Enregistré dans:
Détails bibliographiques
Auteurs principaux: Schlotthauer, Joel, Kroos, Christian, Hinze, Chris, Hangya, Viktor, Hahn, Luzian, Küch, Fabian
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:https://arxiv.org/abs/2507.08472
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866918100847296512
author Schlotthauer, Joel
Kroos, Christian
Hinze, Chris
Hangya, Viktor
Hahn, Luzian
Küch, Fabian
author_facet Schlotthauer, Joel
Kroos, Christian
Hinze, Chris
Hangya, Viktor
Hahn, Luzian
Küch, Fabian
contents Optimizers play a decisive role in reducing pre-training times for LLMs and achieving better-performing models. In this study, we compare three major variants: the de-facto standard AdamW, the simpler Lion, developed through an evolutionary search, and the second-order optimizer Sophia. For better generalization, we train with two different base architectures and use a single- and a multiple-epoch approach while keeping the number of tokens constant. Using the Maximal Update Parametrization and smaller proxy models, we tune relevant hyperparameters separately for each combination of base architecture and optimizer. We found that while the results from all three optimizers were in approximately the same range, Sophia exhibited the lowest training and validation loss, Lion was fastest in terms of training GPU hours but AdamW led to the best downstream evaluation results.
format Preprint
id arxiv_https___arxiv_org_abs_2507_08472
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Pre-Training LLMs on a budget: A comparison of three optimizers
Schlotthauer, Joel
Kroos, Christian
Hinze, Chris
Hangya, Viktor
Hahn, Luzian
Küch, Fabian
Machine Learning
Artificial Intelligence
Optimizers play a decisive role in reducing pre-training times for LLMs and achieving better-performing models. In this study, we compare three major variants: the de-facto standard AdamW, the simpler Lion, developed through an evolutionary search, and the second-order optimizer Sophia. For better generalization, we train with two different base architectures and use a single- and a multiple-epoch approach while keeping the number of tokens constant. Using the Maximal Update Parametrization and smaller proxy models, we tune relevant hyperparameters separately for each combination of base architecture and optimizer. We found that while the results from all three optimizers were in approximately the same range, Sophia exhibited the lowest training and validation loss, Lion was fastest in terms of training GPU hours but AdamW led to the best downstream evaluation results.
title Pre-Training LLMs on a budget: A comparison of three optimizers
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2507.08472