Text this: Activation-Guided Low-Rank Parameter Adaptation for Efficient Model Fine-Tuning