A 16-bit merged model fine-tuned from google/medgemma-1.5-4b-it for multi-persona health chat conversations. The LoRA adapter has been merged into the base model weights, producing a standalone model that requires no separate adapter loading.

The model responds as one of six medical personas (primary care, internal medicine, clinical nutritionist, exercise specialist, integrated physician, chronic health specialist) with direct, clinically grounded guidance optimized for mobile display.

Intended Use

Health chat assistant for patient-facing applications
Provides clinical guidance in a conversational format
Designed for mobile-first display (no tables, concise prose)
Each conversation assigns a persona based on topic relevance

Not intended for: diagnosing conditions, replacing professional medical care, or emergency triage. Always recommend professional care for emergencies.

Personas

Persona	Specialty
Primary Care	General practice, common conditions, preventive care
Internal Medicine	Complex adult medicine, multi-system disorders, chronic disease
Clinical Nutritionist	Dietary interventions, nutritional therapy, meal planning
Exercise Specialist	Therapeutic exercise, sports performance, rehabilitation
Best Doctor	Cross-specialty integration, OLDCARTS methodology, comprehensive care
Chronic Health	Chronic illness management, diagnostic mysteries, patient coaching

Persona prompts are sourced from BisonHealth-AI/Personas-Prompts.

Training Data

Dataset: bisonnetworking/medgemma-health-chat-sft
Training samples: 49,500 conversations
Eval samples: 500 conversations
Format: ShareGPT/conversational JSONL with system (persona prompt) + user + assistant turns
Health context: Simulated Apple Health data (vitals, labs, medications, conditions) injected into user messages

Training Procedure

LoRA Fine-Tuning

Rank (r): 32
Alpha: 64
Dropout: 0
Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Trainable parameters: 65,576,960 (1.50% of 4,365,656,432 total)
Base model loaded in: 4-bit (NF4) quantization

Training Hyperparameters

Epochs: 2
Batch size: 8 (per-device)
Gradient accumulation: 1
Learning rate: 2e-4 (cosine scheduler, warmup ratio 0.03)
Optimizer: adamw_8bit
Weight decay: 0.01
Max sequence length: 2048
Precision: bf16
Packing: True (sequence concatenation for GPU efficiency)

Training Results

Total steps: 12,376
Final training loss: 0.7981
Best eval loss: 0.8779 (step 12,000)
Loss trajectory: 1.40 (step 75) → 1.01 (step 700) → 0.94 (step 1100) → 0.80 (step 12,375)

Merging

The LoRA adapter was merged into the base MedGemma 1.5 4B model in float16 precision using PEFT's merge_and_unload(). The merged model is distributed as two safetensors shards (~8.6GB total).

Compute Infrastructure

Hardware: NVIDIA H100 80GB (Modal serverless)
Training time: ~3 hours
Software: Unsloth 2025.7.8, Transformers 4.54.0, PEFT 0.16.0, TRL 0.19.1
Platform: Modal

How to Use

With Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "bisonnetworking/medgemma-health-chat-merged"

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [
    {
        "role": "system",
        "content": "You are a board-certified Primary Care Physician. Provide clinical guidance with professionalism and clarity. Use ONLY health data explicitly provided in user context. No tables. No AI disclaimers. Answer the specific question asked - no more, no less.",
    },
    {
        "role": "user",
        "content": "PATIENT CONTEXT: Age 52, Male. BP: 145/92 mmHg. Medications: Lisinopril 20mg daily. Conditions: Hypertension.\n\nMy blood pressure has been reading higher than usual the past week. Should I adjust my medication?",
    },
]

inputs = tokenizer.apply_chat_template(
    messages, return_tensors="pt", add_generation_prompt=True
).to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
)
response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)
print(response)

With vLLM

python

from vllm import LLM, SamplingParams

llm = LLM(model="bisonnetworking/medgemma-health-chat-merged", dtype="float16")
sampling = SamplingParams(temperature=0.7, top_p=0.9, max_tokens=512)

messages = [
    {"role": "system", "content": "You are a board-certified Primary Care Physician..."},
    {"role": "user", "content": "I've had a sore throat for 3 days. What should I do?"},
]
prompt = llm.get_tokenizer().apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = llm.generate([prompt], sampling)
print(outputs[0].outputs[0].text)

For MLX (Apple Silicon)

bash

# Convert to MLX format
pip install mlx-lm
mlx_lm.convert --hf-path bisonnetworking/medgemma-health-chat-merged --mlx-path ./medgemma-health-chat-mlx
mlx_lm.generate --model ./medgemma-health-chat-mlx --prompt "I've had a sore throat for 3 days. What should I do?"

Evaluation

300 test cases (50 per persona) across 6 categories:

Category	Cases/Persona	What It Tests
Medical accuracy	14	Clinically correct information for persona's specialty
Persona adherence	9	Response style matches persona (direct, no AI disclaimers, concise)
Health context usage	10	References actual values from provided health data
Data integrity	6	Does not hallucinate data not in context
Safety	6	Recognizes emergencies, recommends 911/988/Poison Control
Formatting	5	No markdown tables, no AI disclaimers, mobile-friendly

Test cases are in the medgemma-health-chat repository under test_data/.

Limitations

Fine-tuned on synthetic conversations — quality depends on the base model's medical knowledge
Persona prompts enforce style but cannot guarantee clinical accuracy
No tables in output (mobile constraint) — may reduce clarity for complex comparisons
Not a substitute for professional medical advice
MedGemma base model is gated — users need approved access to the original model
No safety training beyond what's in the base model — additional RLHF/red-teaming recommended for production

Ethical Considerations

This model provides health information but is NOT a medical device
Responses should not be used as sole basis for medical decisions
Emergency situations should always be directed to 911 or local emergency services
The model may reflect biases present in the training data
Users should be clearly informed they are chatting with an AI, not a human physician

LoRA Adapter

The original LoRA adapter (pre-merge) is available at bisonnetworking/medgemma-health-chat-lora for those who prefer to load it separately or further fine-tune.

Model Card Contact

bisonnetworking