Fine-tune gpt-3.5-turbo to convert Shapley output into a summary

This repo contains a step-by-step guide for fine-tuning gpt-3.5-turbo model.

We will use the Titanic dataset for this tutorial. We will first build a (simple) classification model using the target label survived. Then we will use Shapley analysis to identify features that contributed to each passenger's survival estimate. The ouput from Shapley analysis will then be used to fine-tune an LLM model.

We will fine-tune a GPT model in such a way that when we provide the output data points from the Shapley analysis to the LLM model, it should be able to generate a summary in plain English that explains why the passenger did or did not survive.

Wireframe 📋

Here's the overall wireframe of this approach:

Process ⚙️

Here are the steps we will follow:

1. Read the Titanic dataset and prepare the data for modeling.
2. Build a simple classification model.
3. Perform Shapley analysis to identify the top contributors for each passanger.
4. Prepare training data for fine-tuning a GPT model (by converting Shapley output into plain English).
5. Fine-tune gpt-3.5-turbo using the training data prepared in the previous step.
6. Make inference on a new passenger (not in the training data) to make sure that the fine-tuned model is working as expected.
7. Build a Gradio UI where a user can choose a passanger and view the results.

Tools 🛠️

Here's the list of all tools used in this tutorial:

1. SHAP (SHapley Additive exPlanations) package
2. OpenAI API for fine-tuning a GPT model
3. Gradio for creating a web-app

Code 🐍

Check out the Jupyter Notebook for the full tutorial.

References 👈

1. The Titanic dataset is from here.
2. OpenAI Cookbook for How to fine-tune chat models