FDAPT: Federated Domain-adaptive Pre-training for Language Models
This work addresses the challenge of applying foundation models to specific domains with limited data and computational resources, offering a federated learning solution that is incremental but practical for real-world deployments.
The paper tackles the problem of domain-adaptive pre-training for language models in federated settings, demonstrating that federated domain-adaptive pre-training (FDAPT) maintains competitive downstream task performance compared to centralized baselines, and proposing a novel algorithm (FFDAPT) that improves computational efficiency by 12.1% on average with performance fluctuations under 1%.
Foundation models (FMs) have shown prominent success in a wide range of tasks. Their applicability to specific domain-task pairings relies on the availability of, both, high-quality data and significant computational resources. These challenges are not new to the field and, indeed, Federated Learning (FL) has been shown to be a promising solution in similar setups. This paper tackles the specific case of Domain-Adaptive Pre-Training (DAPT), a key step in the application of FMs. We conduct the first comprehensive empirical study to evaluate the performance of Federated Domain-Adaptive Pre-Training (FDAPT). We demonstrate that FDAPT can maintain competitive downstream task performance to the centralized baseline in both IID and non-IID situations. Finally, we propose a novel algorithm, Frozen Federated Domain-Adaptive Pre-Training (FFDAPT). FFDAPT improves the computational efficiency by 12.1% on average and exhibits similar downstream task performance to vanilla FDAPT, with general performance fluctuations remaining less than 1%.