Omaratef3221/llama-3.1-8b-s1-full-s2-lora-medarabench

Base model: meta-llama/Llama-3.1-8B
Training stage: Stage 2 — Task Fine-tuning (MedAraBench)
Stage 1 method: full
Stage 2 method: lora

Paper

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

{
  "model_name": "meta-llama/Llama-3.1-8B",
  "stage1_method": "full",
  "stage2_method": "lora",
  "stage": "task_specific",
  "num_epochs": 3,
  "config_path": "/home/oelgendy/arabic-medical-llm/script/configs/lora.yaml",
  "stage1_checkpoint": "outputs/exp04_llama_full_lora/stage1",
  "train_samples": 16756,
  "val_samples": 882
}