README.md

---
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: cc-by-nc-4.0
task_categories:
- image-text-to-text
size_categories:
- 1K<n<10K
dataset_info:
  features:
  - name: images
    list:
      image:
        decode: false
  - name: id
    dtype: string
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      list:
      - name: type
        dtype: string
      - name: text
        dtype: string
  - name: origin_dataset
    dtype: string
  - name: raw_metadata
    dtype: string
  splits:
  - name: train
    num_bytes: 2118998617
    num_examples: 4342
  download_size: 2118414692
  dataset_size: 2118998617
---

# AgroBench

AgroBench is a vision-language model (VLM) benchmark for agriculture, annotated by expert agronomists. It covers seven agricultural topics spanning 203 crop categories and 682 disease categories, with 4,342 question-answer examples pairing images with multiple-choice questions across tasks such as crop identification, disease diagnosis, pest identification, and weed identification.

This dataset has been standardized to the HF `image_text_to_text` format: one conversational `messages` schema, imagefolder-native images, and (if present) a `text_only` parquet config.

This dataset is indexed on https://project-agml.github.io/ as part of the AgML python library.

## Usage

```python
from datasets import load_dataset

# Single image folder -> one default config
ds = load_dataset("Project-AgML/AgroBench")

# Stream without downloading
ds = load_dataset("Project-AgML/AgroBench", streaming=True)
```
Every record shares the SAME columns so heterogeneous AgML datasets concatenate cleanly: `id`, `file_names` (one image per row), `messages`, `origin_dataset`, and `raw_metadata`. `raw_metadata` is a JSON-encoded string holding source fields not folded into `messages`/`file_names` (here: the original `source_image` filename and parsed `crop_prefix`); restore it with `json.loads(row["raw_metadata"])`. Image placeholders in `messages` align 1:1 with `file_names`. Using a JSON string (not a native struct) is what lets `concatenate_datasets([...])` work across datasets whose raw fields differ in type. Multi-image rows return images as a list aligned to the `{"type": "image"}` placeholders in messages.

# Citation

```bibtex
@misc{shinoda2025agrobench,
      title={AgroBench: Vision-Language Model Benchmark in Agriculture},
      author={Shinoda, Risa and Inoue, Nakamasa and Kataoka, Hirokatsu and Onishi, Masaki and Ushiku, Yoshitaka},
      year={2025},
      eprint={2507.20519},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.20519}
}

Shinoda, Risa; Inoue, Nakamasa; Kataoka, Hirokatsu; Onishi, Masaki; Ushiku, Yoshitaka (2025), "AgroBench: Vision-Language Model Benchmark in Agriculture", arXiv:2507.20519
```