PandyaJeet/gemma4-plant-pathology-3k

🌿 Gemma 4 Multimodal Plant Pathology & Spatial Grounding Dataset (3.75K)

gemma4-plant-pathology-3k is a curated, class-balanced, multi-domain Vision-Language Dataset engineered specifically for fine-tuning Gemma 4 E4B-IT (and similar multimodal Vision-Language Models) on visual plant disease diagnosis, spatial lesion localization, differential pathology, and out-of-distribution (OOD) safety guardrails.

Unlike traditional classification datasets that only map an image to a single integer label, this dataset formats every sample as a structured Visual Chain-of-Thought (VCoT) conversation (system $\rightarrow$ user $\rightarrow$ assistant). The model is trained to first describe macroscopic visual symptoms, localize diseased regions using normalized [ymin, xmin, ymax, xmax] (0–1000 scale) bounding boxes, rule out look-alike conditions, and emit a deterministic disease_code for downstream remediation lookup.

---

📊 Dataset Summary & Split Architecture

Metric	Specification
Total Curated Samples	3,750 RGB Images + Multi-Turn VLM Conversations
Training Split (train)	3,375 samples (exactly 225 per class across 15 classes)
Validation Split (validation)	375 samples (exactly 25 per class across 15 classes)
Target Crops	Tomato (Solanum lycopersicum), Potato (Solanum tuberosum), Corn / Maize (Zea mays)
Total Canonical Classes	15 (13 Crop Health/Disease Classes + 2 Safety Guardrail Classes)
Bounding Box Format	Normalized Integer [ymin, xmin, ymax, xmax] on a [0, 1000] grid
Diseased Samples with Boxes	2,750 / 2,750 (100% coverage; zero full-frame [0, 0, 1000, 1000] boxes)
Healthy / Invalid with [] Boxes	1,000 / 1,000 (100% verified empty [] arrays)

---

🧠 Why Multi-Domain Curation? (Solving Real-World VLM Failure Modes)

Models trained exclusively on studio datasets (like PlantVillage) achieve >98% lab accuracy but drop below 35% in real farm environments due to background shortcut learning. Conversely, models trained without rejection classes hallucinate diseases when shown insects, soil, or blurry photos.

This dataset fuses four distinct visual domains into a single unified distribution:

1. Textbook Morphological Anchors (PlantVillage / Studio Lab):
   • High-contrast studio leaves on neutral backgrounds teach the Vision Transformer's attention heads fine-grained lesion geometry (e.g., concentric target rings in Early Blight vs. pale grey pycnidia specks in Septoria Leaf Spot).
2. In-The-Wild Field Captures (PlantDoc / Outdoor Farm):
   • Noisy outdoor captures featuring overlapping foliage, harsh sunlight glare, soil backgrounds, and human hands force the model to attend strictly to the leaf blade and ignore background clutter.
3. Abiotic & Pest Guardrail (Agricultural Pests):
   • Field images of mechanical insect chewing, aphids, beetles, and leaf miners train the model to block false-positive fungicide prescriptions (GUARDRAIL_NUTRIENT_PEST).
4. Out-of-Distribution & Defocus Shield (Natural Images + Synthetic Gaussian Defocus):
   • Combines non-plant everyday objects (vehicles, animals, indoor scenes) with severely out-of-focus leaf photos ($\sigma = 15\text{ px}$ Gaussian blur) to train an explicit rejection circuit (GUARDRAIL_INVALID_IMAGE).

---

🗂️ The 15 Canonical Classes

Each of the 15 classes contains exactly 250 total samples (225 train / 25 validation):

#	Canonical disease_code	Crop / Category	Condition Name	Status	Box Policy
1	TOMATO_HEALTHY	Tomato	Healthy Leaf	Healthy	Empty []
2	TOMATO_EARLY_BLIGHT	Tomato	Early Blight (Alternaria solani)	Diseased	[ymin, xmin, ymax, xmax]
3	TOMATO_LATE_BLIGHT	Tomato	Late Blight (Phytophthora infestans)	Diseased	[ymin, xmin, ymax, xmax]
4	TOMATO_SEPTORIA_SPOT	Tomato	Septoria Leaf Spot (Septoria lycopersici)	Diseased	[ymin, xmin, ymax, xmax]
5	TOMATO_BACTERIAL_SPOT	Tomato	Bacterial Spot (Xanthomonas spp.)	Diseased	[ymin, xmin, ymax, xmax]
6	TOMATO_YELLOW_CURL_VIRUS	Tomato	Tomato Yellow Leaf Curl Virus (TYLCV)	Diseased	[ymin, xmin, ymax, xmax]
7	POTATO_HEALTHY	Potato	Healthy Leaf	Healthy	Empty []
8	POTATO_EARLY_BLIGHT	Potato	Early Blight (Alternaria solani)	Diseased	[ymin, xmin, ymax, xmax]
9	POTATO_LATE_BLIGHT	Potato	Late Blight (Phytophthora infestans)	Diseased	[ymin, xmin, ymax, xmax]
10	CORN_HEALTHY	Corn (Maize)	Healthy Leaf	Healthy	Empty []
11	CORN_COMMON_RUST	Corn (Maize)	Common Rust (Puccinia sorghi)	Diseased	[ymin, xmin, ymax, xmax]
12	CORN_NORTHERN_BLIGHT	Corn (Maize)	Northern Corn Leaf Blight (Exserohilum turcicum)	Diseased	[ymin, xmin, ymax, xmax]
13	CORN_GRAY_LEAF_SPOT	Corn (Maize)	Gray Leaf Spot (Cercospora zeae-maydis)	Diseased	[ymin, xmin, ymax, xmax]
14	GUARDRAIL_NUTRIENT_PEST	Guardrail	Pest Damage or Abiotic / Nutrient Stress	Abiotic_Or_Pest_Stress	[ymin, xmin, ymax, xmax]
15	GUARDRAIL_INVALID_IMAGE	Guardrail	Invalid, Non-Plant, or Severely Blurry Image	Invalid_Image	Empty []

---

🎯 Multi-Tier Spatial Bounding Box Pipeline

To ensure high-precision spatial grounding without lazy full-frame [0, 0, 1000, 1000] shortcuts, bounding boxes for all 2,750 symptomatic images were curated through a 3-stage verification pipeline:

1. Human-Annotated Ground Truth (PlantDoc Pascal VOC XMLs):
   • Extracted and normalized 1,092 multi-box field annotations directly from official Pascal VOC XML files into [ymin, xmin, ymax, xmax] (0–1000 scale).
2. Open-Vocabulary Zero-Shot Detection (google/owlv2-base-patch16-ensemble):
   • Prompts targeting "diseased leaf spot", "necrotic lesion", "fungal blight", and "leaf rust" localized high-confidence lesions across unboxed samples.
3. HSV Chlorosis/Necrosis Hotspot & Contour Refinement:
   • A custom computer-vision verification gate scanned all 3,750 records and replaced 903 full-frame canopy boxes with tight symptomatic sub-boxes targeting yellow chlorotic halos and brown necrotic tissue (Hue 8–36, morphological closing, and border-margin clamping).

---

💬 Data Schema & Example Payload

Every row in the dataset contains two top-level columns:

• image: Decoded PIL.Image (RGB) embedded directly inside the Parquet shard.
• messages: Standard OpenAI/HuggingFace multimodal chat array (system, user, assistant) compatible out-of-the-box with Unsloth's UnslothVisionDataCollator and TRL's SFTTrainer.

Sample assistant JSON Ground Truth:

{
  "visual_evidence": "Outdoor field capture of Tomato exhibiting Severe upward curling and cupping of leaflets with prominent marginal and interveinal yellow chlorosis and thickened, crumpled leaf texture.",
  "lesion_boxes": [
    {
      "box_2d": [186, 55, 426, 453],
      "label": "tomato_yellow_curl_virus"
    },
    {
      "box_2d": [15, 570, 166, 758],
      "label": "tomato_yellow_curl_virus"
    }
  ],
  "species": {
    "common_name": "Tomato",
    "scientific_name": "Solanum lycopersicum"
  },
  "diagnosis": {
    "disease_code": "TOMATO_YELLOW_CURL_VIRUS",
    "condition_name": "Tomato Yellow Leaf Curl Virus (TYLCV)",
    "status": "Diseased",
    "ruled_out": "Ruled out physiological heat curl and nutrient deficiency due to severe upward cup-like curling, stunted leaflet size, and interveinal yellowing."
  },
  "immediate_summary": "Viral infection vectored by whiteflies detected. Rogue infected plants and control whitefly population immediately."
}