# Federated Data-Efficient Instruction Tuning for Large Language Models

**Authors:** Zhen Qin, Zhaomin Wu, Bingsheng He, Shuiguang Deng  
**arXiv:** [2410.10926](https://arxiv.org/abs/2410.10926)  
**Date:** 2024-10-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2410.10926v1/

## Key points

- Instruction tuning is crucial for enhancing large language models' responsiveness to human instructions.
- Federated learning leverages diverse client-side data sources to enhance LLM tuning.
- Traditional approaches to federated LLM tuning can lead to excessive computational overhead and overfitting local data.
- FedHDS is a novel approach that uses a representative subset of edge-side data (coreset) for fine-tuning LLMs.
- FedHDS reduces redundancy in data samples at both intra-client and inter-client levels through hierarchical data selection.
- Extensive experiments have shown that FedHDS significantly reduces the volume of data required for fine-tuning while improving responsiveness to unseen tasks in various scenarios.
- FedHDS has the potential to optimize LLM performance by efficiently utilizing instructional data within a federated learning framework.

## Summary

In the realm of large language models (LLMs), instruction tuning plays a crucial role in enhancing their responsiveness to human instructions. This improvement is largely attributed to the utilization of diverse instruction data. Federated learning has emerged as a powerful technique that leverages varied client-side data sources to further enhance LLM tuning, making it a popular choice in the field. However, traditional approaches to federated LLM tuning often involve exhaustive traversal of all local data during training, leading to excessive computational overhead and the potential risk of overfitting local data. To address these challenges, there is a growing need for a federated data-efficient instruction tuning approach that minimizes the amount of data required from the entire dataset. In response to this demand, a novel approach known as FedHDS has been introduced. FedHDS makes use of a representative subset of edge-side data called coreset to fine-tune LLMs. By implementing a hierarchical data selection framework, FedHDS effectively reduces redundancy in data samples at both intra-client and inter-client levels. This process involves jointly selecting a small number of representative data samples for local training without sharing raw data. Extensive experiments conducted across six scenarios involving various LLMs, datasets, and data partitions have demonstrated the efficacy of FedHDS. Notably, this approach significantly reduces the volume of data required for fine-tuning while simultaneously enhancing the responsiveness of instruction-tuned LLMs to unseen tasks. The findings underscore the potential impact of FedHDS in optimizing LLM performance through efficient utilization of instructional data within a federated learning framework.