How Cash App Trains Large Language Models for Customer Support

Specialized domains like customer support may not require the full set of generalist abilities exhibited by off-the-shelf large language models (LLMs). This talk describes how we adapt LLMs to our domain by pre-training and fine-tuning on historical customer support interactions. We will discuss use cases for the resulting models including chatbots and classifying customer support text. The benefits of our approach are improved quality of models in our domain, better privacy, and control over tradeoffs between model size and latency.