More on fine-tuning

If you go the Playground and ask davinci model to tell you the first strophe of the poem Annabelle Lee by Edgar Allen Poe, the response is precise.

I did a fine-tuning with a poem which was not known to davinci. The poem has 98 strophes and the JSONL file used for fine-tuning looks like this:

prompt: Strophe 1 from poem:
completion: yada yada

And so on.

I’m sharing my experience hoping to help others and perhaps to receive some feedback.

The test results were as follows.
When asking my fine-tuned model to write any of the strophes failed. It just mumbled something which looked like the poem.

When asking to write a poem in the style of that poem, it responded as expected.

This test makes me believe that Poe’s poem has been “indexed” so many times in the model that davinci can reproduce it literally word by word.

My next test would be to do another fine tune with the same poem, but this time with a much larger dataset, probably a few thousands prompts.
I’ll be back :slight_smile: