Chasing Random: Instruction Selection Strategies Fail to Generalize

TL;DR

This paper critically examines instruction selection strategies for language models, revealing that they often fail to generalize well and may not be cost-effective compared to using full datasets or random subsets.

Contribution

It provides a comprehensive analysis of popular instruction selection methods across various datasets and benchmarks, highlighting their limited generalization and cost-effectiveness.

Findings

Selection strategies often do not outperform random baselines.

Data selection can be more costly than fine-tuning on full datasets.

Limited gains from data selection strategies in many scenarios.

Abstract

Prior work has shown that language models can be tuned to follow user instructions using only a small set of high-quality instructions. This has accelerated the development of methods that filter a large, noisy instruction-tuning datasets down to high-quality subset which works just as well. However, typically, the performance of these methods is not demonstrated across a uniform experimental setup and thus their generalization capabilities are not well established. In this work, we analyze popular selection strategies across different source datasets, selection budgets and evaluation benchmarks: Our results indicate that selection strategies generalize poorly, often failing to consistently outperform even random baselines. We also analyze the cost-performance trade-offs of using data selection. Our findings reveal that data selection can often exceed the cost of fine-tuning on the full dataset, yielding only marginal and sometimes no gains compared to tuning on the full dataset or a random subset.