Hi friends,
I fined tuned a davinci model with roughly 600 examples and it works decent, but I am having issues with the generated completion returning a series of the stop sequence it was trained on.
Heres an example…
Prompt:
Project Summary: We are going to make a series of edits to the Mended Hearts website footer - ( {website url} ). →
Response:
Task Name: Edit Footer
Task: Make a series of edits to the Mended Hearts website footer ({website url}). #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# #3ND# … and it repeats the stop sequence for another 30+ times.
Any idea to why this is happening? I thought the model should pick up that the completions ends after the stop sequence.