hllzmz/medgemma-mentalist

MedGemma-27B Mentalist

MedGemma Mentalist is a specialized Mental Health Assistant model fine-tuned on top of Google's medgemma-27b. It has been trained using with high-quality synthetic client-therapist dialogues.

This model is designed to understand users' emotional states, interpret described experiences through the lens of mental health and clinical standards, and provide supportive guidance using a warm, empathetic tone.

Critical Disclaimer

This model is NOT a licensed medical professional.

• It cannot provide definitive medical diagnoses.
• It cannot prescribe medication.
• It cannot replace emergency services in crisis situations (e.g., suicide, self-harm, harm to others).
• It is intended solely for educational, research, and preliminary informational purposes.

Model Capabilities

• Empathetic Dialogue: Listens to the user without judgement and validates their feelings (Active Listening).
• Symptom Analysis: correlates user-described experiences with clinical terminology and criteria.
• Safety First: Prioritizes safety planning and refers users to professional help when risk signals are detected.

How to Use

This model can be loaded directly using 'transformers'. For optimal performance and role adherence, please strictly follow the parameters below.

1. Install Dependencies

pip install torch transformers

2. Python Inference Script

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Model from HF
model_id = "hllzmz/medgemma-mentalist"

# Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Load Model
# Use load_in_4bit=True for VRAM efficiency (requires bitsandbytes)
# Or use torch_dtype=torch.bfloat16 for full precision if you have enough VRAM
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    dtype=torch.bfloat16, 
    load_in_4bit=True 
)

SYSTEM_PROMPT = """
    You are MedGemma Mentalist, an advanced AI mental health assistant designed to provide empathetic support, scientifically grounded psychoeducation, and guidance.
    Your goal is to be a bridge to professional help and a source of reliable mental health information.
    You should NEVER diagnose or stigmatize the user directly. 
    """

def generate_response(user_input):
    # Construct the conversation history
    messages = [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": user_input}
    ]
    
    # Apply Chat Template (Gemma Format)
    input_ids = tokenizer.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_tensors="pt"
    ).to(model.device)

    # Generate Response
    outputs = model.generate(
        input_ids=input_ids,
        max_new_tokens=1024,
        temperature=0.5, 
        top_p=0.9,
        repetition_penalty=1.1,
    )
    
    # Decode
    response = tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)
    return response

# Test Run
print(generate_response("I feel anxious and sweaty when I am in crowded places."))

Recommended Parameters

To prevent the model from hallucinating or accidentally roleplaying as the client (User), the following generation settings are highly recommended:

Temperature: 0.3 - 0.5 (Lower values ensure the model remains objective and grounded in mental health knowledge).

Repetition Penalty: 1.1 (Prevents the model from getting stuck in loops).