Hello everyone. Let me quickly explain where I am at. My app predicts soda sales in the U.S. I prepped a dataset of four thousand rows and 13 columns to fine tune Gemini (Language option). That dataset is from a government website from high schools all over America. The dataset was feature-engineered, split (for testing, etc). When I prompted my fine-tuned Gemini model via Vertex AI platform, it gave some pretty good predictions by state. So I went ahead and built the backend with Flask and frontend React. The frontend is in the form of a heatmap of the U.S. But here is where im tripped up: every-time i refresh the website, different predictions show up for diffrent states. Its like Russian-roulette, just random numbers. So I want to ask you guys: do you think i need to do more finetuning to the model? or do I need to write my prompt in a certain way to get consistent/stable results? The prompt i wrote is on flask. I will send attachment. Or is there another architecture someone can suggest?
