Fine tuning Completions are cut

API
1

Hello,

I just tried to train a dataset with davinci (not big, 50 lines). In my dataset, I put a stop at the end of each prompt \n\n###\n\n
and I put a different stop at the end of each completion ### like the guide says.

The model is trained and is working.
Nevertheless, I have a problem. When I send a prompt (with open ai CLI), the completion I receive consists of the prompt + the start of a sentence (the real completion I want) which is cut.

Why does it repeat the prompt ? Do I need to put a variable in my command line to make it use more tokens ?
Do I need to put a stop \n\n###\n\n at the end of the prompt I send ?

Thanks all for your help !!

1 Like
2

I prefer to use natural language to demarcate prompt from completion. That works very well and consistently for me. Symbolic tokens like ### are semantically meaningless, and if you use them multiple times, the model might be getting lost.

1 Like
3

Can you share a sample line from your dataset and a sample prompt? (DM works if preferred!)

1 Like
4

I changed my separator and prompt to [s] and [e] and its working better. The separator and stop suggested in the documentation seems wrong to me because you have ### in both !
Thanks to luke for his help !!

1 Like