SCAR: Data Selection via Style Consistency-Aware Response Ranking for Efficient Instruction-Tuning of Large Language Models

TL;DR

This paper introduces SCAR, a method for selecting high-quality, style-consistent training data for large language models, which improves performance while using significantly less data.

Contribution

The paper proposes Style Consistency-Aware Response Ranking (SCAR), a novel approach to prioritize training examples based on stylistic consistency, enhancing LLM fine-tuning efficiency.

Findings

Using only 0.7% of data, models achieve comparable or better performance.

Higher stylistic consistency correlates with improved LLM performance.

SCAR outperforms baseline data selection methods.

Abstract

Recent studies emphasize that manually ensuring a consistent response style and maintaining high data quality in training sets can significantly improve the performance of fine-tuned Large Language Models (LLMs) while reducing the number of training examples needed. However, the precise definition of style and the relationship between style, data quality, and LLM performance remains unclear. This research identifies two key stylistic elements in responses: linguistic form and instructional surprisal. We find that, among training data of comparable quality, higher consistency in these response elements leads to better LLM performance. Inspired by this, we introduce Style Consistency-Aware Response Ranking (SCAR), which automatically prioritizes instruction-response pairs in the training set based on their response stylistic consistency. By selecting the most style-consistent examples, using only 0.7% of the full dataset in the best case, the fine-tuned LLMs can match or even surpass the performance of models trained on the entire dataset in coding and open-ended question-answering benchmarks. Code and data are available at https://github.com/zhuang-li/SCAR .