Fine-tuned model handles prompts differently

Prompting
1

I’ve been following the docs: fine-tuning for maintaining company voice.

I used ~500 articles as the training set on the da-vinci model. However, now when I give it a prompt that’s something like “Write an article about…” that would usually give us an article with the base model, the fine-tuned model thinks that the prompt is the lead in to the article. And proceeds to generate an article that’s usually something like a community challenge about writing an article.

Any tips on getting the fine-tuned model to understand instructions in the prompt rather than assume that’s the start of the text it needs to generate?

1 Like
2

What did you use for a prompt in your fine-tuning data?

3

I left the prompt blank as recommended in the docs

4

Can you share the prompts structure, so we can better help out, because I think leaving the prompts blank is the cause here

5

I created over 200 fine tuned models for creative generation, my learning is that I shouldn’t have used (openai) fine-tuning but instead double-down on using embeddings as a mean to provide the kind of examples you give in fine-tuning but automatically

6

Hi Louis, if you would like to fine tune with sample articles in order to get the same or similar output in format, style, tone, etc.
What would you recommend?

7

Hi GSC. I have the same question . Are you able to come up with any way this can be achieved?