Efficient Low-Level Medical Question Answering (NLP Project)

Quantized LoRA Fine-Tuned LLM for MedQA This project implements a Quantization + Low-Rank Adaptation (QLoRA) pipeline for fine-tuning a lightweight large language model (LLM) on the MedQA dataset. The goal is to train a parameter-efficient, domain-adapted model capable of answering medical examination-style questions under strict hardware constraints (single GPU).

Objective: Fine-tune a compact, quantized LLM that maintains high accuracy on medical QA reasoning tasks while dramatically reducing GPU memory usage and training cost.

Key Features:

🔹 4-bit Quantization (bitsandbytes) — Compresses model weights to 4-bit without major performance loss. 🔹 LoRA Adapters (PEFT) — Injects small trainable matrices into attention layers for efficient adaptation. 🔹 Task-Specific Fine-Tuning (MedQA) — Adapts the base LLM to medical question-answering, focusing on reasoning and terminology. 🔹 Low GPU Memory Usage — Fine-tunes 7B-class models on a single RTX 3060 / 3090 GPU. 🔹 Evaluation Metrics — Reports accuracy, loss, and reasoning trace for each validation step.

Base Model (4-bit) ---> LoRA Injected Layers ---> Fine-Tuning on MedQA │ │ ▼ ▼ Quantization (bitsandbytes) Adaptation (PEFT)

Dataset:

MedQA (USMLE-format medical exam) Multiple-choice medical questions Domain: physiology, pathology, pharmacology, diagnosis Data split: train / validation / test Tokenized using SentencePiece or LLaMA tokenizer

Model	Params	Precision	GPU	Accuracy (MedQA)	Memory
Base LLaMA-2-7B	7B	FP16	RTX 3090	43%	~28 GB
QLoRA Fine-Tuned	7B	4-bit + LoRA	RTX 3060	47%	<8 GB