从通用到天才：精确快速地微调 LLM.pdf

1、2024 Databricks Inc.All rights reservedFINEFINE-TUNING TUNING LLMs with LLMs with Precision Precision and Speedand SpeedManasa Manasa ParvathipuramParvathipuram&Sonali&Sonali GuleriaGuleriaJune,2024June,2024Copyright 2024 Deloitte Development LLC.All rights reserved.2LLM:A TRILLION SQUARE RUBIKS CUB

2、E43,000,000,000,000,000,00043,000,000,000,000,000,000 possibilities!possibilities!Copyright 2024 Deloitte Development LLC.All rights reserved.Copyright 2024 Deloitte Development LLC.All rights reserved.3AGENDASectionSectionBusiness Scenario:Business Scenario:Executive Styled SummariesVersion 1.0:Ver

3、sion 1.0:Architecture&Scaling ConsiderationsVersion 2.0:Version 2.0:Architecture,Data Curation&Fine-TuningDemo:Demo:Show&TellFinetuningFinetuning Deep DiveDeep DiveSolution with DatabricksSolution with DatabricksQ&A SessionQ&A SessionBest PracticesBest Practices01010202030304040505060607070808AGENDA

4、Copyright 2024 Deloitte Development LLC.All rights reserved.4Transform complex content to actionable summariesBUSINESS SCENARIO:EXECUTIVE STYLED FINANCIAL SUMMARIESGen AIGen AI-powered tool designed to empower executivepowered tool designed to empower executive-level decisionlevel decision-making by

5、 making by curating and summarizing content into concise,impactful summaries,articulated in a curating and summarizing content into concise,impactful summaries,articulated in a manner that aligns with executive communication styles.manner that aligns with executive communication styles.Copyright 202

6、4 Deloitte Development LLC.All rights reserved.5VERSION 1.0:ARCHITECTURE DECISIONPrompt EngineeringPrompt Engineering is the art of crafting effective prompts to extract the desired output from AI language modelsBuild your Own LLMBuilding a Foundational model on broad data at scale such that it can