jmdevita/medical-wayfinder-gemma-4-e2b

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:5.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:May 18, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

The jmdevita/medical-wayfinder-gemma-4-e2b is a 5.1 billion parameter Gemma 4 E2B model fine-tuned by jmdevita for on-device healthcare facility wayfinding. This model specializes in generating step-by-step directions, including landmarks and accessibility information, in both English and Spanish. It is optimized for local execution on mobile devices, providing structured JSON responses for navigation queries without sending Protected Health Information (PHI) off-device.

Loading preview...

Medical Wayfinder: On-Device Healthcare Navigation

The jmdevita/medical-wayfinder-gemma-4-e2b is a specialized 5.1 billion parameter Gemma 4 E2B model, fine-tuned by jmdevita for on-device healthcare facility wayfinding. Developed as a submission for the Gemma 4 Good Hackathon, its core function is to provide step-by-step directions within medical facilities in English and Spanish.

Key Capabilities

  • Structured Wayfinding: Given a system prompt, facility context, and user query (e.g., "cardiology," "MRI"), the model emits a strict JSON response for multi-step walking guides.
  • Multilingual Support: Provides directions and handles queries in both English and Spanish, with Spanish performance notably outscoring English in evaluations.
  • On-Device Operation: Designed for local execution on mobile devices (e.g., iOS via llama.cpp + Metal GPU) using GGUF Q4_K_M quantization, ensuring no Protected Health Information (PHI) leaves the device.
  • Intent Classification & Hedging: Handles user intent, multilingual phrasing, and accessibility-aware step formatting, while a separate orchestrator manages path-finding.
  • Performance Improvement: Achieved significant improvements over the base model in evaluation metrics, including a +0.36 increase in mean rubric score and a +10 percentage point increase in strict pass rate, particularly in "Scope Handling."

Intended Use Cases

  • Hospital/Clinic Navigation: Ideal for generating internal wayfinding instructions for patients and visitors within healthcare facilities.
  • On-Device Applications: Suitable for mobile applications requiring localized, privacy-preserving navigation assistance.

Limitations

  • Not a Medical Advisor: Explicitly out of scope for medical advice, diagnosis, or clinical decisions.
  • Synthetic Training Data: Primarily trained on 100% synthetic data, though anchored by curated real-world directional phrases.
  • Evaluation Bias: The 100-case evaluation suite was authored alongside the data contract, potentially introducing bias.

Finetunes
1 models