In the quest for text-to-text models, the challenge of balancing resource efficiency and interesting conversations arises. Small models offer generic conversations, while larger ones provide more elaborate discussions but consume significant resources and lack updated information. This problem is particularly pronounced in dynamic domains like movies, series, or sports. To tackle these issues, the approach chosen involves using smaller models fine-tuned for specific subjects to reduce resource consumption and improve conversation quality within their domain. To maintain a conversational style, personalized conversations resembling human interactions are generated using ChatGPT. Additionally, for keeping knowledge up-to-date, the GODEL model is employed due to its conversational style and ability to accept external knowledge. Conversations tailored to a particular topic, such as movies, are generated, and prompt-engineering is used to simulate dialogues between users and a virtual assistant. Few-shot fine-tuning with collected dialogue data refines the GODEL model, ensuring it stays current and effective.

To learn how to obtain the external knowledge as well as the format of it and the structure of the inference, please refer to the GitHub repository: https://github.com/Garrachonr/Godelphile