deepguess/Qwen2.5-VL-7B-Meteorology

Weather Analysis Vision-Language Model (Qwen2.5-VL-7B)

A specialized vision-language model for meteorological image analysis, fine-tuned from Qwen2.5-VL-7B-Instruct.

Model Details

• —Architecture: Qwen2.5-VL (Vision-Language Model)
• —Parameters: 7.6B
• —Base Model: Qwen/Qwen2.5-VL-7B-Instruct
• —Fine-tuning Method: LoRA (Low-Rank Adaptation)
• —Rank (r): 32
• —Alpha: 32
• —Target modules: ['vproj', 'downproj', 'gateproj', 'kproj', 'upproj', 'qproj', 'o_proj']
• —Training Data: Specialized weather and meteorological imagery dataset
• —Checkpoint: checkpoint-7000

Training Statistics

{
  "global_step": 7000,
  "epoch": 2.911837350180693,
  "total_flos": 4.786937654858951e+18,
  "train_loss": ".751"
}

Image Preprocessing Note

This model was trained with images preprocessed to 448x448 resolution. While Qwen2.5-VL supports dynamic resolution:

• —Best performance may be achieved with 448x448 images
• —The model will still work well with other resolutions
• —Native support for images from 56x56 to 3584x3584

Quick Start

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from PIL import Image
import torch

# Load model and processor
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "qwen25-vl-weather-7b",
    torch_dtype=torch.float16,
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("qwen25-vl-weather-7b")

# Prepare your weather image
image = Image.open("weather_image.jpg")

# Create a prompt
prompt = "Analyze this weather image and describe the meteorological conditions."

# Format the message
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": prompt}
        ]
    }
]

# Process the input
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(
    text=[text],
    images=[image],
    padding=True,
    return_tensors="pt"
).to(model.device)

# Generate response
generated_ids = model.generate(**inputs, max_new_tokens=512)
generated_ids_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)[0]

print(output_text)

Intended Use

This model is designed for:

• —Weather Analysis: Interpreting meteorological imagery and data
• —Educational Applications: Teaching weather concepts
• —Research Support: Assisting in weather data analysis
• —Operational Meteorology: Supporting weather forecasting workflows

Capabilities

The model excels at analyzing:

• —Radar Imagery: Reflectivity, velocity, dual-polarization products
• —Satellite Data: Visible, infrared, water vapor imagery
• —Surface Charts: Weather maps, station plots, frontal analysis
• —Upper Air Data: Soundings, constant pressure charts
• —Model Output: Forecast charts, ensemble data
• —Observational Data: Surface observations, meteograms

Example Prompts

Professional Analysis:

• —"Analyze the radar reflectivity patterns and identify any supercell characteristics."
• —"What does this water vapor imagery reveal about the jet stream position?"
• —"Describe the atmospheric stability based on this sounding."

Educational:

• —"Explain this weather pattern in simple terms."
• —"What safety precautions should people take given these conditions?"

Limitations

• —Specialized for meteorological imagery; may not perform well on general images
• —Best with standard meteorological data formats and visualizations
• —Responses reflect training data biases toward certain weather phenomena

Hardware Requirements

• —Minimum VRAM: 16GB (with 8-bit quantization)
• —Recommended VRAM: 24GB+ (for full precision)
• —Optimal Performance: NVIDIA A100/H100 or RTX 4090/3090

Citation

@misc{weather-qwen25vl-2025,
  title={Weather Analysis Vision-Language Model based on Qwen2.5-VL-7B},
  author={Deepguess},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/qwen25-vl-weather-7b}
}

Acknowledgments

• —Base model: Qwen team for Qwen2.5-VL
• —Training framework: Unsloth for efficient fine-tuning
• —Dataset: Custom curated weather imagery dataset

License

This model follows the license terms of Qwen2.5-VL. See the license file for details.