🏥 Zima - Geriatric Health Assistant

A fine-tuned AI model to provide compassionate, actionable health advice for elderly individuals (70+).

🎯 Project Overview

Model: Qwen 2.5 1.5B fine-tuned with LoRA
Dataset: 10,743 synthetic samples
Perplexity: 1.51 (excellent)

🔗 HuggingFace

• Model: YsK-dev/zima-qwen-geriatric-1.5b
• Dataset: YsK-dev/geriatric-health-advice

📁 Project Structure

zima/
├── data_creation/           # Data generation scripts
│   ├── data_creation_lightning.py  # Main generation script
│   ├── setup_lightning.sh          # Lightning.ai setup
│   └── LIGHTNING_AI_GUIDE.md       # Usage guide
│
├── training/                # Model training
│   ├── prepare_data.py      # Data preprocessing
│   ├── train_unsloth.py     # Unsloth fine-tuning
│   ├── evaluate_model.py    # Model evaluation
│   └── setup_training.sh    # Environment setup
│
├── generated_data/          # Training data
│   └── synthetic_geriatric_data (2).jsonl
│
├── seed_data/               # Initial seed prompts
│
├── trainde_model/Model files (LoRA adapter):
│   ├── adapter_model.safetensors
│   ├── adapter_config.json
│   └── tokenizer files
│
├── MODEL_CARD.md            # HuggingFace model docs
├── DATASET_CARD.md          # HuggingFace dataset docs

🚀 Quick Start

Load from HuggingFace

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "YsK-dev/zima-qwen-geriatric-1.5b",
    max_seq_length=512,
    load_in_4bit=True,
)

Replicate Training

# 1. Generate data (Lightning.ai)
cd data_creation
./setup_lightning.sh
python data_creation_lightning.py

# 2. Train model (Lightning.ai)
cd training
./setup_training.sh
python prepare_data.py
python train_unsloth.py
python evaluate_model.py

📊 Results

Metric	Value
Training Loss	0.32
Validation Loss	0.40
Perplexity	1.51
Samples	10,743

📄 License

Apache 2.0