Accuracy is Not All You Need

TL;DR

This paper reveals that model accuracy alone is insufficient to evaluate compressed LLMs, as they often exhibit significant qualitative differences and answer flips despite similar accuracy, advocating for additional distance metrics.

Contribution

The paper introduces the importance of using distance metrics like KL-Divergence and flips to better evaluate compressed models beyond accuracy.

Findings

Compressed models show many answer flips despite similar accuracy.

Qualitative evaluation reveals significant differences from baseline models.

KL-Divergence correlates well with answer flips as a metric.

Abstract

When Large Language Models (LLMs) are compressed using techniques such as quantization, the predominant way to demonstrate the validity of such techniques is by measuring the model's accuracy on various benchmarks.If the accuracies of the baseline model and the compressed model are close, it is assumed that there was negligible degradation in quality.However, even when the accuracy of baseline and compressed model are similar, we observe the phenomenon of flips, wherein answers change from correct to incorrect and vice versa in proportion.We conduct a detailed study of metrics across multiple compression techniques, models and datasets, demonstrating that the behavior of compressed models as visible to end-users is often significantly different from the baseline model, even when accuracy is similar.We further evaluate compressed models qualitatively and quantitatively using MT-Bench and show that compressed models are significantly worse than baseline models in this free-form generative task.Thus, we argue that compression techniques should also be evaluated using distance metrics.We propose two such metrics, KL-Divergence and flips, and show that they are well correlated.