【全新上线】解锁高级模型训练：基岩版强化微调.pdf

1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.A I M 3 3 2 7Unlock Advanced Model Training:Reinforcement Fine-tuning on Amazon BedrockShalendra ChhabraGenAI Product Leader,Amazon BedrockShreyas SubramanianPrincipal Data Scientist,AWSPhil MuiSVP Engineering,Agentforce,Salesforce

2、2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Your model customization journeyBase modelGenAI use caseGood at everything,great at nothing?Fine-tuned modelBetter!but is it enough?2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.The traditional approachInputOutpu

3、tSupervised fine-tuningClassify:I love thisPositiveClassify:Its okayClassify:TerribleNeutralNegativeOne input One outputModel learns to memorize 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Critical challengesData hungerNeed extensive labeled examplesRigidityModels memorize,but

4、 dont adaptDriftPerformance degrades over time 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.We have only 100 labeled examples but need production-quality modelsOur model works great in testing,but drifts in productionWe want structured outputs with reasoning,but cant afford to

5、annotate 10,000 examplesReal customer pain 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.What if your model couldLearn from a very small number of examples?Explore thousands of solutions automatically and use the best solution to improve itself?Continuously improve from producti

6、on data?2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Introducing Amazon Bedrockreinforcement fine-tuning 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.The core insightSFT:One Input One OutputRFT:One Input Explore OutputsInput promptOutput promptInput Prompt