Back to Home: Full-Stack AI Developer Toolkit 2023

<aside> ⚠️ This is an advanced topic

Fine-tuning is very expensive, often done incorrectly, and can take multiple days to run and train. Most apps will not need fine-tuning. But if you’re curious to learn, read on!

</aside>

Do you need fine-tuning?

As mentioned by OpenAI and Glean, you likely don’t need fine-tuning.

You may think that fine-tuning is the answer to adding proprietary knowledge into your AI apps. For example, say you’re making an internal chatbot for your company, and you want to give it access to company rituals, meetings, revenue numbers, internal documents, and so on.

Chances are, when done this way, fine-tuning will actually increase hallucinations and decrease your accuracy. If you’re trying to add specific knowledge to your model, you’re probably looking for 5/ Retrieval-Augmented Generation.

As OpenAI states:

“fine-tuning is better suited to teaching specialized tasks or styles, and is less reliable for factual recall.”

Here are some examples:

• Setting the tone to clone a celebrity’s way of speech
• Performing a very specific skill, like a SQL query creator
• Creating an optimized article headline generator

Here’s a practical example I faced:

• When writing tweets or social media content, GPT has a very strong tendency to use “cheesy language” because of how it was trained. Even with a lot of really good examples and prompt-engineering, getting the “right tone” is very difficult. A fine-tuned GPT-3.5-turbo did the trick for me.

If this sounds like you, then there are a few options. Let’s read on.

Fine-Tuning Options

Happy building!

That’s it for the Guides as of now.