Train your cake and eat it too! Repurposing collaborative training to tailor LLMs to private data without sharing

Radovič, Boris (author), Aljahdali, Mohammed (author), Canini, Marco (author), Pejović, Veljko (author), Khayyat, Zuhair (author)

In the emerging field of large language models (LLMs), a significant challenge arises when organizations with vast datasets lack the computational resources to independently train and fine-tune models. This issue stems from privacy, compliance, and resource constraints: organizations cannot share their sensitive data but still need external computational assistance for model training. In this paper, we implement, enhance, and empirically compare several methods, including Split Learning (SL) and select Federated Learning (FL) methods, which enable data-rich yet compute-poor clients to offload LLM training without sharing raw data. Our study evaluates these methods across multiple dimensions, including model quality and training time.

• Language: English
• Keywords: energy usage, digital data
• Work type: Other
• Typology: 1.08 - Published Scientific Conference Contribution
• Organization: FRI - Faculty of Computer and Information Science
• Publication status: Published
• Publication version: Version of Record
• Year: 2024
• Number of pages: Str. 1-10
• PID: 20.500.12556/RUL-166792
• UDC: 004
• COBISS.SI-ID: 221660675

Copyright

Licenca Creative Commons je navedena na pristajalni strani članka (glej izvorni URL): "Licensed under Crative Commons Atribution 4.0 International". (Datum opombe 24. 1. 2025)

• Publication date in RUL: 24.01.2025

Record is a part of a monograph

• Title: Workshop on Efficient Systems for Foundation Models II : ES-FoMo-II 2024
• Place of publishing: [Massachusetts
• Publisher: OpenReview
• Year: 2024
• COBISS.SI-ID: 221659651

Licences

• License: CC BY 4.0, Creative Commons Attribution 4.0 International
• Link: http://creativecommons.org/licenses/by/4.0/
• Description: This is the standard Creative Commons license that gives others maximum freedom to do what they want with the work as long as they credit the author.

Secondary language

• Language: Slovenian
• Keywords: poraba energije, digitalni podatki