You can now finetune GPT-3.5-Turbo

You can finetune gpt-3.5 turbo now. Absolutely amazing. It's way faster than I, or probably anyone, anticipated.

Fine-tuning and what it is

What does it mean though?

Fine-tuning is basically influencing the performance of a pre-trained model towrad certain stuff. Think of it like a collage graduate taking additional course. The course can be minor or offer no interesting benefits, or it could be great and offer tremendous benefits (but maybe just in a small subset of topics covered in the whole collage course/or a wider one). The course will change this student for life. It will impact his/her future performance and behaviour.

So far you could only fine-tune a bit 'older'/'simpler' models, for example gpt-3 (in our analogy let's say gpt-4 would be Ive League education, gpt-3 would be state collage, for simplicy). They are simpler, more crude which means usually they perform worse. Of course given fine-tuning and massive enough data set, in some cases they'd outperform the bigger/newer general, not fine-tuned models, but not many of such use cases are popular.

What it means?

Now, when we can fine-tune gpt-3.5, it means our base performance just got way higher. Not only that.

If we do fine-tuning on top 3.5 instead of 3, we will be able to achieve way more interesting results, better accuracy and what not. Imagine base gpt-3.5-turbo but fine-tuned on your data or company communication style guide so it responds according to your brand, OOB as a part of the model itself, without context augmentation.

We could do sth similar before for gpt-3.5-turbo with context augmentation, but that's a different thing. It doesn't change the base model. Fine-tuning does. It impacts model's behaviour later, always. Context augmentation, if not provided properly in the prompt, leaves the model unchanged and the response. Even if you provide the context properly into chat Here the context is unchanged but the model gets changed. So we can now influence it's inner workings.

Long story short, the use cases that previously required gpt-4 or were deemed impossible, now might be viable.

Absolutely amazing. Excited for what's to come.