TomaMMU

TomaMMU is a tomato leaf disease VQA dataset built by researchers at FPT University (Vietnam) and Korea University. The motivation behind it is pretty simple: most tomato disease models are trained and tested on clean, lab-quality images, but accuracy falls off a cliff (30-40%) once you point them at a real field photo with messy backgrounds and inconsistent lighting. TomaMMU tries to close that gap by pairing field- and lab-collected photos with human-annotated questions that go well beyond "sick or healthy" — asking models to identify the crop, count leaves, name the pathogen, and even give the scientific name of the disease.

There are two papers behind this work. The first introduces the dataset and a benchmark (TomaBench) for evaluating VLMs on it. The second, SOLAR, builds a fine-tuned model on top of it.

• TomaMMU: A Comprehensive Multimodal Understanding Benchmark for Tomato Leaf Diseases — Truong, Quoc, Quach (arXiv:2608.08727)
• A Multi-Modal Generative Model for Tomato Disease Leaves Understanding — Quoc, Tran, Truong, Quach (arXiv:2609.19555) — the SOLAR model paper, trained on this dataset
• Original repository: huggingface.co/datasets/enalis/TomaMMU

What's in the papers

The images come from LeafNet and TOM2024, mixing field shots with lab photos, and cover 15 disease categories (14 diseases plus healthy) with 25 distinct symptoms. Every question was human-annotated rather than machine-generated, which the authors call out as a deliberate contrast to lab-only benchmarks like PlantVillage.

Questions fall into seven types, organized into three difficulty tiers:

Level	Task types
1: Basic Perception	HDC (Healthy-Diseased Classification), LC (Leaf Count), CSI (Crop Species Identification)
2: Pathology Understanding	SI (Symptom Identification), PC (Pathogen Classification)
3: Expert Diagnosis	DC (Disease Classification), SNC (Scientific Name Classification)

The first paper's own eval setup (TomaBench) splits things into 124,367 pairs for fine-tuning plus a held-out set of 42,626 multiple-choice and 37,816 open-ended questions. They benchmarked a wide range of models — Gemini 2.0 Flash, InternVL3, Qwen2.5/3, Gemma4, SmolVLM2, LLaVA-1.5, LFM2.5-VL, plus CLIP-family models (CLIP-ViT-L-14, SigLIP2, BioCLIP, SCOLD). The best zero-shot model only hit 61.3% MCQ accuracy, and open-ended answers were a lot worse (~2.5/5 on GPT-graded scoring across the board). Fine-tuning closes most of that gap: their TomaLLaVA model reached 96.09% on the MCQ set.

The second paper (SOLAR) picks up from there and builds an actual diagnostic model rather than just benchmarking existing ones — a vision encoder (SCOLD) feeding into a sparse mixture-of-experts layer that routes attention based on the question, then an LFM-VL 2.5 decoder for generating the answer. It hits 97.68% on TomaMMU and holds up reasonably (69% average) across three other external tomato-disease datasets it wasn't trained on. Interestingly, symptom identification is still the weakest link (7.8% error rate) and mistakes there tend to cascade into wrong disease/pathogen answers downstream — worth keeping in mind if you're using this data to train or evaluate something similar.

Neither paper nor the original dataset card lists a license, so this repo's card leaves it as unspecified rather than guessing.

What's standardized here

This repo has the full 213,119-pair corpus in a single train config — it doesn't reproduce the paper's own 60/20/20 train/eval split, since the source CSV has no split column to begin with.

Config	Rows (QA pairs)	Unique images	Image data
train	213,119	28,112	0.31GB (1 shard)

One thing worth flagging: the source Tomato.csv lists 36,993 distinct image_path strings, but a good chunk of those are just case differences (Tomato___Bacterial_Floundering vs Tomato___Bacterial_floundering) that point at the same file on a case-insensitive filesystem. Rather than dedup by path string, images here are keyed by a SHA256 hash of their actual bytes, which sidesteps the casing issue entirely. That's how 36,993 path strings end up as 28,112 unique stored images, matching the paper's reported 28,808 real files once you account for the hash collapsing true duplicates.

Question type breakdown:

Question type	Options
SI (Symptom Identification)	4
HDC (Healthy-Diseased Classification)	2
CSI (Crop Species Identification)	4
LC (Leaf Count)	2
DC (Disease Classification)	4
PC (Pathogen Classification)	4
SNC (Scientific Name Classification)	4

All seven types are multiple-choice — there's no open-ended detection-style task in this corpus (unlike the paper's own OEQ eval set, which isn't part of what's standardized here).

Layout

train-0000-of-0001.parquet          # metadata rows
images/
  image-train-000-of-001.zip        # image shard (ZIP_STORED, uncompressed)
  path_to_shard.parquet             # maps image SHA256 hash -> shard_file

Images sit uncompressed inside zip shards, keyed by hash, capped at 5GB per shard (this dataset only needed one).

Schema

Column	Type	Description
images	list<struct<bytes, path>>	bytes is null; path is the SHA256 hash, looked up in images/path_to_shard.parquet
id	string	tomammu_train_<n>
messages	chat-style list	user turn is the image plus the question, with whichever lettered options are actually present in the source (some rows only have A/B); assistant turn combines the answer's letter with its text, e.g. "B. Bacterial Floundering", rather than just the bare text
raw_metadata	JSON string	image_path (the original source path), question_id, question_type

Each row is one image-question pair, so the same image shows up across as many as seven rows.

Citations

@article{truong2026tomammu,
  title={TomaMMU: A Comprehensive Multimodal Understanding Benchmark for Tomato Leaf Diseases},
  author={Truong, Gia-Han and Quoc, Khang Nguyen and Quach, Luyl-Da},
  journal={arXiv preprint arXiv:2608.08727},
  year={2026}
}

@article{quoc2026solar,
  title={A Multi-Modal Generative Model for Tomato Disease Leaves Understanding},
  author={Quoc, Khang Nguyen and Tran, Minh-Phuoc and Truong, Gia-Han and Quach, Luyl-Da},
  journal={arXiv preprint arXiv:2609.19555},
  year={2026}
}

---

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