Measuring the Instability of Fine-Tuning

TL;DR

This paper critically examines various measures of instability in fine-tuning pre-trained language models, proposing a systematic framework to evaluate their validity and comparing their effectiveness across different mitigation methods.

Contribution

It introduces a comprehensive framework to assess multiple instability measures and analyzes their consistency, advancing the understanding of how to accurately quantify fine-tuning instability.

Findings

SD alone is insufficient to characterize instability

Six additional measures provide a more granular view of instability

Different measures vary in their sensitivity and reliability

Abstract

Fine-tuning pre-trained language models on downstream tasks with varying random seeds has been shown to be unstable, especially on small datasets. Many previous studies have investigated this instability and proposed methods to mitigate it. However, most studies only used the standard deviation of performance scores (SD) as their measure, which is a narrow characterization of instability. In this paper, we analyze SD and six other measures quantifying instability at different levels of granularity. Moreover, we propose a systematic framework to evaluate the validity of these measures. Finally, we analyze the consistency and difference between different measures by reassessing existing instability mitigation methods. We hope our results will inform the development of better measurements of fine-tuning instability.