Large language models (LLMs) achieve strong performance on a wide range of NLP tasks, yet conventional fine‑tuning relies on centralized data collection, raising privacy concerns. Federated learning (FL) enables collaborative LLM fine‑tuning without sharing raw client data, but deployment over bandwidth‑limited wireless networks is hampered by the high communication cost of transmitting model parameters. Low‑Rank Adaptation (LoRA) reduces the number of trainable parameters, yet its communication cost still grows with model size. Knowledge distillation avoids parameter sharing by using output logits, but token‑level logits in LLMs are costly due to long sequences and large vocabularies. Reducing logits dimensions cuts bandwidth but weakens supervision and harms accuracy. To overcome these limits, we propose FLoKD, an adaptive knowledge‑distillation framework for federated LoRA fine‑tuning of LLMs over wireless networks. FLoKD communicates intermediate LoRA activations as the distillation signal instead of logits or full parameters. To further cut communication, we introduce a transformer block importance scoring scheme that selectively transmits the most informative blocks, and two public‑dataset selection strategies that discard samples deviating from the local data distribution and prioritize those most beneficial for distillation. Extensive experiments on WikiText‑103, PTB, and dialogue datasets show that FLoKD reduces communication overhead by roughly 50%–65% while achieving rapid convergence to competitive perplexity compared with baselines.
Review