# LESS: Selecting Influential Data for Targeted Instruction Tuning

**Authors:** Mengzhou Xia, Sadhika Malladi, Suchin Gururangan, Sanjeev Arora, Danqi Chen  
**arXiv:** [2402.04333](https://arxiv.org/abs/2402.04333)  
**Date:** 2024-02-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.04333v1/

## Key points

- Authors discuss challenges in developing specialized capabilities in large language models (LLMs) for real-world applications
- LESS algorithm offers a practical and efficient solution for targeted instruction tuning in LLMs
- Algorithm constructs a gradient datastore with low-dimensional features for effective reuse and transferability
- LESS demonstrates superior performance by selecting examples based on similarity to few-shot instances representing specific capabilities
- Selected data exhibits high transferability, enabling smaller models to identify useful data for larger models and different model families

## Summary

, , , , 

In their paper titled "LESS: Selecting Influential Data for Targeted Instruction Tuning," authors Mengzhou Xia, Sadhika Malladi, Suchin Gururangan, Sanjeev Arora, and Danqi Chen discuss the challenges of developing specialized capabilities in large language models (LLMs) for real-world applications. LESS offers a practical and efficient solution for targeted instruction tuning in LLMs, showcasing its potential to enhance model performance and applicability in diverse real-world scenarios. The algorithm constructs a gradient datastore with low-dimensional features that can be reused and transferred effectively. By selecting examples based on their similarity to few-shot instances representing specific capabilities, LESS demonstrates superior performance when training on a selected 5% subset of data compared to using the entire dataset across various downstream tasks. Importantly, the selected data exhibits high transferability, enabling smaller models to identify useful data for larger models and models from different families. <ks>Influence of Data Selection on LLMs</ks>

While instruction tuning has enabled the development of general-purpose chatbots by leveraging combined datasets, tasks requiring specific skills such as reasoning necessitate a targeted approach to data selection. The authors introduce LESS, an algorithm designed to estimate data influences and perform Low-rank gradiEnt Similarity Search for selecting instruction data efficiently. LESS is characterized by its adaptability to the Adam optimizer and variable-length instruction data. <ks>Efficient Selection Process with LESS</ks>

Through qualitative analysis, the authors highlight that LESS goes beyond surface-level cues to identify data exemplifying essential reasoning skills required for intended downstream applications. Overall, LESS offers a practical and efficient solution for targeted instruction tuning in LLMs, showcasing its potential to enhance model performance and applicability in diverse real-world scenarios.