LoRA vs. Full Fine-Tuning for Arabic Medical Question Answering: A Systematic Comparison Across General-Purpose and Arabic-Centric Large Language Models

Training data

Stage	Dataset	Samples
Stage 1	AraMed (open-ended Arabic medical QA)	~110K
Stage 2	MedAraBench (Arabic MCQ)	~17.6K (cleaned)

Evaluation

Evaluated on MedAraBench test set (4,959 MCQ samples) using log-probability selection. Metrics: Accuracy and Macro F1 across answer classes A–E.

Experiment metadata

json

{
  "model_name": "meta-llama/Llama-3.1-8B",
  "method": "full",
  "stage": "domain_adaptation",
  "num_epochs": 1,
  "config_path": "/home/oelgendy/arabic-medical-llm/script/configs/full_ft.yaml",
  "train_samples": 107637,
  "val_samples": 2197
}