The problem
Fine-tuning a language model on private documents gives much better answers, but no organisation wants to send its data to another. LoRA makes fine-tuning cheap by training small adapters; how to merge adapters from different organisations with different data is still open.
Who needs it
Healthcare, legal and finance groups, multi-campus universities; data-protection laws forbid pooling the data.
Idea of the approach
- Small model (e.g. Qwen2.5-0.5B/1.5B) with LoRA, 3–5 simulated organisations with different instruction datasets.
- Implement federated averaging of adapters; compare with local-only and centralized fine-tuning.
- Study heterogeneous ranks and personalization (keep part of the adapter local).
- Measure quality, communication and GPU time.
What you will build and deliver
- Training code and experiments.
- Paper.
Skills needed
Python, PyTorch, Hugging Face PEFT; GPU access needed.
Research contribution & where to publish
Contribution: personalized and rank-heterogeneous federated LoRA. Venues: FL workshops at NeurIPS/ICML, ICASSP, IEEE Transactions on Mobile Computing.
Key references — start reading here
- E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen. LoRA: Low-Rank Adaptation of Large Language Models. ICLR, 2022.
- J. Zhang, S. Vahidian, M. Kuo, C. Li, R. Zhang, T. Yu, G. Wang, Y. Chen. Towards Building the FederatedGPT: Federated Instruction Tuning. IEEE ICASSP, 2024.
- H. Djigal et al.. ProLoRA: Resource-Efficient Personalized Federated Learning for Sensor Based Human Activity Recognition. , 2024.
- T. Dettmers, A. Pagnoni, A. Holtzman, L. Zettlemoyer. QLoRA: Efficient Finetuning of Quantized LLMs. NeurIPS, 2023.
Interested?
Send the topic title and the courses you have taken.
Contact Dr. Hamza Djigal