STLM Engineering Report: Dropout

TL;DR

This paper investigates the effectiveness of dropout in small and large language models, revealing its continued relevance for overfitting and potential benefits for model fitting even with abundant data, challenging existing explanations.

Contribution

It provides new insights into dropout's role in language models under different data regimes, especially for models under 100M parameters.

Findings

Dropout remains effective for overfitting scenarios.

Dropout may improve model fit even with large datasets.

Existing explanations for dropout's mechanism are not fully applicable to language models.

Abstract

In this work we explore the relevance of dropout for modern language models, particularly in the context of models on the scale of <100M parameters. We explore it's relevance firstly in the regime of improving the sample efficiency of models given small, high quality datasets, and secondly in the regime of improving the quality of its fit on larger datasets where models may underfit. We find that concordant with conventional wisdom, dropout remains effective in the overfitting scenario, and that furthermore it may have some relevance for improving the fit of models even in the case of excess data, as suggested by previous research. In the process we find that the existing explanation for the mechanism behind this performance gain is not applicable in the case of language modelling.