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

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

🌐 Project Page · 💻 GitHub

Overview

Multimodal foundation models are used as generative assistants: shown a photo of a crop, an animal, an insect, or a weed, they must answer with its name. AgriTaxon tests whether they can, in two settings on the same entities: recognition (choose the label among four options) and naming (state the name without options). Models recognize far more entities than they can name; AgriTaxon measures this seeing-without-naming gap.

Every entity is linked through Wikidata to an agricultural knowledge base (FAO Ecocrop, FAO DAD-IS, or the EPPO Global Database) and is labelled at species level or finer (for livestock, usually a breed).

Setting	Purpose	Entities	Images
AgriTaxon-Main	Default setting	6,049	6,049 Wiki images
AgriTaxon-Hard	Extended: entities that nearly all models fail	510	510 Wiki images
AgriTaxon-Wild	Extended: field images of the same entities	2,826	14,130 field images

Track	Entities	Knowledge base
Crop	1,960	FAO Ecocrop (Wikidata P4753)
Livestock	172	FAO DAD-IS (Wikidata P3380)
Pest	2,415	EPPO Global Database (Wikidata P3031)
Weed	1,502	EPPO Global Database (Wikidata P3031)
Total	6,049

Dataset Structure

├── annotations/{track}.jsonl          # AgriTaxon, open-ended naming
├── annotations/{track}_mc.jsonl       # AgriTaxon, multiple choice: options, answer, and `hard` (AgriTaxon-Hard)
├── images/{track}/                    # Wiki images, one per entity
├── splits/hard.json                   # AgriTaxon-Hard QIDs (510)
├── wild/                              # AgriTaxon-Wild: annotations, field images, per-photo attributions and licenses
├── coarse_subset/                     # 1,365 genus- and family-level entities, evaluated separately (paper appendix)
├── construction/                      # the five construction steps, with removed ids and Wikidata ranks
├── judge/                             # LLM-judge prompt and its human-annotated validation sample
├── outputs/                           # model outputs with scores
│   ├── main/<model>/{track}_{multiple-choice,open-ended}.jsonl      # 15 models
│   └── wild/k{1,3,5}/<model>/{track}_{multiple-choice,open-ended}.jsonl
└── metadata.json

Tracks are crop, livestock, pest, and weed. Annotation entries carry qid, label, image, track, the knowledge-base identifier (ecocropID, faoID, or eppoCode), and Wikipedia links when available; _mc files add options and answer. Output rows carry qid, subset (main or coarse; Wild rows omit it), track, setting, label, prediction, raw_response, and the score: correct for multiple choice; exact_match, judge_accepts, judge_reason, and correct for open-ended naming.

Every entity is labelled at species level or finer. Entities whose knowledge-base label is a genus or family are kept apart in coarse_subset/, because a photo shows one species and a genus label would score a correct species-level answer as wrong. construction/README.md documents how the 6,049 entities were selected.

Evaluation

• MC accuracy: share of multiple-choice items answered with the correct option.
• OE-EM: the prediction matches the label after normalization; a name more specific than the label (e.g. a subspecies of the labelled species) also counts.
• OE-Acc: additionally credits aliases (common names, taxonomic synonyms, spelling variants) accepted by an LLM judge: the open-weight Qwen3.8-Flash-Next (FP8 checkpoint, reasoning disabled, temperature 0; prompt in judge/judge_prompt.md). On a stratified sample of 579 predictions, the judge agrees with human annotation on 97.8% (OE-EM: 92.8%).

Summary scores are the unweighted mean over the four tracks.

Main Results

Accuracy (%), ranked by OE-Acc within each group. Hard is MC accuracy on AgriTaxon-Hard, which was selected against the 14 models other than Qwen3.8-Flash-Next.

Model	MC accuracy	OE-EM	OE-Acc	Hard
Proprietary models
gemini-3-pro-preview	85.8	38.7	47.3	13.7
doubao-seed-2-0-pro	82.6	36.8	45.8	10.8
doubao-seed-2-0-lite	80.8	32.4	40.5	10.8
gemini-3-flash-preview	85.8	25.0	39.5	11.4
gpt-5	81.4	23.0	31.6	12.9
gpt-5-mini	73.2	16.8	22.2	4.7
claude-haiku-4-5	63.1	5.5	9.2	7.6
Open-source models
kimi-k2.5	76.6	23.6	32.0	3.7
qwen3.8-flash-next	76.4	18.9	25.8	–
qwen3.5-397b-a17b	74.0	16.3	21.3	14.1
qwen3-vl-235b-a22b	70.7	15.1	20.2	3.5
glm-4.6v	65.9	12.3	17.1	2.2
glm-4.6v-flashx	61.7	8.6	12.5	5.9
qwen3-vl-30b-a3b	62.3	7.0	11.7	3.1
qwen3.5-35b-a3b	70.0	5.2	10.6	7.8

AgriTaxon-Wild Results

Model	Images	MC accuracy	OE-EM	OE-Acc
doubao-seed-2-0-lite	Wiki image	81.0	37.9	48.4
	K=1	80.6	31.0	40.3
	K=3	88.1	47.0	56.3
	K=5	89.4	49.7	61.1
qwen3.5-35b-a3b	Wiki image	75.5	7.1	23.1
	K=1	66.2	7.2	15.2
	K=3	75.5	16.0	25.6
	K=5	76.7	18.5	28.8

One field image is less informative than the Wiki image; several field images overtake it.

Usage

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

import json
crop = [json.loads(l) for l in open("AgriTaxon/annotations/crop.jsonl")]
hard = set(json.load(open("AgriTaxon/splits/hard.json"))["qids"])
print(len(crop), "crop entities;", len(hard), "AgriTaxon-Hard entities")

Licensing and Attribution

• Benchmark annotations and metadata: CC BY 4.0.
• Wiki 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.
• Knowledge-base 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	11,837	CC BY-NC-SA	196
CC BY	1,391	CC BY-NC-ND	163
CC0	360	CC BY-SA	143
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 203 NoDerivatives photographs must not be modified.