This repository contains an ultra-portable LoRA Adapter (15 MB) for Gemma-3-270m-it, fine-tuned natively on Apple Silicon for clinical dialogue-to-SOAP-note summarization.
🚀 The Efficiency Advantage
With a tiny 15 MB footprint, this Adapter-only release is optimized for high-speed inference on edge devices, mobile hardware, and local workstations. It acts as a lightweight "medical brain patch" for the foundational Gemma-3-270m-it model.
🎯 Performance Gain
ROUGE-L Improvement: +64.4% gain over the base 270M model.
Optimal Use Case: Extreme speed and low-memory clinical documentation for embedded and mobile applications.
🌍 Cross-Platform Support
While trained with MLX, these adapters are architecture-agnostic:
Linux/Windows: Use Hugging Face transformers + PEFT.
Edge/CPU: Convert to GGUF (llama.cpp) for use on mobile and local CPUs.
📂 Training & Evaluation
Training Dataset:MedSynth (Synthetic Medical Dialogue-Note Pairs).
Evaluation Dataset:MTS-Dialog (Real clinical dialogues from the Task 1 benchmark).
⚖️ Licensing
MTS-Dialog:CC BY 4.0 (Commercial use allowed with attribution).
MedSynth: No explicit license mentioned; intended for research purposes.
