Temporal Scaling Law for Large Language Models

Recently, Large Language Models (LLMs) have been widely adopted in a wide range of tasks, leading to increasing attention towards the research on how scaling LLMs affects their performance. Existing works, termed Scaling Laws, have discovered that the final test loss of LLMs scales as power-laws wit...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Xiong, Yizhe, Chen, Xiansheng, Ye, Xin, Chen, Hui, Lin, Zijia, Lian, Haoran, Su, Zhenpeng, Niu, Jianwei, Ding, Guiguang
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page
container_issue
container_start_page
container_title
container_volume
creator Xiong, Yizhe
Chen, Xiansheng
Ye, Xin
Chen, Hui
Lin, Zijia
Lian, Haoran
Su, Zhenpeng
Niu, Jianwei
Ding, Guiguang
description Recently, Large Language Models (LLMs) have been widely adopted in a wide range of tasks, leading to increasing attention towards the research on how scaling LLMs affects their performance. Existing works, termed Scaling Laws, have discovered that the final test loss of LLMs scales as power-laws with model size, computational budget, and dataset size. However, the temporal change of the test loss of an LLM throughout its pre-training process remains unexplored, though it is valuable in many aspects, such as selecting better hyperparameters \textit{directly} on the target LLM. In this paper, we propose the novel concept of Temporal Scaling Law, studying how the test loss of an LLM evolves as the training steps scale up. In contrast to modeling the test loss as a whole in a coarse-grained manner, we break it down and dive into the fine-grained test loss of each token position, and further develop a dynamic hyperbolic-law. Afterwards, we derive the much more precise temporal scaling law by studying the temporal patterns of the parameters in the dynamic hyperbolic-law. Results on both in-distribution (ID) and out-of-distribution (OOD) validation datasets demonstrate that our temporal scaling law accurately predicts the test loss of LLMs across training steps. Our temporal scaling law has broad practical applications. First, it enables direct and efficient hyperparameter selection on the target LLM, such as data mixture proportions. Secondly, viewing the LLM pre-training dynamics from the token position granularity provides some insights to enhance the understanding of LLM pre-training.
doi_str_mv 10.48550/arxiv.2404.17785
format Article
fullrecord <record><control><sourceid>arxiv_GOX</sourceid><recordid>TN_cdi_arxiv_primary_2404_17785</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2404_17785</sourcerecordid><originalsourceid>FETCH-LOGICAL-a675-48dd8a1dd1393041bd433d75a319b8d46fa31fae0d3db9974ad0e35523cd890b3</originalsourceid><addsrcrecordid>eNotjjkOwjAQRd1QoMABqMgFEmzGxnaJEJsUREH6aMI4UaQsyIjt9oSl-e9VX4-xieCxNErxGfpndY_nkstYaG3UkMWpay6dxzo8nbGu2jJM8BEWne_pS9dvW96wl0NHrr6O2KDA-urGfwYs3azT1S5Kjtv9aplEuNAqkobIoCASYIFLkZMEIK0QhM0NyUXRW4GOE1BurZZI3IFScziTsTyHgE1_t9_g7OKrBv0r-4Rn33B4AzC-PIs</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>Temporal Scaling Law for Large Language Models</title><source>arXiv.org</source><creator>Xiong, Yizhe ; Chen, Xiansheng ; Ye, Xin ; Chen, Hui ; Lin, Zijia ; Lian, Haoran ; Su, Zhenpeng ; Niu, Jianwei ; Ding, Guiguang</creator><creatorcontrib>Xiong, Yizhe ; Chen, Xiansheng ; Ye, Xin ; Chen, Hui ; Lin, Zijia ; Lian, Haoran ; Su, Zhenpeng ; Niu, Jianwei ; Ding, Guiguang</creatorcontrib><description>Recently, Large Language Models (LLMs) have been widely adopted in a wide range of tasks, leading to increasing attention towards the research on how scaling LLMs affects their performance. Existing works, termed Scaling Laws, have discovered that the final test loss of LLMs scales as power-laws with model size, computational budget, and dataset size. However, the temporal change of the test loss of an LLM throughout its pre-training process remains unexplored, though it is valuable in many aspects, such as selecting better hyperparameters \textit{directly} on the target LLM. In this paper, we propose the novel concept of Temporal Scaling Law, studying how the test loss of an LLM evolves as the training steps scale up. In contrast to modeling the test loss as a whole in a coarse-grained manner, we break it down and dive into the fine-grained test loss of each token position, and further develop a dynamic hyperbolic-law. Afterwards, we derive the much more precise temporal scaling law by studying the temporal patterns of the parameters in the dynamic hyperbolic-law. Results on both in-distribution (ID) and out-of-distribution (OOD) validation datasets demonstrate that our temporal scaling law accurately predicts the test loss of LLMs across training steps. Our temporal scaling law has broad practical applications. First, it enables direct and efficient hyperparameter selection on the target LLM, such as data mixture proportions. Secondly, viewing the LLM pre-training dynamics from the token position granularity provides some insights to enhance the understanding of LLM pre-training.</description><identifier>DOI: 10.48550/arxiv.2404.17785</identifier><language>eng</language><subject>Computer Science - Computation and Language</subject><creationdate>2024-04</creationdate><rights>http://arxiv.org/licenses/nonexclusive-distrib/1.0</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>228,230,780,885</link.rule.ids><linktorsrc>$$Uhttps://arxiv.org/abs/2404.17785$$EView_record_in_Cornell_University$$FView_record_in_$$GCornell_University$$Hfree_for_read</linktorsrc><backlink>$$Uhttps://doi.org/10.48550/arXiv.2404.17785$$DView paper in arXiv$$Hfree_for_read</backlink></links><search><creatorcontrib>Xiong, Yizhe</creatorcontrib><creatorcontrib>Chen, Xiansheng</creatorcontrib><creatorcontrib>Ye, Xin</creatorcontrib><creatorcontrib>Chen, Hui</creatorcontrib><creatorcontrib>Lin, Zijia</creatorcontrib><creatorcontrib>Lian, Haoran</creatorcontrib><creatorcontrib>Su, Zhenpeng</creatorcontrib><creatorcontrib>Niu, Jianwei</creatorcontrib><creatorcontrib>Ding, Guiguang</creatorcontrib><title>Temporal Scaling Law for Large Language Models</title><description>Recently, Large Language Models (LLMs) have been widely adopted in a wide range of tasks, leading to increasing attention towards the research on how scaling LLMs affects their performance. Existing works, termed Scaling Laws, have discovered that the final test loss of LLMs scales as power-laws with model size, computational budget, and dataset size. However, the temporal change of the test loss of an LLM throughout its pre-training process remains unexplored, though it is valuable in many aspects, such as selecting better hyperparameters \textit{directly} on the target LLM. In this paper, we propose the novel concept of Temporal Scaling Law, studying how the test loss of an LLM evolves as the training steps scale up. In contrast to modeling the test loss as a whole in a coarse-grained manner, we break it down and dive into the fine-grained test loss of each token position, and further develop a dynamic hyperbolic-law. Afterwards, we derive the much more precise temporal scaling law by studying the temporal patterns of the parameters in the dynamic hyperbolic-law. Results on both in-distribution (ID) and out-of-distribution (OOD) validation datasets demonstrate that our temporal scaling law accurately predicts the test loss of LLMs across training steps. Our temporal scaling law has broad practical applications. First, it enables direct and efficient hyperparameter selection on the target LLM, such as data mixture proportions. Secondly, viewing the LLM pre-training dynamics from the token position granularity provides some insights to enhance the understanding of LLM pre-training.</description><subject>Computer Science - Computation and Language</subject><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2024</creationdate><recordtype>article</recordtype><sourceid>GOX</sourceid><recordid>eNotjjkOwjAQRd1QoMABqMgFEmzGxnaJEJsUREH6aMI4UaQsyIjt9oSl-e9VX4-xieCxNErxGfpndY_nkstYaG3UkMWpay6dxzo8nbGu2jJM8BEWne_pS9dvW96wl0NHrr6O2KDA-urGfwYs3azT1S5Kjtv9aplEuNAqkobIoCASYIFLkZMEIK0QhM0NyUXRW4GOE1BurZZI3IFScziTsTyHgE1_t9_g7OKrBv0r-4Rn33B4AzC-PIs</recordid><startdate>20240427</startdate><enddate>20240427</enddate><creator>Xiong, Yizhe</creator><creator>Chen, Xiansheng</creator><creator>Ye, Xin</creator><creator>Chen, Hui</creator><creator>Lin, Zijia</creator><creator>Lian, Haoran</creator><creator>Su, Zhenpeng</creator><creator>Niu, Jianwei</creator><creator>Ding, Guiguang</creator><scope>AKY</scope><scope>GOX</scope></search><sort><creationdate>20240427</creationdate><title>Temporal Scaling Law for Large Language Models</title><author>Xiong, Yizhe ; Chen, Xiansheng ; Ye, Xin ; Chen, Hui ; Lin, Zijia ; Lian, Haoran ; Su, Zhenpeng ; Niu, Jianwei ; Ding, Guiguang</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-LOGICAL-a675-48dd8a1dd1393041bd433d75a319b8d46fa31fae0d3db9974ad0e35523cd890b3</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2024</creationdate><topic>Computer Science - Computation and Language</topic><toplevel>online_resources</toplevel><creatorcontrib>Xiong, Yizhe</creatorcontrib><creatorcontrib>Chen, Xiansheng</creatorcontrib><creatorcontrib>Ye, Xin</creatorcontrib><creatorcontrib>Chen, Hui</creatorcontrib><creatorcontrib>Lin, Zijia</creatorcontrib><creatorcontrib>Lian, Haoran</creatorcontrib><creatorcontrib>Su, Zhenpeng</creatorcontrib><creatorcontrib>Niu, Jianwei</creatorcontrib><creatorcontrib>Ding, Guiguang</creatorcontrib><collection>arXiv Computer Science</collection><collection>arXiv.org</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext_linktorsrc</fulltext></delivery><addata><au>Xiong, Yizhe</au><au>Chen, Xiansheng</au><au>Ye, Xin</au><au>Chen, Hui</au><au>Lin, Zijia</au><au>Lian, Haoran</au><au>Su, Zhenpeng</au><au>Niu, Jianwei</au><au>Ding, Guiguang</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>Temporal Scaling Law for Large Language Models</atitle><date>2024-04-27</date><risdate>2024</risdate><abstract>Recently, Large Language Models (LLMs) have been widely adopted in a wide range of tasks, leading to increasing attention towards the research on how scaling LLMs affects their performance. Existing works, termed Scaling Laws, have discovered that the final test loss of LLMs scales as power-laws with model size, computational budget, and dataset size. However, the temporal change of the test loss of an LLM throughout its pre-training process remains unexplored, though it is valuable in many aspects, such as selecting better hyperparameters \textit{directly} on the target LLM. In this paper, we propose the novel concept of Temporal Scaling Law, studying how the test loss of an LLM evolves as the training steps scale up. In contrast to modeling the test loss as a whole in a coarse-grained manner, we break it down and dive into the fine-grained test loss of each token position, and further develop a dynamic hyperbolic-law. Afterwards, we derive the much more precise temporal scaling law by studying the temporal patterns of the parameters in the dynamic hyperbolic-law. Results on both in-distribution (ID) and out-of-distribution (OOD) validation datasets demonstrate that our temporal scaling law accurately predicts the test loss of LLMs across training steps. Our temporal scaling law has broad practical applications. First, it enables direct and efficient hyperparameter selection on the target LLM, such as data mixture proportions. Secondly, viewing the LLM pre-training dynamics from the token position granularity provides some insights to enhance the understanding of LLM pre-training.</abstract><doi>10.48550/arxiv.2404.17785</doi><oa>free_for_read</oa></addata></record>
fulltext fulltext_linktorsrc
identifier DOI: 10.48550/arxiv.2404.17785
ispartof
issn
language eng
recordid cdi_arxiv_primary_2404_17785
source arXiv.org
subjects Computer Science - Computation and Language
title Temporal Scaling Law for Large Language Models
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2024-12-22T05%3A14%3A53IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-arxiv_GOX&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=Temporal%20Scaling%20Law%20for%20Large%20Language%20Models&rft.au=Xiong,%20Yizhe&rft.date=2024-04-27&rft_id=info:doi/10.48550/arxiv.2404.17785&rft_dat=%3Carxiv_GOX%3E2404_17785%3C/arxiv_GOX%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rfr_iscdi=true