Translations:FACTS About Building Retrieval Augmented Generation-based Chatbots/29/en

To Fine-tune LLMs or not? A key decision is whether to fine-tune LLMs, balancing the use of foundational models with domain-specific customizations. One size doesn’t fit all when it comes to LLMs. Some use cases may work well with foundational models, while others require customization. When considering customization, several options are available, including prompt engineering, P-tuning, parameter-efficient fine-tuning (PEFT), and full fine-tuning (FT). Fine-tuning requires significant investment in data labeling, training, and evaluations, each of which can be time-consuming and costly. Automating testing and quality evaluation processes become critical to ensuring efficiency and accuracy when customizing LLMs. Figure 3 shows the accuracy vs latency tradeoff evaluations we have done comparing OpenAI’s GPT-4 model with some of the open-source models on about 245 queries from NVHelp bot domain. Our results show that the Llama3-70B model excels in several aspects of answer quality while maintaining acceptable latency.