On the Effect of Pretraining Corpora on In-context Learning by a Large-scale Language Model

TL;DR

This paper analyzes how the source and size of pretraining data affect in-context learning in large language models, revealing that domain relevance and data combination influence performance more than corpus size or perplexity.

Contribution

It provides an in-depth analysis of pretraining corpus effects on in-context learning, highlighting factors beyond size and perplexity that impact performance.

Findings

In-context learning depends heavily on corpus domain source.

Combining multiple corpora can enable in-context learning even if individual corpora do not.

Pretraining on task-related data does not always improve in-context learning for that task.

Abstract

Many recent studies on large-scale language models have reported successful in-context zero- and few-shot learning ability. However, the in-depth analysis of when in-context learning occurs is still lacking. For example, it is unknown how in-context learning performance changes as the training corpus varies. Here, we investigate the effects of the source and size of the pretraining corpus on in-context learning in HyperCLOVA, a Korean-centric GPT-3 model. From our in-depth investigation, we introduce the following observations: (1) in-context learning performance heavily depends on the corpus domain source, and the size of the pretraining corpus does not necessarily determine the emergence of in-context learning, (2) in-context learning ability can emerge when a language model is trained on a combination of multiple corpora, even when each corpus does not result in in-context learning on its own, (3) pretraining with a corpus related to a downstream task does not always guarantee the competitive in-context learning performance of the downstream task, especially in the few-shot setting, and (4) the relationship between language modeling (measured in perplexity) and in-context learning does not always correlate: e.g., low perplexity does not always imply high in-context few-shot learning performance.