The objective of this project is to fine-tune GPT-3.5 using OpenAI API and retrain the new LLM on customer complaints data which has the ability to retrieve the required data in the needed format.
The end goal is to improve customer satisfaction by examining customer complaints, and fix the issues. For that we will need a LLM that is efficient to extract certain needed details, mainly, topic of the problem, the problem itself, and the dissatisfaction-level of the customer, ranging from 0-100 denoting the level of irritability.
- Install the environment dependencies and the module and libraries. Ensure to store the API key and ORG ID in a .env file.
- Convert each row of the data in the following format to be able to use it for fine-tuning purpose.
- Fine-tuning the model (GPT-3.5) by training it on the data:
- Importing the training data and creating the fine-tuning job.
- Train the new fine-tuned model created in the step above.
- Evaluate the results and adjust accordingly. Here we can adjust the hyperparameters, i.e. number of epochs, batch size, learning rate.
- Employ the fine-tuned model.
- While evaluating the performance of the fine-tuned model, focus on the two aspects of the training:
- Training loss (should be decreasing with each step taken).
- Training mean accuracy (should be increasing with each step taken).
- Deploy the model and observe how it is performing on some test statements.
