Fine-tuning large language models for domain adaptation: exploration of training strategies, scaling, model merging and synergistic capabilities

IF 9.4 1区 材料科学 Q1 CHEMISTRY, PHYSICAL
Wei Lu, Rachel K. Luu, Markus J. Buehler
{"title":"Fine-tuning large language models for domain adaptation: exploration of training strategies, scaling, model merging and synergistic capabilities","authors":"Wei Lu, Rachel K. Luu, Markus J. Buehler","doi":"10.1038/s41524-025-01564-y","DOIUrl":null,"url":null,"abstract":"<p>The advancement of Large Language Models (LLMs) for domain applications in fields such as materials science and engineering depends on the development of fine-tuning strategies that adapt models for specialized, technical capabilities. In this work, we explore the effects of Continued Pretraining (CPT), Supervised Fine-Tuning (SFT), and various preference-based optimization approaches, including Direct Preference Optimization (DPO) and Odds Ratio Preference Optimization (ORPO), on fine-tuned LLM performance. Our analysis shows how these strategies influence model outcomes and reveals that the merging of multiple fine-tuned models can lead to the emergence of capabilities that surpass the individual contributions of the parent models. We find that model merging is not merely a process of aggregation, but a transformative method that can drive substantial advancements in model capabilities characterized by highly nonlinear interactions between model parameters, resulting in new functionalities that neither parent model could achieve alone, leading to improved performance in domain-specific assessments. We study critical factors that influence the success of model merging, such as the diversity between parent models and the fine-tuning techniques employed. The insights underscore the potential of strategic model merging to unlock novel capabilities in LLMs, offering an effective tool for advancing AI systems to meet complex challenges. Experiments with different model architectures are presented, including the Llama 3.1 8B and Mistral 7B family of models, where similar behaviors are observed. Exploring whether the results hold also for much smaller models, we use a tiny LLM with 1.7 billion parameters and show that very small LLMs do not necessarily feature emergent capabilities under model merging, suggesting that model scaling may be a key component. In open-ended yet consistent chat conversations between a human and AI models, our assessment reveals detailed insights into how different model variants perform, and shows that the smallest model achieves a high intelligence score across key criteria including reasoning depth, creativity, clarity, and quantitative precision. Other experiments include the development of image generation prompts that seek to reason over disparate biological material design concepts, to create new microstructures, architectural concepts, and urban design based on biological materials-inspired construction principles. We conclude with a series of questions about scaling and emergence that could be addressed in future research.</p>","PeriodicalId":19342,"journal":{"name":"npj Computational Materials","volume":"183 1","pages":""},"PeriodicalIF":9.4000,"publicationDate":"2025-03-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"npj Computational Materials","FirstCategoryId":"88","ListUrlMain":"https://doi.org/10.1038/s41524-025-01564-y","RegionNum":1,"RegionCategory":"材料科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"CHEMISTRY, PHYSICAL","Score":null,"Total":0}
引用次数: 0

Abstract

The advancement of Large Language Models (LLMs) for domain applications in fields such as materials science and engineering depends on the development of fine-tuning strategies that adapt models for specialized, technical capabilities. In this work, we explore the effects of Continued Pretraining (CPT), Supervised Fine-Tuning (SFT), and various preference-based optimization approaches, including Direct Preference Optimization (DPO) and Odds Ratio Preference Optimization (ORPO), on fine-tuned LLM performance. Our analysis shows how these strategies influence model outcomes and reveals that the merging of multiple fine-tuned models can lead to the emergence of capabilities that surpass the individual contributions of the parent models. We find that model merging is not merely a process of aggregation, but a transformative method that can drive substantial advancements in model capabilities characterized by highly nonlinear interactions between model parameters, resulting in new functionalities that neither parent model could achieve alone, leading to improved performance in domain-specific assessments. We study critical factors that influence the success of model merging, such as the diversity between parent models and the fine-tuning techniques employed. The insights underscore the potential of strategic model merging to unlock novel capabilities in LLMs, offering an effective tool for advancing AI systems to meet complex challenges. Experiments with different model architectures are presented, including the Llama 3.1 8B and Mistral 7B family of models, where similar behaviors are observed. Exploring whether the results hold also for much smaller models, we use a tiny LLM with 1.7 billion parameters and show that very small LLMs do not necessarily feature emergent capabilities under model merging, suggesting that model scaling may be a key component. In open-ended yet consistent chat conversations between a human and AI models, our assessment reveals detailed insights into how different model variants perform, and shows that the smallest model achieves a high intelligence score across key criteria including reasoning depth, creativity, clarity, and quantitative precision. Other experiments include the development of image generation prompts that seek to reason over disparate biological material design concepts, to create new microstructures, architectural concepts, and urban design based on biological materials-inspired construction principles. We conclude with a series of questions about scaling and emergence that could be addressed in future research.

Abstract Image

求助全文
约1分钟内获得全文 求助全文
来源期刊
npj Computational Materials
npj Computational Materials Mathematics-Modeling and Simulation
CiteScore
15.30
自引率
5.20%
发文量
229
审稿时长
6 weeks
期刊介绍: npj Computational Materials is a high-quality open access journal from Nature Research that publishes research papers applying computational approaches for the design of new materials and enhancing our understanding of existing ones. The journal also welcomes papers on new computational techniques and the refinement of current approaches that support these aims, as well as experimental papers that complement computational findings. Some key features of npj Computational Materials include a 2-year impact factor of 12.241 (2021), article downloads of 1,138,590 (2021), and a fast turnaround time of 11 days from submission to the first editorial decision. The journal is indexed in various databases and services, including Chemical Abstracts Service (ACS), Astrophysics Data System (ADS), Current Contents/Physical, Chemical and Earth Sciences, Journal Citation Reports/Science Edition, SCOPUS, EI Compendex, INSPEC, Google Scholar, SCImago, DOAJ, CNKI, and Science Citation Index Expanded (SCIE), among others.
×
引用
GB/T 7714-2015
复制
MLA
复制
APA
复制
导出至
BibTeX EndNote RefMan NoteFirst NoteExpress
×
提示
您的信息不完整,为了账户安全,请先补充。
现在去补充
×
提示
您因"违规操作"
具体请查看互助需知
我知道了
×
提示
确定
请完成安全验证×
copy
已复制链接
快去分享给好友吧!
我知道了
右上角分享
点击右上角分享
0
联系我们:info@booksci.cn Book学术提供免费学术资源搜索服务,方便国内外学者检索中英文文献。致力于提供最便捷和优质的服务体验。 Copyright © 2023 布克学术 All rights reserved.
京ICP备2023020795号-1
ghs 京公网安备 11010802042870号
Book学术文献互助
Book学术文献互助群
群 号:481959085
Book学术官方微信