability and ranked by OE-Acc within each group.

Model	MC Macro	OE-EM	OE-Acc	Hard
Proprietary models
gemini-3-pro-preview	82.5	44.0	51.2	9.0
doubao-seed-2-0-pro	79.4	44.2	48.8	6.0
gemini-3-flash-preview	83.0	33.4	48.1	8.7
doubao-seed-2-0-lite	77.6	37.7	44.2	5.6
gpt-5	78.6	29.6	37.6	8.4
gpt-5-mini	71.6	22.0	27.4	7.7
claude-haiku-4-5	60.4	11.2	14.7	4.7
Open-source models
kimi-k2.5	73.7	30.0	38.1	2.1
glm-4.6v	65.1	22.9	30.5	6.2
qwen3-vl-235b-a22b	68.4	21.7	27.6	2.2
qwen3.5-397b-a17b	71.9	21.9	27.1	9.2
qwen3-vl-30b-a3b	59.9	16.0	22.5	2.2
glm-4.6v-flashx	60.3	15.3	19.9	5.8
qwen3.5-35b-a3b	67.8	11.4	17.6	4.2

The Hard column reports accuracy on the performance-conditioned AgriTaxon-Hard subset and should not be interpreted as average benchmark performance.

Main Results on AgriTaxon-Wild

Macro-averaged accuracy (%) across the four tracks. K is the number of field images provided for each entity.

Model	Setting	MC	OE-EM	OE-Acc
Doubao-Seed-2.0-Lite	Curated	78.0	45.0	50.2
	Wild K=1	75.6	36.2	39.9
	Wild K=3	82.1	51.0	54.3
	Wild K=5	84.3	55.7	59.4
Qwen3.5-35B-A3B	Curated	66.2	12.7	19.4
	Wild K=1	63.2	13.6	17.6
	Wild K=3	68.5	19.5	22.0
	Wild K=5	70.4	21.3	23.6

A single field image is harder than the curated image for both evaluated models. With three or five independent field views, both multiple-choice recognition and open-ended naming improve beyond the corresponding curated baseline. These findings apply to the two models evaluated on AgriTaxon-Wild and are not generalized to all 14 Core models.

Licensing and Attribution

• Benchmark annotations and metadata: CC BY 4.0.
• Core images (images/): Wikimedia Commons images under their original licenses, predominantly CC BY and CC BY-SA.
• Wild images (wild/images/): iNaturalist photographs under their individual licenses, predominantly CC BY-NC. Per-photo licenses and credits are provided in wild/attributions.jsonl and summarized in wild/LICENSES.md.
• Source code, prompts, and project documentation: MIT License.
• Authority identifiers: Wikidata QIDs are available under CC0; FAO and EPPO identifiers are used for reference linking only.

Wild license distribution:

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

Because AgriTaxon-Wild includes NonCommercial and NoDerivatives images, the repository metadata uses a mixed license. Users must consult the per-photo license and attribution record before reuse. In particular, the Wild subset should be treated as non-commercial, and the 215 NoDerivatives photographs must not be modified.