Decentrai is an open-source framework for hierarchical federated learning, to be released soon. Clients train on their own data, regional coordinators aggregate their clients’ updates, and a global orchestrator combines the coordinators’ results into the next global model. Raw data never leaves the clients.
It includes 18 aggregation algorithms (FedAvg, FedProx, SCAFFOLD, Krum and others), differential privacy, and LoRA fine-tuning of language models across sites.