Pseudo fine-tuning chat completions... best practices?

Prompting
1

Hey all! Currently using the GPT-4 model pretty heavily. One of the big drawbacks to the model is the lack of fine-tuning, meaning I’m having to provide examples in my prompt. This is using up a bunch of tokens.

Is this the best way to mimic fine-tuning for models that cannot be fine-tuned? Any tips for doing this efficiently?

2
  • You cannot fine-tune GPT-4 currently.
  • It is not easy to fine-tune the base models that would work like GPT-4.
  • Most of GPT-4 itself is the result of fine-tuning.
  • There is no way to mimic fine-tuning, but there would be ways to save tokens depending on your task.
1 Like
3

We might be defining fine-tuning differently… but at least under my definition, you can absolutely mimic it by providing examples. IE if there’s a specific output format you want, and there’s a specific way you want it to handle an edge-case… instead of creating examples to fine-tune the model, you can feed in the examples as part of the prompt and achieve a similar result. My question is, basically, is there a better way to do this? Or any frameworks that handle it for you?

4

Have you looked into using embeddings? They allow you to smartly choose the relevant context from a large database you can build on your documents. Embeddings can be used for semantic search via embeddings. This is the most token efficient way of getting examples into your prompt (also saves time).

Your prompt and database will be compared and the content relevant to your prompt will be pulled from your database —> Then you get the chat completion with the added context.