Dataset Card for TomaMMU

Dataset Summary

TomaMMU (Tomato leaf disease MultiModal Understanding) is a large-scale, standardized dataset and multimodal benchmark designed to evaluate Vision-Language Models (VLMs) on plant pathology, specifically tomato leaf diseases. Unlike traditional datasets that frame diagnosis as a simple visual classification problem, TomaMMU emphasizes multimodal disease understanding, reasoning, and real-world agricultural evaluation. The dataset includes TomaBench, a comprehensive benchmarking framework organizing agricultural tasks into a hierarchical taxonomy spanning from basic perception to expert diagnosis.

Dataset Structure

Data Instances

The dataset is composed of high-quality image data paired with human-annotated Visual Question Answering (VQA) pairs. The data covers 15 distinct categories (14 tomato diseases and 1 healthy category) and includes 25 different tomato disease symptoms. The images are formatted as 224x224 RGB.

Data Splits

The raw metadata dataset is proportionally divided into three splits:

• Train: 60% of the data
• Validation: 20% of the data
• Test: 20% of the data

In terms of VQA pairs, the final processed dataset comprises 213,119 total conversations:

• Training Set: 124,367 VQA pairs specialized for fine-tuning VLMs.
• TomaBench (Evaluation Set): 42,626 Multiple-Choice Questions (MCQs) and 37,816 Open-Ended Questions (OEQs) designed for zero-shot benchmarking.

Task Taxonomy

The dataset evaluates models across seven specific agricultural tasks, which are grouped into three hierarchical domains:

1. Basic Perception

• Healthy-Diseased Classification (HDC): Identifies the basic presence of pathology.
• Leaf Count (LC): Evaluates leaf condition and quantity.
• Crop Species Identification (CSI): Evaluates host crop species identification.

2. Pathology Understanding

• Symptoms Identification (SI): Focuses on fine-grained visual manifestations like spots, chlorosis, and mosaic symptoms.
• Pathogen Classification (PC): Identifies the pathogen category responsible for visual symptoms.

3. Disease Diagnosis

• Disease Classification (DC): Targets specific pathological conditions.
• Scientific Name Classification (SNC): Predicts the formal scientific name of the disease.

Dataset Creation

Curation Rationale

TomaMMU was built to address the generalization gap and environmental variability issues found in models trained solely on uniform, laboratory-acquired images. It is designed to combine visual observation with background diagnostic knowledge to test a model's true pathological reasoning.

Source Data

The images are collected from two primary sources:

• In-the-wild sources: The majority of the images are captured directly from real farms to reflect natural visual variability, background clutter, and diverse illumination.
• In-house collections: Incorporates data from existing authoritative sources such as LeafNet.

Annotation Process

The dataset was constructed through a rigorous three-stage pipeline:

1. Data Collection: Gathering in-situ and laboratory images.
2. Human Annotation and Categorization: Mislabeled data was removed, and human domain experts strictly executed a metadata extraction pipeline. Experts manually extracted diagnostic knowledge, symptom descriptions, and verified every data point for agronomic reliability.
3. QA Generation: Single images were associated with up to seven distinct question types to balance agricultural knowledge representation. MCQs were created using the extracted facts as ground truth against three incorrect options, while OEQs were designed to require short-form responses.

Considerations for Using the Data

Baseline Evaluations

TomaBench is highly challenging for current state-of-the-art models. The evaluation of 14 advanced VLMs (including Gemini 2.0 Flash, Qwen-based models, and BioCLIP) revealed significant performance gaps in fine-grained recognition and factually grounded reasoning without task-specific training. Simple fine-tuning utilizing this dataset (e.g., the resulting TomaLLaVA model) achieved up to 96.09% accuracy on the challenging MCQs, demonstrating the dataset's efficacy for domain adaptation.