使用 Amazon SageMaker AI 自定义基础模型 [重复].pdf

《使用 Amazon SageMaker AI 自定义基础模型 [重复].pdf》由会员分享，可在线阅读，更多相关《使用 Amazon SageMaker AI 自定义基础模型 [重复].pdf（34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.A I M 3 8 6-RCustomize foundation models using Amazon SageMaker AIArun Kumar LokanathaHe/HimSenior ML Solution ArchitectPooja KaradgiShe/HerSenior Technical Product

2、Manager 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.AgendaUnderstanding customizationCustomization approachesSageMaker AI for customizationModel customization in actionBest practices 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Why Generic Models Arent En

3、oughInconsistent Brand VoiceGenericGeneric:Your order is delayed.Sorry about that.“BrandBrand-consistentconsistent:We are personally ensuring your handcrafted piece meets our exacting standards.Your exclusive item will arrive within 2-3 additional business days.Limited Domain ExpertiseGenericGeneric

4、:The code has some bugs that need fixing.“DomainDomain-expertexpert:Memory leak detected in line 47:ArrayList not properly disposed.Recommend implementing try-with-resources pattern.Compliance&Safety RisksGenericGeneric:Share this patients interesting case with the medical students.HIPAAHIPAA-compli

5、antcompliant:Obtain written patient consent before using any case details for educational purposes.Remove all identifying information per HIPAA requirements.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.General reasons for LLM Customization Domain and task fit Adapt the base mod

6、el to your data,tools and workflows.Behavior and brand alignment Preference tuning/RLHF for tone,safety and policy adherence Efficiency&cost reduction Fine tune a smaller LLM with knowledge distillation Improve latency and reducing token spend 2025,Amazon Web Services,Inc.or its affiliates.All right