lfm2.5-vrsbench-firewatch-sft-lora-450m

LFM2.5-VL-450M Fire Expert (Sentinel-2)

Fine-tuned vision-language model for wildfire detection and burn scar analysis in Sentinel-2 satellite imagery.

Model Description

This model is a fine-tuned version of LFM2.5-VL-450M on the NuTonic/firewatch-sft-v1 dataset combined with additional fire/no-fire satellite imagery. The model specializes in detecting and analyzing wildfires, burn scars, and fire-related features in Sentinel-2 satellite imagery.

Base Model

• Base: liquidai/LFM2.5-VL-450M (450M parameter vision-language model)
• Pre-trained on: VRSBench dataset for general satellite image understanding
• Fine-tuned on: Fire watch dataset with Sentinel-2 imagery

Training Data

The model was trained on 520 samples combining:

• firewatch-sft-v1: 200 training samples with Sentinel-2 pre/post fire image pairs
• Local fire data: 60 samples (35 fire + 25 no-fire images)
• Augmented data: 260 samples (flipped versions for data augmentation)

Dataset Features

• firewatch-sft-v1: RGB composites (True Color: B4=Red, B3=Green, B2=Blue), pre/post fire pairs with bounding boxes
• Local fire data: Fire-optimized composites (B12=SWIR2, B8=NIR, B4=Red) for enhanced fire detection
• Binary fire/no-fire classification
• Burn scar extent analysis

Usage

With Transformers (PyTorch)

from transformers import AutoModelForVision2Seq, AutoProcessor
from PIL import Image

# Load model and processor
model = AutoModelForVision2Seq.from_pretrained("5ch4um1/lfm2.5-vrsbench-fire-sentinel2-lora-450m")
processor = AutoProcessor.from_pretrained("5ch4um1/lfm2.5-vrsbench-fire-sentinel2-lora-450m")

# Prepare input
image = Image.open("sentinel2_fire_image.jpg")
messages = [
    {"role": "user", "content": [
        {"type": "image", "image": image},
        {"type": "text", "text": "Is there any wildfire or burn scar in this Sentinel-2 image? Provide bounding boxes if fire is detected."}
    ]}
]

# Process and generate
inputs = processor(messages, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
response = processor.decode(outputs[0], skip_special_tokens=True)
print(response)

With GGUF (llama.cpp)

# Download GGUF model
wget https://huggingface.co/5ch4um1/lfm2.5-vrsbench-fire-sentinel2-lora-450m/resolve/main/fire-expert-f16.gguf

# Run with llama.cpp
./llama-cli -m fire-expert-f16.gguf -p "Analyze this satellite image for fire damage" --image sentinel2_image.jpg

With ONNX Runtime

ONNX exports are available in the onnx/ directory with multiple precision options:

• fp16 - Half precision (faster inference)
• q4 - 4-bit quantized (smallest size)
• q8 - 8-bit quantized (balanced)

import onnxruntime as ort
import numpy as np
from PIL import Image

# Load ONNX model
session = ort.InferenceSession("onnx/vision_encoder_fp16.onnx")
# ... process image and run inference

Model Performance

Trained for 7 epochs with:

• LoRA rank: 16
• LoRA alpha: 32
• Learning rate: 2e-5
• Batch size: optimized for multi-GPU training

Limitations

• Designed specifically for Sentinel-2 satellite imagery
• May not generalize to other satellite platforms
• Trained on mixed composites: RGB (True Color) and fire-optimized (SWIR2+NIR+Red)
• Requires clear imagery for accurate fire detection

Training Configuration

base_model: liquidai/LFM2.5-VL-450M
dataset:
  - NuTonic/firewatch-sft-v1
  - local_fire_data
  - augmented_samples
epochs: 7
lora_rank: 16
lora_alpha: 32
learning_rate: 2e-5

Repository Structure

├── config.json              # Model configuration
├── model.safetensors       # Merged model weights
├── processor_config.json    # Image processor config
├── tokenizer.json          # Tokenizer
├── fire-expert-f16.gguf    # GGUF format (FP16)
└── onnx/                   # ONNX exports
    ├── vision_encoder_fp16.onnx
    ├── decoder_model_merged_fp16.onnx
    └── embed_tokens_fp16.onnx

Citation

If you use this model, please cite:

@misc{lfm25vl-fire-expert,
  title={LFM2.5-VL-450M Fire Expert for Sentinel-2},
  author={AI-in-Space},
  year={2026},
  howpublished={\url{https://huggingface.co/5ch4um1/lfm2.5-vrsbench-fire-sentinel2-lora-450m}}
}

License

This model follows the license of the base model (LFM2.5-VL-450M). Check the base model repository for license details.