5ch4um1/lfm2.5-vrsbench-MADOS-maritime-lora-450m

LFM2.5-VL-450M VRSBench + MADOS Maritime Expert

Model Description

This is a fine-tuned version of LiquidAI's LFM2.5-VL-450M vision-language model, specialized for maritime object detection and ship analysis in satellite imagery. The model was trained in two stages:

1. VRSBench Training: Base training on VRSBench dataset
2. MADOS Fine-tuning: Additional training on MADOS (Maritime Detection in Satellite Imagery) dataset

The model can detect and analyze ships and maritime objects in satellite images, providing bounding boxes and classification.

Training Details

Stage 1: VRSBench Pre-training

• Base Model: LFM2.5-VL-450M
• Dataset: VRSBench
• Epochs: 1
• Method: LoRA (r=16, alpha=32)

Stage 2: MADOS Fine-tuning

• Base Model: VRSBench-trained model
• Dataset: MADOS (Maritime Detection in Satellite Imagery)
• Epochs: 6
• Method: LoRA (r=16, alpha=32)
• Hardware: Local training (no Ray/distributed)

Evaluation Results

MADOS Benchmark

Metric	Score
IoU@0.5	~2%
IoU@0.25	~4-6%

Note: The model shows limited performance on maritime object detection benchmarks. This is expected as:

• MADOS is a challenging dataset with small objects
• The model is primarily trained for classification tasks
• Object detection requires different architectural approaches

For production maritime detection, consider using specialized object detection models.

Usage

With llama.cpp

# Download Q4_K_M quantized version (recommended)
wget https://huggingface.co/5ch4um1/lfm2.5-vrsbench-mados-maritime-lora-450m/resolve/main/lfm2.5-vrsbench-mados-maritime-lora-450m-q4_k_m.gguf

# Run inference
./llama-cli -m lfm2.5-vrsbench-mados-maritime-lora-450m-q4_k_m.gguf \
  --image satellite_ship_image.jpg \
  -p "Describe any ships or maritime objects visible in this satellite image."

With Transformers

from transformers import AutoModelForVision2Seq, AutoProcessor
from PIL import Image

model = AutoModelForVision2Seq.from_pretrained(
    "5ch4um1/lfm2.5-vrsbench-mados-maritime-lora-450m",
    torch_dtype="auto",
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("5ch4um1/lfm2.5-vrsbench-mados-maritime-lora-450m")

image = Image.open("satellite_ship_image.jpg")
prompt = "Describe any ships or maritime objects visible in this satellite image."

inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(processor.decode(outputs[0], skip_special_tokens=True))

GGUF Quantizations

Version	Size	Description
F16	679 MB	Full precision (16-bit)
Q8_0	362 MB	8-bit quantization
Q4_K_M	219 MB	4-bit quantization (recommended for most use cases)

Model Sources

• Base Model: LiquidAI/LFM2.5-VL-450M
• MADOS Dataset: Maritime Detection in Satellite Imagery

Limitations

• Model performance on maritime object detection is limited (IoU@0.5: ~2%)
• Designed more for classification and description rather than precise object detection
• Works best with visible ships in satellite imagery
• May not generalize to all maritime scenarios without additional training

Training Environment

• Framework: Transformers + PEFT (LoRA)
• Hardware: Local GPU (CUDA)
• Training Scripts: Available in the cookbook repository