AgroBench

AgroBench is a benchmark that tests vision-language models on agriculture. It has 4,342 multiple-choice questions across 7 topics, annotated by agriculture experts. It covers 203 crop categories, 682 disease categories, 134 pest categories and 108 weed categories.

The authors found that current models struggle with fine-grained identification. Weed identification is the hardest task, and most open-source models score close to random on it.

The weed images come from the 4 datasets listed in source_datasets. The other images come from websites supervised by plant pathologists and from Creative Commons or publicly available datasets.

• Paper: AgroBench: Vision-Language Model Benchmark in Agriculture (ICCV 2025)
• Original dataset: risashinoda/AgroBench
• Code: dahlian00/AgroBench
• License: the authors do not state a license.

Tasks

Task	source	Rows
Disease Identification	did	1,502
Pest Identification	pid	544
Weed Identification	wid	609
Crop Management	cmn	411
Disease Management	dmn	569
Machine Usage QA	mqa	303
Traditional Management	tm	404

Structure

train-0000-of-0001.parquet
images/
  image-train-000-of-001.zip
  path_to_shard.parquet

Each row has:

• images: list with one {"bytes": null, "path": <sha256>}. The path is the SHA256 of the image bytes and points into a zip shard through images/path_to_shard.parquet.
• id: agrobench_<source>_<counter>, for example agrobench_did_0.
• messages: one user turn (image, then question with options) and one assistant turn (answer).
• raw_metadata: JSON string with the source fields category, source, id and raw_image_paths. Depending on the task it also has crop, category2 and reasoning (a short explanation of the answer).

Example:

{
  "id": "agrobench_did_0",
  "messages": [
    {"role": "user", "content": [
      {"type": "image", "text": null},
      {"type": "text", "text": "What disease is affecting the plant?\nA. Northern Leaf Blight\nB. Pythium root rot\nC. Witches'-broom\nD. Purple stain\nE. Bacterial dieback of nectarine"}
    ]},
    {"role": "assistant", "content": [{"type": "text", "text": "E. Bacterial dieback of nectarine"}]}
  ],
  "raw_metadata": "{\"category\": \"Bacteria\", \"crop\": \"Nectarine\", \"source\": \"did\", \"id\": \"729\", \"raw_image_paths\": [\"png/did_729.png\"]}"
}

Changes from the original data

• Built from the original dataset.jsonl and png/ files. They have the same rows and images as the original parquet files, plus the reasoning and category2 fields.
• All 7 tasks are in one train split. The source field in raw_metadata gives the task.
• Options are added to the question as lines like A. text. All questions have 5 options except one, which has 6.
• Answers combine the letter and the option text, for example E. Bacterial dieback of nectarine. One question lists the correct answer twice (A and E), so A is used.
• 6 answers do not exactly match any option text. They are kept as in the source, without a letter.
• raw_metadata keeps the source fields except question, options, answer and image, which are already in messages. The image path is kept in raw_image_paths. The reasoning text stays in raw_metadata and is not added to the answer.
• Images are stored once, deduplicated by SHA256. The 4,342 rows use 3,742 unique images.

Usage

from agml import loadImageTextToTextDataset

dataset = loadImageTextToTextDataset("Project-AgML/AgroBench")

sample = dataset["train"][0]
sample["images"]    # list of PIL images
sample["messages"]

Citation

@InProceedings{Shinoda_2025_ICCV,
  author    = {Shinoda, Risa and Inoue, Nakamasa and Kataoka, Hirokatsu and Onishi, Masaki and Ushiku, Yoshitaka},
  title     = {AgroBench: Vision-Language Model Benchmark in Agriculture},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  month     = {October},
  year      = {2025},
  pages     = {7634-7644}
}

This dataset is indexed and structured on https://project-agml.github.io/ as part of Project AgML @ UC Davis. This dataset was reformatted from its original format to match HuggingFace's Imagefolder standards but requires an external module (agml) that processes and returns a HF Dataset object faster than HF module functions.