Qv3-Plant_spotter-Vl-8b

Qv3-Plant_spotter-Vl-8b

Plant Disease Identification Vision-Language Model

Qv3-Plant_spotter-Vl-8b is a fine-tuned multimodal model for plant disease identification, built upon Qwen3-VL-8B-Instruct. It accepts an image of a plant leaf and identifies the disease or condition present.

Model Architecture

Language Model

Vision Encoder

LoRA Configuration

Training Details

Datasets

The model was fine-tuned on the following plant-related datasets, all in sharegpt multimodal format with the prompt template <image>Identify the plant disease from this image.:

All datasets were converted to a unified sharegpt format with <image> placeholder for visual input.

Training Hyperparameters

Training Procedure

• Vision Tower: Frozen (preserves general visual understanding)
• Multi-modal Projector: Unfrozen (adapts visual-to-text mapping to plant domain)
• LoRA Target: All linear layers (both LLM and vision encoder)

Training Metrics

Loss Trajectory: The loss decreased from 4.36 at step 1 to 0.008 at step 181, with a final step loss of 0.07, demonstrating stable convergence under cosine LR schedule.

Usage

Python (Transformers)

import torch
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration
from PIL import Image

model_path = "Qv3-Plant_spotter-Vl-8b"
processor = AutoProcessor.from_pretrained(model_path)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_path, torch_dtype=torch.bfloat16, device_map="auto"
)

image = Image.open("leaf.jpg").convert("RGB")
prompt = "<image>Identify the plant disease from this image."
inputs = processor(images=image, text=prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=50)
print(processor.decode(output[0], skip_special_tokens=True))

llama.cpp (GGUF)

./llama-cli \
    -m Qv3-Plant_spotter-Vl-8b-q4_k_m.gguf \
    -mmproj mmproj-Qv3-Plant_spotter-Vl-8b.gguf \
    --image leaf.jpg \
    -p "<image>Identify the plant disease from this image." \
    -n 50

Available Formats

License & Attribution

• Base Model: Qwen3-VL-8B-Instruct (Apache 2.0)
• Fine-tuned Weights: Apache 2.0
• Training Datasets:
  • PlantVillage: CC BY 4.0 (Zenodo)
  • PlantDoc: CC BY 4.0 (HuggingFace)

Limitations

• Trained primarily on controlled/lab and field images; complex real-world scenarios may reduce accuracy
• Disease classes are limited to ~120 categories covered by the training datasets
• The model may inherit biases from the base model and training data distribution
• Not a substitute for professional agricultural diagnosis

Citation

@misc{qv3-plant-spotter-vl-8b,
  author = {},
  title = {Qv3-Plant_spotter-Vl-8b: Plant Disease Identification via LoRA Fine-tuned Qwen3-VL},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/Litxiong/Qv3-Plant_spotter-Vl-8b}
}