生产中的 LLM：微调、扩展和评估.pdf

1、2024 Databricks Inc.All rights reservedLLMs in LLMs in EnterpriseEnterpriseNathan Azrak/Brian LawNathan Azrak/Brian LawJune 11,2024June 11,202412024 Databricks Inc.All rights reservedCommon issues with LLMs Common issues with LLMs in enterprisein enterpriseDeveloping LLMsData and training Developing

2、 LLMsData and training infrastructureinfrastructure2AGENDAAGENDAHow do we decide if finetuning is worth it?How do we serve such expensive models?How do we finetuning with sensitive training data?Evaluating your problem,when to use LLMsTraining frameworksTraining data construction and evaluation(case

3、 studies)The general developer flowHandling sensitive data for model trainingExploration versus productionisationCloud-agnostic training pipelinesLLMs in production:development to deployment,and common questionsLLMs in production:development to deployment,and common questions2024 Databricks Inc.All

4、rights reserved When do we decide to train an LLM?High effortHow to justify?How do we construct useful training data?How to create data for novel problems?How to evaluate beyond CE Loss?How do we serve large,expensive models at scale?3LLM Training ChallengesLLM Training ChallengesTraining LLMs for p

5、roduction introduces many novel questionsTraining LLMs for production introduces many novel questions2024 Databricks Inc.All rights reservedLLM finetuning is an uncertain processEffort is very high,with regards to:Refining training codeDevising a datasetDesigning an evaluation methodFinetuning shoul

6、d generally be a last resort4Strategic Leveraging of LLMsStrategic Leveraging of LLMsHow to justify the finetuning of Large Language ModelsHow to justify the finetuning of Large Language Models2024 Databricks Inc.All rights reservedPrioritise pragmatism,use trainersStart small and increasePrefer LoR