README.md

---
license: apache-2.0
dataset_info:
  features:
  - name: question_id
    dtype: string
  - name: image_folder
    dtype: string
  - name: image_name
    dtype: string
  - name: question_type
    dtype: string
  - name: question
    dtype: string
  - name: A
    dtype: string
  - name: B
    dtype: string
  - name: C
    dtype: string
  - name: D
    dtype: string
  - name: answer
    dtype: string
  - name: image
    dtype: image
  splits:
  - name: test
    num_bytes: 3497427279
    num_examples: 54580
  download_size: 3267412760
  dataset_size: 3497427279
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
extra_gated_fields:
  Full Name: text
  Email Address: text
  Institution or Company: text
  Country: country
  Intended Use of the Dataset:
    type: select
    options: 
      - Academic Research
      - Educational Purpose
  I agree to use this dataset for non-commercial purposes only: checkbox
---
# LeafBench 2.0

**LeafBench 2.0** is a visual question answering (VQA) benchmark for evaluating **fine-grained plant disease understanding** in vision-language models (VLMs). Derived directly from [LeafNet 2.0](https://huggingface.co/datasets/your-username/LeafNet2.0), the benchmark consists of multiple-choice questions spanning **9 complementary plant pathology tasks**, designed to assess disease understanding beyond coarse category recognition.

LeafBench 2.0 was evaluated across **16 VLMs** including 7 CLIP-based models, 7 generative VLMs, and 2 proprietary models (GPT-4o, Gemini 2.5 Pro), revealing substantial performance gaps between coarse recognition tasks and fine-grained pathological reasoning.

This benchmark accompanies the paper:

> **LeafNet 2.0: A Multiregional Image–Text Dataset for Vision-Language Modeling and Reasoning of Plant Diseases**  
> Trang V. Nguyen, Khang Nguyen Quoc, David Harwath, Phuong D. Dao  
> The University of Texas at Austin · Korea University

---

## Benchmark at a Glance

| Property | Value |
|---|---|
| Total benchmark instances | 6,361 |
| Evaluation tasks | 9 |
| Answer format | Multiple choice (A / B / C / D) |
| Distractors per question | 3 |
| Source dataset | LeafNet 2.0 (255,825 images) |
| Models evaluated | 16 (7 CLIP-based, 7 generative, 2 proprietary) |
| Best overall accuracy | 67.78% (Gemini 2.5 Pro) |
| Best open-source accuracy | 60.02% (Qwen3-VL-4B) |

---

## The 9 Benchmark Tasks

| Task | Abbreviation | Description |
|---|---|---|
| Disease Identification | DI | Identify the disease present in the leaf image |
| Pathogen Classification | PC | Classify the causal pathogen type (fungus, bacterium, virus, etc.) |
| Crop Species Identification | CSI | Identify the crop species shown in the image |
| Symptom Identification | SI | Identify the specific visible symptom(s) |
| Healthy/Diseased Classification | HDC | Determine whether the leaf is healthy or diseased |
| Scientific Name Classification | SNC | Assign the correct scientific name to the disease or pathogen |
| Lesion Identification | LI | Identify lesion type, morphology, or distribution pattern |
| Leaf Symptom Detection | LSD | Detect the presence or location of symptomatic regions |
| Disease Severity Classification | DSC | Classify the severity level of the visible disease |

Tasks range from coarse recognition (HDC, CSI) to fine-grained pathological reasoning (PC, SNC, LI), providing a comprehensive evaluation spectrum for plant disease AI systems.

---

## Dataset Structure

```
leafbench2.0/
├── images/
│   ├── [crop1]-[disease1]/
│   │   ├── [crop1]-[disease1]-id_001.jpg
│   │   ├── [crop1]-[disease1]-id_002.jpg
│   │   └── ...
│   ├── [crop2]-[disease2]/
│   │   └── ...
│   └── ...
└── leafbenchv2.csv       ← full benchmark annotation file
```

### `leafbenchv2.csv` columns

| Column | Description |
|---|---|
| `image_path` | Relative path to the benchmark image |
| `task` | Task category (e.g., `DI`, `PC`, `CSI`, `SI`, `HDC`, `SNC`, `LI`, `LSD`, `DSC`) |
| `question` | Task-specific multiple-choice question |
| `choice_a` | Answer option A |
| `choice_b` | Answer option B |
| `choice_c` | Answer option C |
| `choice_d` | Answer option D |
| `ground_truth` | Correct answer label (`A`, `B`, `C`, or `D`) |

---

## Benchmark Results

Evaluation was conducted across 16 VLMs using classification accuracy per task. Results reveal clear specialization differences across model families.

### Overall Accuracy (Top Models)

| Model | Type | Overall Accuracy |
|---|---|---|
| Gemini 2.5 Pro | Proprietary | **67.78%** |
| GPT-4o | Proprietary | 64.72% |
| Qwen3-VL-4B | Generative (open) | 60.02% |
| AgriCLIP | CLIP-based (domain) | ~65% (DI task) |
| SCOLD | CLIP-based (domain) | Competitive on SI tasks |

### Task-Level Highlights

| Task | Easiest Model | Hardest for Most Models |
|---|---|---|
| Healthy/Diseased Classification (HDC) | GPT-4o: **93.50%** | No — most models perform well |
| Crop Species Identification (CSI) | Gemini 2.5 Pro: **76.80%** | Moderate difficulty |
| Leaf Symptom Detection (LSD) | Gemma4-8B: **92.76%** | No |
| Pathogen Classification (PC) | All models struggle | **Yes** — requires subtle discrimination |
| Scientific Name Classification (SNC) | All models struggle | **Yes** — requires domain knowledge |
| Lesion Identification (LI) | All models struggle | **Yes** — subtle morphological cues |

**Key finding:** Agriculture-adapted models (AgriCLIP, SCOLD) consistently outperformed several larger general-domain architectures on symptom-oriented tasks (SI, LSD), demonstrating the value of domain-specific pretraining. Most models achieved >90% on HDC but dropped substantially on PC, SNC, and LI, confirming that LeafBench 2.0 captures meaningful fine-grained complexity beyond coarse disease recognition.

---

## Usage

### Load benchmark annotations

```python
from datasets import load_dataset

ds = load_dataset("your-username/LeafBench2.0", name="benchmark")
print(ds["test"][0])
# → {
#     "image_path": "coffee-miner/coffee-miner-id_001.jpg",
#     "task": "DI",
#     "question": "What disease is visible on this leaf?",
#     "choice_a": "Coffee Leaf Miner",
#     "choice_b": "Coffee Rust",
#     "choice_c": "Brown Eye Spot",
#     "choice_d": "Healthy",
#     "ground_truth": "A"
#   }
```

### Load with images for direct model evaluation

```python
ds = load_dataset("your-username/LeafBench2.0", name="benchmark_with_images")
print(ds["test"][0])
# → {"image": <PIL.Image>, "task": "PC", "question": "...", ..., "ground_truth": "B"}
```

### Filter by task

```python
ds = load_dataset("your-username/LeafBench2.0", name="benchmark")
pc_subset = ds["test"].filter(lambda x: x["task"] == "PC")
```

---

## Evaluation Script

The official evaluation code, model implementations, and reproduction scripts are available at:

🔗 **[https://github.com/EnalisUs/LeafBench](https://github.com/EnalisUs/LeafBench)**

**Environment:**
```
Python 3.10
torch==2.7.0
transformers==4.51.3
opencv-python==4.11.0.86
accelerate==1.8.1
torchvision==0.22.0
peft==0.15.0
```

---

## Intended Use

LeafBench 2.0 is designed for:

- **Fine-grained VLM evaluation** — assessing plant disease understanding across 9 pathology tasks with varying difficulty levels.
- **Agricultural domain adaptation benchmarking** — comparing general-domain and agriculture-adapted models on symptom-level reasoning.
- **Diagnostic reasoning research** — studying whether multimodal models learn biologically meaningful symptom features or rely on superficial visual correlations.
- **Zero-shot and few-shot evaluation** — testing model generalization to unseen crop-disease combinations or geographic distributions.
- **Multimodal reasoning studies** — examining causal interpretation, uncertainty estimation, and differential diagnosis in plant pathology.

---

## Relationship to LeafNet 2.0

LeafBench 2.0 is derived directly from the LeafNet 2.0 evaluation subset (6,361 image–caption pairs), preserving the same variability in:

- Imaging devices and conditions (smartphones, digital cameras, controlled/natural backgrounds)
- Geographic and environmental diversity (9 regions)
- Disease severity and progression stages (early/late)
- Crop and disease coverage (37 species, 197 classes)

This ensures that benchmark performance reflects real-world agricultural conditions rather than idealized controlled settings.

---

## Limitations

- The benchmark covers 9 tasks but does not include open-ended generation tasks (e.g., free-form caption generation or differential diagnosis). Future versions may extend to these settings.
- Performance on PC, SNC, and LI tasks is generally low across all current architectures, suggesting these remain open research challenges rather than solved problems.
- Task difficulty is inherently tied to disease and crop distribution in LeafNet 2.0; rare classes may be underrepresented in the benchmark.
- As with LeafNet 2.0, a small proportion of images exhibit ambiguous stage-specific features that may affect ground-truth reliability for the DSC task.

---

## License

This dataset is released under the [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/) license.