Saved in:
Bibliographic Details
Main Authors: Takase, Sho, Kiyono, Shun, Kobayashi, Sosuke, Suzuki, Jun
Format: Preprint
Published: 2023
Subjects:
Online Access:https://arxiv.org/abs/2312.16903
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913959334903808
author Takase, Sho
Kiyono, Shun
Kobayashi, Sosuke
Suzuki, Jun
author_facet Takase, Sho
Kiyono, Shun
Kobayashi, Sosuke
Suzuki, Jun
contents Loss spikes often occur during pre-training of large language models. The spikes degrade the performance of large language models and sometimes ruin the pre-training. Since the pre-training needs a vast computational budget, we should avoid such spikes. Based on the assumption that the loss spike is caused by the sudden growth of the gradient norm, we explore factors to keep the gradient norm small through an analysis of the spectral norms of the Jacobian matrices for the sub-layers. Our findings suggest that stabilizing the pre-training process requires two conditions: small sub-layers and large shortcut. We conduct various experiments to empirically verify our theoretical analyses. Experimental results demonstrate that methods satisfying the conditions effectively prevent loss spikes during pre-training.
format Preprint
id arxiv_https___arxiv_org_abs_2312_16903
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Spike No More: Stabilizing the Pre-training of Large Language Models
Takase, Sho
Kiyono, Shun
Kobayashi, Sosuke
Suzuki, Jun
Computation and Language
Artificial Intelligence
Loss spikes often occur during pre-training of large language models. The spikes degrade the performance of large language models and sometimes ruin the pre-training. Since the pre-training needs a vast computational budget, we should avoid such spikes. Based on the assumption that the loss spike is caused by the sudden growth of the gradient norm, we explore factors to keep the gradient norm small through an analysis of the spectral norms of the Jacobian matrices for the sub-layers. Our findings suggest that stabilizing the pre-training process requires two conditions: small sub-layers and large shortcut. We conduct various experiments to empirically verify our theoretical analyses. Experimental results demonstrate that methods satisfying the conditions effectively prevent loss spikes during pre-training.
title Spike No More: Stabilizing the Pre-training of Large Language Models
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2312.16903