Scholarlink

The rapid advancement of artificial intelligence has enabled language models to excel across many domains. Traditional centralized training, however, raises privacy concerns and demands extensive computational resources. This paper investigates federated learning (FL) for training small language models (SLMs), introducing a stage-wise hybrid strategy that applies different FL algorithms at different training phases. Two SLMs, Gemma-2B and Qwen2.5-1.5B, were evaluated during instruction fine-tuning and alignment fine-tuning with seven FL algorithms, FedAvg, SCAFFOLD, FedProx, FedYogi, FedAvgM, FedAdagrad, and FedAdam, compared to local baselines. Results show that Gemma performs best with FedAvg and FedAdagrad, while Qwen excels with SCAFFOLD and FedAvgM in the respective stages, demonstrating that FL effectiveness depends on both model design and training phase. The proposed hybrid strategy improves training efficiency, enables privacy-preserving deployment on resource-limited devices, and provides a practical framework for applying SLMs in medical, educational, and distributed environments.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Federated Learning for Small Language Models: A Cutting-Edge AI Paradigm for Privacy and Efficiency

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mohamad Naji,
  • Ahmed Freidoon Fadhil,
  • Jawad Khanafer,
  • Alaa Farhat,
  • Ali Anaissi

摘要

The rapid advancement of artificial intelligence has enabled language models to excel across many domains. Traditional centralized training, however, raises privacy concerns and demands extensive computational resources. This paper investigates federated learning (FL) for training small language models (SLMs), introducing a stage-wise hybrid strategy that applies different FL algorithms at different training phases. Two SLMs, Gemma-2B and Qwen2.5-1.5B, were evaluated during instruction fine-tuning and alignment fine-tuning with seven FL algorithms, FedAvg, SCAFFOLD, FedProx, FedYogi, FedAvgM, FedAdagrad, and FedAdam, compared to local baselines. Results show that Gemma performs best with FedAvg and FedAdagrad, while Qwen excels with SCAFFOLD and FedAvgM in the respective stages, demonstrating that FL effectiveness depends on both model design and training phase. The proposed hybrid strategy improves training efficiency, enables privacy-preserving deployment on resource-limited devices, and provides a practical framework for applying SLMs in medical, educational, and distributed environments.