How to Train BERT with an Academic Budget

TL;DR

This paper presents a cost-effective method for pretraining BERT-like models within 24 hours on modest hardware, making large language models more accessible to resource-constrained settings.

Contribution

It introduces a practical recipe combining software, design, and hyperparameter optimizations for affordable BERT pretraining.

Findings

Models achieve competitive GLUE performance

Pretraining time reduced to 24 hours

Cost lowered significantly compared to traditional methods

Abstract

While large language models a la BERT are used ubiquitously in NLP, pretraining them is considered a luxury that only a few well-funded industry labs can afford. How can one train such models with a more modest budget? We present a recipe for pretraining a masked language model in 24 hours using a single low-end deep learning server. We demonstrate that through a combination of software optimizations, design choices, and hyperparameter tuning, it is possible to produce models that are competitive with BERT-base on GLUE tasks at a fraction of the original pretraining cost.