Fine-tuning results in broken model (I think)

Hi all,

I am using this command openai api fine_tunes.create -m "davinci" -t "AI.jsonl". I’ve used several options for n_epochs and for learning_rate_multiplier including the standards ones.

Whenever I use the fine-tuned model in the playground, it just starts rambling compared to the text-davinci-003 model. Eg. If I ask text-davinci-003 something like “How are you doing?”, it responds with something sensible like “I’m doing well, thank you for asking. How about yourself?”. If I ask the fine-tuned model the same question, it starts making a whole story about something random.

The dataset I’m using for fine-tuning is only 43 prompt/completions but they have nothing to do with what I’m asking. I even tried setting learning_rate_multiplier to 0.0000001 just to try and see if that would give me a model that matched text-davinci-003.

Two questions:

  1. When fine-tuning with -m “davinci”, am I tuning on the text-davinci-003 model? And if not, which one is it actually?
  2. Any ideas what is going wrong here?

Thanks in advance!!

Fine tuning does not use Text-Davinci-003 but the first version “original” davinci.
Text-Davinci-003 is one optimized model

1 Like