ResearchBenchmarksTraining Time on QQP (Paraphrase Detection)Follow112.35Training Time (hours)FedCIO109.5464128.4707147.395166.3193Nov 28, 2025Evaluation ResultsMethodMethodLinksTraining Time (hours)Per-Sequence Training Cost (hours)FedCIOModel=RoBERTLARGEModel=RoBERTLARGE2025.11112.35—FedSGTModel=RoBERTLARGE, B (...Model=RoBERTLARGE, B (sequences)=102025.1118118.1FedAvgModel=RoBERTLARGEModel=RoBERTLARGE2025.11182.44—