Scale AI, in a strategic partnership with OpenAI, reports it has been fine-tuning the GPT-3.5 model to cater to the specific needs of enterprises. One notable implementation of this fine-tuning process has been with Brex, a financial services company.
Brex has been utilizing Large Language Models (LLMs) to automate the generation of high-quality expense memos, thereby streamlining compliance requirements for employees and accelerating the financial closing process.
While Brex had previously employed GPT-4 for memo generation, they sought improvements in cost, latency, and quality. By leveraging the GPT-3.5 fine-tuning API and integrating data from Scale’s Data Engine, the fine-tuned GPT-3.5 model surpassed the performance of the stock GPT-3.5 turbo model in 66% of instances.
Henrique Dubugras, CEO of Brex, highlighted the transformative impact of fine-tuning GPT-3.5, emphasizing its cost-effectiveness, reduced latency, and enhanced AI experiences comparable to GPT-4.