mshz88/FADA-Mobile-GGUF

FADA-SKD 0.8B (GGUF)

Fetal Anatomy Delineation and Analysis - Selective Knowledge Distillation

On-device vision-language model for fetal ultrasound image analysis, optimized for mobile deployment via llama.cpp.

Model Overview

Property	Value
Base Model	Qwen3.5-VL 0.8B (via Unsloth)
Fine-tuning	LoRA (rank 16, alpha 32) via Unsloth SKD
Training Strategy	Offline Selective Knowledge Distillation
Architecture	Qwen3.5 (text) + Qwen3-VL Merger (vision)
Context Length	4096 tokens
Text Model	Q4_K_M quantized (~517 MB)
Vision Encoder	FP16 (~196 MB)
Format	GGUF v3
Total Size	~713 MB

Teacher Models (SKD)

Teacher	Weight	Role
FetalCLIP	0.40	Fetal anatomy representation alignment
UltraSAM	0.25	Ultrasound segmentation knowledge
USF-MAE	0.20	Self-supervised ultrasound features
UltraFedFM	0.15	Federated foundation model knowledge

Training: 21,144 steps, 3 epochs, fusion MSE loss, 19,000+ images across 32+ anatomical classes.

Files

File	Description	Size
gguf/fada-skd-0.8b-Q4_K_M.gguf	Text model (Q4_K_M quantized)	517 MB
gguf/fada-skd-0.8b-mmproj-f16.gguf	Vision encoder (FP16 mmproj)	196 MB
tokenizer.json	Tokenizer vocabulary	-
tokenizer_config.json	Tokenizer configuration	-
chat_template.jinja	Chat template for inference	-

Usage with llama.cpp

Text + Vision Inference

# Using llama-mtmd-cli from llama.cpp
./llama-mtmd-cli \
  -m fada-skd-0.8b-Q4_K_M.gguf \
  --mmproj fada-skd-0.8b-mmproj-f16.gguf \
  -p "Analyze this fetal ultrasound image" \
  --image ultrasound.jpg

Mobile Deployment

This model is designed for on-device mobile deployment using the FADA Android app with llama.cpp's multimodal (mtmd) library for native inference without cloud connectivity.

Capabilities

FADA-SKD performs 5-phase fetal ultrasound analysis:

1. Interpretation: Structured clinical report (imaging plane, anatomical structures, gestational age, image quality, normality assessment)
2. Classification: Anatomical view identification (BPD plane, four-chamber, Doppler, etc.)
3. Mapping: Intelligent routing to relevant detection/segmentation classes
4. Detection: Bounding box localization of anatomical structures (normalized 0-1000 coordinates)
5. Segmentation: Polygon mask delineation of anatomical regions

Supported Anatomy (32+ classes)

• Brain: BPD, CSP, lateral ventricle, brain parenchyma
• Cardiac: Heart chambers, thorax
• First Trimester: CRL, NT, nasal bone/skin/tip
• Doppler: Arteries, veins, liver, stomach
• Pelvimetry: Fetal head, pubic symphysis
• Body/Pose: Abdomen, limbs, head
• Keypoints: CRL endpoints, NT caliper points, scale bar

Citation

If you use this model in your research, please cite:

@article{fada2026,
  title={FADA: Knowledge-Distilled Vision-Language Models for Accessible Fetal
         Ultrasound Interpretation in Low-Resource Obstetric Settings},
  author={Alzubaidi, Mahmood and Agus, Marco},
  journal={Arxiv},
  year={2026},
  note={Submitted to the "Digital Health in Low-Resource Settings" Collection}
} year={2026}
}

License

Apache 2.0

Disclaimer

This is a research prototype. Not for clinical diagnostic use. All outputs should be reviewed by qualified medical professionals.