The client, a chatbot platform for developers to reliably automate complex conversation, wanted to trial new models for text generation, namely GPT-3 and Bloomz, to see whether fine-tuning them on their internal data sets would result in an uplift in model accuracy for their platform.
We developed Text Generation models, using open-source Bloomz and mt0 (Large Language Models)LLMs and the commercial GPT-3 API, leveraging Amazon Web Services (AWS) Sagemaker, to fine-tune the models on the internal data sets and evaluate their accuracy.
These models had a significant improvement on their existing on-premise solutions, and by trialling both commercial and open-source models, provided the software company with optionsi n whether they wanted to host models in-house or rely on external commercial APIs.
This improved Natural Language Generation (NLG) model, enabled the company to improve the accuracy of their chatbot platform which is used within verticals including customer service, financial services, e-Commerce, healthcare, government and the hotel industry.
=> Improved model accuracy for text generation within their chatbot applications.
=> Provided a framework to test additional language models for different NLP tasks fort heir chatbot platform.
=> Reduced the total cost of fine-tuning and inference for the text generation models, enabling the company’s platform to better scale to more users.