Optimize low-resource text classification. Compare Bangla PLM fine-tuning via AdamW with zero-shot and few-shot (5 to 15-shot) generative prompt engineering.