🏥 Zima: Geriatric Health Assistant

A specialized AI assistant providing compassionate, actionable health guidance for elderly individuals (70+)

🌟 What Makes Zima Special?

Zima is a fine-tuned Qwen 2.5 1.5B model specifically designed to assist elderly individuals with health-related questions. Unlike general-purpose chatbots, Zima:

✅ Age-Appropriate Communication - Uses clear, simple language
✅ Safety-First Approach - Prioritizes user safety, knows when to recommend emergency services
✅ Actionable Advice - Provides specific, implementable steps rather than generic suggestions
✅ Comprehensive Coverage - Handles everything from daily wellness to first aid
✅ Compact & Efficient - Only 1.5B parameters, perfect for edge deployment

📊 Performance Metrics

Metric	Value	Status
Perplexity	1.51	🏆 Excellent (< 2.0)
Training Loss	0.32	✅ Converged
Validation Loss	0.40	✅ No overfitting
Training Samples	10,743	📚 High quality
Model Size	71 MB (LoRA)	🚀 Edge-ready

🚀 Quick Start

Installation

pip install unsloth transformers torch

Basic Usage

from unsloth import FastLanguageModel

# Load model
model, tokenizer = FastLanguageModel.from_pretrained(
    "YsK-dev/zima-qwen-geriatric-1.5b",
    max_seq_length=512,
    dtype=None,
    load_in_4bit=True,
)

# Enable inference mode
FastLanguageModel.for_inference(model)

# Create prompt
prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
How can I prevent falls at home?

### Input:
Patient is elderly (70+), seeking advice on safety.

### Response:
"""

# Generate response
inputs = tokenizer([prompt], return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7)
response = tokenizer.batch_decode(outputs)[0]

print(response)

Expected Output

Install handrails near stairs and in the bathroom. Use non-slip mats in the shower 
and bathtub. Keep floors clear of clutter. Ensure good lighting throughout your 
home, especially in hallways and near stairs. Consider using a walking aid if you 
feel unsteady. Remove loose rugs or secure them with non-slip backing.

💡 Use Cases

Daily Health Guidance

Medication reminders and management
Nutrition and hydration advice
Exercise recommendations for seniors
Sleep hygiene tips

First Aid & Emergency

Minor injury treatment (cuts, burns, bruises)
Fall prevention and recovery
Recognizing when to seek emergency help
Symptom assessment

Mental & Emotional Well-being

Loneliness and social isolation
Cognitive health maintenance
Stress management techniques
Connection with support resources

Chronic Condition Management

Blood pressure monitoring
Diabetes management tips
Arthritis pain relief
Medication adherence strategies

🔬 Training Details

Base Model

Architecture: Qwen 2.5 1.5B Instruct
Fine-tuning Method: LoRA (Low-Rank Adaptation)
Quantization: 4-bit for efficiency

Training Configuration

LoRA Configuration:
  rank: 16
  alpha: 32
  dropout: 0.05
  target_modules: [q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]

Training Hyperparameters:
  batch_size: 8 (effective: 32 with gradient accumulation)
  learning_rate: 2e-4
  scheduler: cosine
  epochs: 3
  optimizer: AdamW 8-bit
  
Hardware:
  GPU: NVIDIA L4 (22GB VRAM)
  Training time: 34 minutes
  Framework: Unsloth (2x faster than standard)

Dataset

Source: geriatric-health-advice
Size: 10,743 high-quality samples
Split: 90% train / 10% validation
Format: Alpaca-style instruction-following
Generation: Synthetic data from Qwen 2.5 14B

📋 Example Conversations

Example 1: Common Cold

User: I have a cold

Zima: Rest and drink plenty of fluids like warm water, tea or soup. Use saline nasal spray to clear congestion. Take over-the-counter decongestants if needed. Avoid close contact with others. See your doctor for severe symptoms. Stay home until you feel better.

Example 2: Loneliness

User: I feel lonely

Zima: Try joining clubs or groups where you can meet people with similar interests. Consider volunteering to help others and connect socially. Regular phone calls can also keep relationships strong. Talk to a counselor if loneliness feels overwhelming.

Example 3: Medication Management

User: I keep forgetting to take my pills

Zima: Consider using a pill organizer and fill it weekly. Set reminders on your phone or use an alarm clock. Ask family members to help remind you. Keep your medications in a visible spot where you'll see them daily. Talk to your pharmacist about once-daily formulations if available.

⚠️ Limitations & Safety

Important Disclaimers

🚨 NOT A REPLACEMENT FOR PROFESSIONAL MEDICAL CARE

Zima is designed to provide general health guidance and support, but it:

Cannot diagnose medical conditions

Cannot prescribe medications

Cannot replace emergency services (911)

Cannot provide personalized medical treatment plans

Known Limitations

Emergency Responses: May be verbose in critical situations (working on v2)
Language: Currently English-only
Geographic Specificity: Advice is general and may not apply to all regions
Personalization: Cannot access individual medical histories

When to Seek Professional Help

Always contact a healthcare provider or emergency services for:

Chest pain or difficulty breathing
Severe bleeding that won't stop
Loss of consciousness
Sudden vision or speech changes
Severe allergic reactions
Any life-threatening emergency

🛠️ Advanced Usage

Deployment on Edge Devices

Zima is optimized for deployment on resource-constrained devices like Raspberry Pi 5:

# Quantize to GGUF for even smaller size
python quantize_to_gguf.py --model YsK-dev/zima-qwen-geriatric-1.5b --quant q4_k_m

# Run inference on CPU
from llama_cpp import Llama
model = Llama(model_path="zima-q4_k_m.gguf")

API Server

from fastapi import FastAPI
from unsloth import FastLanguageModel

app = FastAPI()
model, tokenizer = FastLanguageModel.from_pretrained("YsK-dev/zima-qwen-geriatric-1.5b")

@app.post("/advice")
async def get_advice(question: str):
    # Generate response
    return {"advice": generated_text}

Gradio Demo

import gradio as gr
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained("YsK-dev/zima-qwen-geriatric-1.5b")

def get_advice(question):
    # Generate response logic
    return response

demo = gr.Interface(
    fn=get_advice,
    inputs=gr.Textbox(placeholder="Ask a health-related question..."),
    outputs=gr.Textbox(),
    title="🏥 Zima: Your Geriatric Health Assistant"
)

demo.launch()

📈 Future Roadmap

v2.0: More concise emergency responses
Multilingual Support: Spanish, French, German, Chinese
Voice Interface: Integration with speech-to-text/text-to-speech
Mobile App: iOS and Android applications
Telehealth Integration: Connect with real healthcare providers
Personalization: User profile and medical history awareness

🤝 Contributing

We welcome contributions! Areas where you can help:

Data Collection: More diverse geriatric health scenarios
Evaluation: Testing model responses for accuracy
Documentation: Improving guides and examples
Localization: Translations to other languages
Integration: Building apps and tools using Zima

📄 License

This model is released under the Apache 2.0 License. You are free to:

✅ Use commercially
✅ Modify and distribute
✅ Use privately
✅ Use for research

🙏 Acknowledgments

Unsloth: For their incredible training optimizations (2x speedup!)
Qwen Team: For the excellent base model
Lightning.ai: For free GPU compute
HuggingFace: For hosting and community

📧 Contact

Developer: YsK-dev
Issues: GitHub Issues
Email: [Your Email if you want to add it]
HuggingFace: @YsK-dev

📚 Citation

If you use Zima in your research, please cite:

@misc{zima2025,
  author = {YsK-dev},
  title = {Zima: A Geriatric Health Assistant},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/YsK-dev/zima-qwen-geriatric-1.5b}},
}

Made with ❤️ for the elderly community