Federated Fine-Tuning

Synchronous federated learning can be bottlenecked by the slowest selected client in each round. Client updates are aggregated with FedAvg weights proportional to local sample counts among selected clients. Cross-device federated deployments are often constrained by uplink communication rather than local computation. The paper’s system model uses a central server and heterogeneous edge clients with non-IID private data distributions. Federated fine-tuning adapts large language models to private user or institutional data without centralizing that data.