bisonnetworking/medgemma-health-chat-merged

Overview

bisonnetworking/medgemma-health-chat-merged is a 4.3 billion parameter model, fine-tuned from Google's MedGemma-1.5-4B-IT. This 16-bit merged model is specifically designed for multi-persona health chat conversations, where it responds as one of six medical personas (e.g., Primary Care, Clinical Nutritionist, Chronic Health Specialist). The LoRA adapter has been merged into the base model, making it a standalone model.

Key Capabilities

• Multi-Persona Responses: Provides clinically grounded guidance from six distinct medical specialties based on topic relevance.
• Mobile-Optimized Output: Generates concise, direct responses without tables, suitable for mobile display.
• Health Context Integration: Utilizes simulated Apple Health data (vitals, labs, medications, conditions) injected into user messages for personalized responses.
• Conversational Clinical Guidance: Acts as a health chat assistant, offering information in a conversational format.

Training Details

The model was fine-tuned using LoRA on 49,500 conversations from the bisonnetworking/medgemma-health-chat-sft dataset. Training involved 2 epochs with a batch size of 8 and a max sequence length of 2048, achieving a final training loss of 0.7981. The base model was loaded in 4-bit (NF4) quantization during training.

Good for

• Patient-facing health chat applications requiring specialized medical personas.
• Providing conversational clinical guidance optimized for mobile interfaces.
• Applications needing to integrate simulated health data into responses.

Limitations

It is crucial to note that this model is not intended for diagnosing conditions, replacing professional medical care, or emergency triage. It is fine-tuned on synthetic conversations, and its quality depends on the base model's medical knowledge. Additional safety training (RLHF/red-teaming) is recommended for production use.