Haseeb9t/CropCare-MLLM-Adapters

🌿 CropCare: Fine-Tuned Multimodal LoRA Adapters for Potato Leaf Disease Diagnosis

Official model weights and fine-tuned LoRA adapters for the paper:
"Evidence-Bound Instruction Tuning for Potato Leaf Disease Classification: A Multi-Stage Vision-Language Pipeline"
by Haseeb Mehmood (Center for Machine Vision and Signal Analysis, University of Oulu).

🔗 GitHub Repository: https://github.com/Lethaldroid/CropCare-MLLM
📊 Kaggle Dataset: Potato Leaf Disease with AI-Generated Captions

---

🏆 Benchmark Performance (Table V in Paper, $N=466$ Held-Out Images)

Model Architecture	Parameters	Method	Strict Accuracy	Early Blight Recall	Late Blight Recall	Multi-Label Adj.	Macro F1
Gemma 4 26-a4b	26B (4B active)	BF16 LoRA	92.1%	91.1%	93.2%	93.3%	92.0%
Qwen 3.8	27B dense	BF16 LoRA	86.0%	88.3%	83.6%	92.1%	86.0%
Llama 3.2 Vision	90B cross-attn	4-bit QLoRA	85.4%	89.1%	81.3%	87.1%	85.4%
Mistral Small 3.1	24B dense	4-bit QLoRA	54.7%	98.8%	5.0%	58.1%	39.6%

---

📦 Hosted Checkpoints & Adapters

This repository hosts the three best fine-tuned LoRA adapter packages:

1. gemma4_31b_potato_best_lora.zip (1.79 GB):
   • Fine-tuned adapter for Gemma 4 26-a4b (google/gemma-4-26-a4b-it).
   • Achieves state-of-the-art 92.1% strict diagnostic accuracy and 92.0% Macro F1.
2. llama32_vision_90b_potato_best_lora.zip (928 MB):
   • Fine-tuned adapter for Llama 3.2 Vision 90B (meta-llama/Llama-3.2-90B-Vision-Instruct).
3. mistral_small_31_24b_potato_best_lora.zip (361 MB):
   • Fine-tuned adapter for Mistral Small 3.1 24B (mistralai/Mistral-Small-24B-Instruct-2501).

Each archive contains the complete PEFT adapter:

• adapter_config.json
• adapter_model.safetensors
• chat_template.jinja
• processor_config.json
• tokenizer.json & tokenizer_config.json
• training_args.bin

---

🚀 How to Load and Use

Unzip the target adapter and load it via Unsloth or PEFT / Transformers:

import zipfile
from unsloth import FastVisionModel

# 1. Unzip the downloaded adapter
with zipfile.ZipFile("gemma4_31b_potato_best_lora.zip", "r") as zip_ref:
    zip_ref.extractall("gemma4_lora_adapter")

# 2. Load the base model with the fine-tuned LoRA adapter
model, tokenizer = FastVisionModel.from_pretrained(
    model_name="gemma4_lora_adapter",
    load_in_4bit=True,
)
FastVisionModel.for_inference(model)

---

📖 Citation

@article{cropcare2026evidence,
  title={Evidence-Bound Instruction Tuning for Potato Leaf Disease Classification: A Multi-Stage Vision-Language Pipeline},
  author={Mehmood, Haseeb},
  journal={Center for Machine Vision and Signal Analysis, University of Oulu},
  year={2026}
}