H-LoRA: Rethinking Rank Selection for Controllable Knowledge Retention in Edge AI

The deployment of specialized language models in resource-constrained edge environments (≤1B parameters, ≤2 GB memory, ≤100 ms latency) faces a critical challenge: Supervised Fine-Tuning (SFT) achieves domain expertise but suffers from irreversible catastrophic forgetting, while traditional Low-Rank Adaptation (LoRA) with conservative ranks (r ≤ 64) often underperforms due to insufficient adaptation capacity. This work introduces H-LoRA (High-Rank LoRA) for edge-deployable models and establishes a fundamental distinction between destructive forgetting and controllable knowledge retention. Through comprehensive experiments on compact models (0.12B Minimind and Qwen-0.5B) across three domains (Human Resources, Medical, Mathematics) using 29,647 samples, we demonstrate that while both SFT and H-LoRA exhibit general capability degradation, they differ fundamentally: SFT completely destroys the original knowledge structure (1% topic retention), while H-LoRA maintains knowledge integrity with 90% topic retention—an 89 percentage point improvement—enabling post-deployment capability recovery. H-LoRA employs simplified scaling and strategic high-rank adaptation at approximately two-thirds of the model’s hidden dimension (r = 512 for d = 768), achieving SFT-level domain performance (99.81% precision) with 5× greater parameter efficiency (20.35% trainable parameters) and robust cross-domain generalization (93.5 ± 6.8% average precision). In addition, H-LoRA reduces over-the-air (OTA) update size from 1.4 GB to 96 MB (≈93%), enabling practical and frequent deployment of specialized models in bandwidth-limited edge environments. Beyond demonstrating effectiveness, this work establishes the first comprehensive framework for characterizing specialization-retention trade-offs in parameter-efficient fine-tuning, providing practical guidance for method selection in real-world deployments.