FedBiOT: LLM Local Fine-tuning in Federated Learning without Full Model

FedBiOT: LLM Local Fine-tuning in Federated Learning without Full Model

25 Jun 2024 | Feijie Wu, Zitao Li, Yaliang Li, Bolin Ding, Jing Gao
FedBiOT is a federated learning approach for local fine-tuning of large language models (LLMs) without requiring access to the full model. The method compresses the LLM into two components: an emulator and an adapter. The emulator simulates the behavior of the full model, while the adapter learns domain-specific linguistic patterns from clients' data. The server compresses the LLM and aligns its performance with the full model, while clients fine-tune the adapter. The problem is formulated as a bi-level optimization problem to minimize the negative impact of data discrepancy between the server and clients. The server and clients update their parameters accordingly. Extensive experiments on LLaMA-2 show that FedBiOT significantly reduces resource consumption compared to existing benchmarks while achieving comparable performance. The method uses parameter-efficient fine-tuning techniques such as LoRA to reduce the number of trainable parameters and communication costs. The emulator is trained on a public dataset to mimic the full model, while the adapter is fine-tuned on clients' data. The proposed FedBiOT achieves better performance than existing methods in tasks such as math problem-solving, code generation, and question answering. The method is efficient in terms of computation and communication overhead, making it suitable for federated learning scenarios where clients have limited resources.FedBiOT is a federated learning approach for local fine-tuning of large language models (LLMs) without requiring access to the full model. The method compresses the LLM into two components: an emulator and an adapter. The emulator simulates the behavior of the full model, while the adapter learns domain-specific linguistic patterns from clients' data. The server compresses the LLM and aligns its performance with the full model, while clients fine-tune the adapter. The problem is formulated as a bi-level optimization problem to minimize the negative impact of data discrepancy between the server and clients. The server and clients update their parameters accordingly. Extensive experiments on LLaMA-2 show that FedBiOT significantly reduces resource consumption compared to existing benchmarks while achieving comparable performance. The method uses parameter-efficient fine-tuning techniques such as LoRA to reduce the number of trainable parameters and communication costs. The emulator is trained on a public dataset to mimic the full model, while the adapter is fine-tuned on clients' data. The proposed FedBiOT achieves better performance than existing methods in tasks such as math problem-solving, code generation, and question answering. The method is efficient in terms of computation and communication overhead, making it suitable for federated learning scenarios where clients have limited resources.
Reach us at info@study.space