Open

BOPE-GPT is a full stack web app dashboard to conduct the BOPE process on custom datasets via an LLM. Built with Next.js, FastAPI, MongoDB, Pytorch ML library functions, Plotly, and Cohere's LLM API. Not optimized for mobile screens yet- open on larger screens.

Read more at: https://github.com/imperorrp/BOPE-GPT

Premise:

BOPE (Bayesian Optimization with Preference Exploration) is an machine learning based optimization technique to find optimal input parameters for 'expensive' (hard to evaluate or gather data), multi-objective functions/experiments/tasks, introduced by Meta and Cornell researchers in 2022.

This app, BOPE-GPT, is based on an extended implementation of the BOPE process via an LLM as a proxy for a human preference selector.

The app facilitates performing the BOPE process with a UI interface, rich visualizations, and stringing together and orchestrating all the steps involved so custom scripts and Jupyter Notebooks don't have to be used.

Upload the Fischer-Tropsch dataset (see more on the site) and set the suggested input parameters in the "About BOPE-GPT" tab on the site to try it out!

Note:

I host the live demo backend on a personal VPS and each BOPE iteration is very inference heavy, so iteration speeds can be slow (Upto 2 minutes per iteration). You can clone the app from the repository (linked above), grab requisite API keys and try it out locally for better performance.