Importance-Aware Data Selection for Efficient LLM Instruction Tuning

TL;DR

This work tackles the inefficiency of instruction-tuning data by introducing MIWV, a metric that quantifies the value of instruction samples via in-context learning discrepancies. The authors propose a three-step universal data-selection pipeline—one-shot example retrieval, MIWV computation, and high-quality data selection—demonstrating that training on a tiny, MIWV-high subset can match or exceed performance achieved with the full dataset across multiple benchmarks and backbones. Extensive experiments on Alpaca and WizardLM show robust improvements and competitive rankings on Open LLM Leaderboard and AlpacaEval, with ablations validating the method's effectiveness across embedding models and architectures. The approach offers a practical, resource-efficient pathway for data utilization in instruction tuning, with broader implications for scalable and targeted LLM enhancement.

Abstract

Instruction tuning plays a critical role in enhancing the performance and efficiency of Large Language Models (LLMs). Its success depends not only on the quality of the instruction data but also on the inherent capabilities of the LLM itself. Some studies suggest that even a small amount of high-quality data can achieve instruction fine-tuning results that are on par with, or even exceed, those from using a full-scale dataset. However, rather than focusing solely on calculating data quality scores to evaluate instruction data, there is a growing need to select high-quality data that maximally enhances the performance of instruction tuning for a given LLM. In this paper, we propose the Model Instruction Weakness Value (MIWV) as a novel metric to quantify the importance of instruction data in enhancing model's capabilities. The MIWV metric is derived from the discrepancies in the model's responses when using In-Context Learning (ICL), helping identify the most beneficial data for enhancing instruction tuning performance. Our experimental results demonstrate that selecting only the top 1\% of data based on MIWV can outperform training on the full dataset. Furthermore, this approach extends beyond existing research that focuses on data quality scoring for data selection, offering strong empirical evidence supporting the effectiveness of our proposed method.

Figures (14)

• Figure 1: Overview of our proposed method.
• Figure 2: Comparing our models trained on selected data with full data. Both (a) and (b) use GPT-4 as the judge.
• Figure 3: The comparison of the win rate between models and the official Alpaca model of ablation study.
• Figure 4: Comparison of Qwen2.5 models trained on selected and full data.
• Figure 5: Case Study on Mathematical Capabilities: LLaMA2-7B Models of Alpaca 1% and Alpaca 100%.