The FlowerTune LLM Leaderboard benchmarks federated fine-tuning of 26 LLMs across four domains, enabling privacy-preserving, domain-specific models through decentralized training without sharing raw data.
Photon is the first full system for federated LLM training, enabling global-scale, low-bandwidth pre-training. It trains 7B models faster than baselines with better perplexity and 64x–512x less communication.
DEPT is a communication-efficient framework for federated LLM pre-training, enabling vocabulary-agnostic training across diverse data while cutting embedding memory 4–5x and improving perplexity up to 20%.
Flower and FLARE integrate to boost the FL ecosystem—Flower supports FL development and research, while FLARE provides a production-ready runtime. Integration enables seamless, modification-free deployment.
WorldLM enables global federated LLM training across diverse legal and data settings via federations of federations, using partial model localization. It outperforms standard setups by up to 1.91×.
Photon enables scalable federated LLM training across institutions, unlocking underused global data and compute. It supports billion-scale models, matching centralized performance with robust convergence.
MedPerf is an open framework for benchmarking medical AI using federated evaluation, enabling privacy-first, large-scale, real-world testing of models across healthcare facilities.
Flower is a new FL framework enabling large-scale, realistic federated learning on heterogeneous edge devices, supporting up to 15M clients and smooth migration from simulation to real devices.
