luisml77/gemma-aerial-12b

Gemma3-Aerial-12B

QLoRA-finetuned Gemma3-12B for generating natural referring expressions in aerial imagery. Distilled from 500 OpenAI o3 samples (~238× cheaper than direct o3 usage).

Links

• 🌐 Project - Project page
• 📝 Paper - arXiv preprint
• 📊 Training Data - O3 500-sample dataset
• 📦 Full Dataset - Aerial-D
• 💻 Code - Complete pipeline

Model Details

• Base: google/gemma-3-12b-it
• Method: QLoRA (rank=64, alpha=16)
• Training Data: 500 OpenAI o3 samples
• Purpose: LLM enhancement for Aerial-D dataset generation

Usage

Used in the dataset generation pipeline to enhance rule-based expressions:

# Start vLLM server
vllm serve luisml77/gemma-aerial-12b --port 8000

# Run enhancement (in another terminal)
cd datagen
python pipeline/7_vllm_enhance.py

Full pipeline at GitHub.

Training

python gemma3_lora_finetune.py \
  --enhanced_data_dir enhanced_annotations_o3_dual \
  --output_dir ./gemma-aerial-12b \
  --lora_r 64 --lora_alpha 16

Citation

@article{marnoto2025aeriald,
  title={Generalized Referring Expression Segmentation on Aerial Photos},
  author={Marnoto, Luís Pedro Soares},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (J-STARS)},
  year={2025},
  note={Submitted}
}