Have you tried this yet? Seems like you can only fine-tune the base models? Curious how the performance of those compares to the production ChatGPT