DinoCU/medical-qa-llama3.1-8b

Medical QA LLM — QLoRA Fine-Tuned Llama 3.1 8B

A medical question-answering model fine-tuned from Meta Llama 3.1 8B-Instruct using QLoRA (4-bit quantization + LoRA adapters) on 107K medical QA examples.

Training Data

Dataset	Examples	Description
ChatDoctor	96,485	Patient-doctor conversations
MedQA-USMLE	10,178	USMLE-style medical exam questions
PubMedQA	1,000	Biomedical research questions
Total	107,663	Train: 101,421 / Val: 5,338

Training Results

• Initial Loss: 3.37
• Final Loss: 1.18
• Best Loss: 1.13 (step 7,190)
• Loss Reduction: 65%

Training Loss Curve

Loss Distribution by Epoch

Evaluation — PubMedQA Benchmark

Model	Accuracy (300 samples)
Base Llama 3.1 8B-Instruct	75.3%
Fine-Tuned (this model)	71.7%

Note: The fine-tuned model was primarily trained on conversational medical data (ChatDoctor ~90% of training set), optimizing for detailed, doctor-style responses rather than terse yes/no/maybe classification. The PubMedQA benchmark measures classification accuracy, which favors the base model's instruction-following format. The fine-tuned model excels at generating comprehensive medical explanations and patient-friendly responses.

Usage

With PEFT (Recommended — uses less memory)

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

# Load in 4-bit for memory efficiency
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    quantization_config=bnb_config,
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, "DinoCU/medical-qa-llama3.1-8b")
tokenizer = AutoTokenizer.from_pretrained("DinoCU/medical-qa-llama3.1-8b")

# Generate
messages = [
    {"role": "system", "content": "You are a helpful medical assistant."},
    {"role": "user", "content": "What are the common side effects of metformin?"},
]
input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Limitations

• This model is for educational and research purposes only — not for clinical decision-making.
• Trained primarily on conversational medical data; may generate verbose responses.
• May produce inaccurate or hallucinated medical information.
• Always consult qualified healthcare professionals for medical advice.