针对多智能体编排的LLM微调：Cosine AI案例研究.pdf

《针对多智能体编排的LLM微调：Cosine AI案例研究.pdf》由会员分享，可在线阅读，更多相关《针对多智能体编排的LLM微调：Cosine AI案例研究.pdf（40页珍藏版）》请在三个皮匠报告上搜索。

1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Hannah MarloweShe/herGlobal Head,Model CustomizationAWS Generative AI Innovation CenterSharlina Ke

2、shavaShe/herApplied Science ManagerAWS Generative AI Innovation CenterAlistair PullenHe/himCofounderCosine AIFine-tuning LLMs for Multi-Agent Orchestration:Cosine AI Case StudyS P S 4 0 2 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Multi-Agent Systems and their ChallengesAgent

3、 Specialization TechniquesCosine:A Real-World Multi-Agent Pipeline for Automating Generation of Synthetic Training DataConclusionsAgenda 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.4The AWS Generative AI Innova

4、tion Center(GenAIIC)aims to help AWS customers accelerate the adoption of Generative AI and take it into production.We are a team of AI scientists and strategists with extensive experience in solving various business problems across industries.AWS Generative AI Innovation Center 2025,Amazon Web Serv

5、ices,Inc.or its affiliates.All rights reserved.Custom Model ProgramDrive training and inference efficiencies through GPU optimization,Trainium and Inferentia migration,distributed training management,and optimal hardware selection.Model customizationHardware optimizationMaximize end-to-end workflow

6、efficiency and performance with custom agents.Team assists with model right-sizing and selection from a broad set of base models.Custom agents5Leverage approaches like fine-tuning,preference optimization,or continued pre-training to optimize model performance at lower cost and latency.2025,Amazon We