nabilfaieaz/tinyllama-med-full

🩺 TinyLlama Medical Assistant (Merged LoRA)

Author: Nabil Faieaz
Base model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
Fine-tuning method: LoRA (Low-Rank Adaptation) using PEFT → merged into base weights
Intended use: Concise, factual, general medical information

---

📌 Overview

This model is a fine-tuned version of TinyLlama 1.1B-Chat adapted for medical question answering.
It has been trained to give brief and accurate answers to medical-related queries, following a consistent Q/A style.

Key features:

• ✅ LoRA fine-tuning for efficient adaptation on limited compute (T4 GPU)
• ✅ Merged LoRA + base into a single standalone model (no separate adapter needed)
• ✅ Optimized for short, factual answers — avoids overly verbose outputs
• ✅ Context-aware: warns users to seek professional medical help for urgent/personal issues

---

⚠️ Disclaimer

This model is for educational and informational purposes only.
It is not a substitute for professional medical advice, diagnosis, or treatment.
Always consult a qualified healthcare provider for medical concerns.

---

🚀 Quick Start

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "nabilfaieaz/tinyllama-med-full"

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_id)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto"
)

# Example prompt
system_prompt = (
    "You are a helpful, concise medical assistant. Provide general information only, "
    "not a diagnosis. If urgent or personal issues are mentioned, advise seeing a clinician."
)

question = "What is hypertension?"
prompt = f"{system_prompt}\n\nQuestion: {question}\nAnswer:"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    do_sample=False,
    temperature=0.0,
    top_p=1.0,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.pad_token_id
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

🧠 Training Details
Base model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
Fine-tuning method: LoRA (via peft)
Target modules: q_proj, k_proj, v_proj, o_proj
LoRA config:
    * r = 16
    * alpha = 16
    * dropout = 0.0
Max sequence length: 512 tokens
Batch size: 2 per device (gradient accumulation for effective batch)
Learning rate: 2e-4
Precision: fp16
Evaluation: periodic eval every 200 steps
Checkpoints: saved every 500 steps, final merge from checkpoint-17000

📊 Intended Use
Intended:
    * Educational explanations of medical terms and concepts
    * Study aid for medical students and healthcare professionals
    * Healthcare-related chatbot demos

Not intended:
    * Real-time clinical decision making
    * Emergency medical guidance
    * Handling sensitive personal medical data (PHI)
⚙️ Technical Notes
    * The model is merged — you don’t need to separately load LoRA adapters.
    * Works with Hugging Face transformers ≥ 4.38.
    * Can be quantized to 4-bit (e.g., QLoRA) for local inference.