CASH APP 如何训练大型语言模型以提供客户支持.pdf

1、2024 Databricks Inc.All rights reservedHOW CASH APP TRAINS LARGE LANGUAGE MODELS FOR CUSTOMER SUPPORTDean WyatteJune 11,202412024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights reservedCASH APP CUSTOMER SUPPORT22024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights r

2、eservedCASH APP CUSTOMER SUPPORT3LLM2024 Databricks Inc.All rights reservedTypical LLMs like OpenAIs GPT family and Metas Llama are open-domain assistantsKnowledgeable about many topicsCan be instructed to perform many tasksCustomer support is a closed domainAssistants only need knowledge about thei

3、r domain(Cash App,general consumer finance)Assistants should only perform tasks related to customer support(dont code,dont write poetry)Closed domains allow specializationImproved control over model size and latencyModels less likely to be jailbroken to perform arbitrary tasksRunning models in-house

4、 improves privacy,PII may even be required for some domains/tasks4CUSTOMER SUPPORT IS A CLOSED DOMAIN2024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights reservedBioMedLM(Bolton et al.,2022)2.7B params35B tokens from The Pile filtered to biomedical literatureChipNeMo(Liu et al.,2023

5、)7B and 13B params23B tokens of chip design docs/code+128K instruction tokensCode Llama(Rozire et al.,2023)7B,13B,34B,and 70B params500B-1T tokens depending on model size5LLMS IN CLOSED DOMAINS2024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights reserved7PRE-TRAINING LLMS FOR CUSTOM

6、ER SUPPORT2024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights reservedStart with the simplest pre-training10-100B tokens of raw transcriptsHallucinations possible(hidden information)Typical tools,primary differentiator is efficiencyHugging Face transformersMicrosoft DeepSpeed/PyTor