HamzaKhan-03/Hamza_the_doctor.ai

🩺 Hamzathedoctor.ai — Medical Chatbot (Fine-tuned Gemma 3 270M)

A lightweight medical conversational model fine-tuned on healthcare QA data using Unsloth’s 4-bit optimization, enabling efficient deployment on consumer GPUs while maintaining strong accuracy for medical dialogue.

🧠 Model Overview

Model Description

• —Developed by: Muhammad Hamza Khan
• —Base model: `unsloth/gemma-3-270m-it-unsloth-bnb-4bit`
• —Task: Medical Question Answering / Chatbot
• —Fine-tuning method: LoRA using PEFT + TRL (SFTTrainer)
• —Language(s): English
• —Frameworks used: Python, Jupyter Notebook (Kaggle), Unsloth, Transformers, PEFT
• —License: Apache 2.0

This model was trained for health-related conversational reasoning — ideal for educational or research use-cases involving patient–doctor-style dialogues.

🧩 Model Sources

• —Repository: HamzaKhan-03/Hamza_the_doctor.ai
• —Dataset: lavita/ChatDoctor-HealthCareMagic-100k

🚀 Usage

from unsloth import FastModel
import torch, os

HF_TOKEN = os.environ["HF_TOKEN"]
HF_REPO_ID = "HamzaKhan-03/Hamza_the_doctor.ai"

# Load model in 4-bit precision
model, tokenizer = FastModel.from_pretrained(
    HF_REPO_ID,
    load_in_4bit=True,
    token=HF_TOKEN,
)
model.to("cuda")

messages = [{"role": "user", "content": "Can you tell me about cancer?"}]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
).removeprefix('<bos>')

inputs = tokenizer(text, return_tensors="pt").to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.95,
    top_k=50,
)

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

🧪 Training Details

Dataset

• —Name: lavita/ChatDoctor-HealthCareMagic-100k
• —Type: Doctor–Patient QA pairs for medical reasoning tasks
• —Preprocessing: Filtered for concise Q&A, cleaned for multi-turn context

Hyperparameters

Hardware

• —Platform: Kaggle TPU / A100 (16 GB)
• —Fine-tuning time: ~30 minutes
• —Framework versions:

• —unsloth ≥ 0.7.1
• —peft 0.17.1
• —transformers ≥ 4.44
• —trl ≥ 0.9.6

📊 Evaluation

⚠️ Note: These metrics are indicative only; not validated for clinical reliability.

🌍 Environmental Impact (Estimated)

🧱 Model Architecture

• —Base: Gemma 3 270 M (Instruction-Tuned)
• —Adapter: LoRA (rank 8)
• —Quantization: 4-bit NF4 for reduced memory footprint
• —Objective: Supervised Fine-Tuning (SFT) with instruction-style prompts

🧾 Citation

If you use this model, please cite:

BibTeX:

@misc{hamza2025doctorai,
  title={Hamza_the_doctor.ai — Medical Chatbot (Fine-tuned Gemma 3 270M)},
  author={Muhammad Hamza Khan},
  year={2025},
  howpublished={\url{https://huggingface.co/HamzaKhan-03/Hamza_the_doctor.ai}},
}

📬 Contact

For technical questions or collaborations: Author: Muhammad Hamza Khan Hugging Face: @HamzaKhan-03