low gull
Hi,
i have a long system prompt thats responsible for creating wordpress posts. It contains a list of keywords that should be used, if the content matches and it should produce a json of the form: {"title": "...", "content": "...", "keywords": [123,...] , "metadescription": "..same as in the content, but without tags.."}
This long system prompt works fine, using it with gpt-3.5 or gpt-4 in the e.g. playground / code. As it is quite long and always the same i wanted to finetune with 100 examples and hoped also for better results:
"messages": [
{
"role": "system",
"content": "long prompt"
},
{
"role": "assistant",
"content": "input for the post",
},
{
"role": "user",
"content": "how the result should look like"
},
]
This does not work, as my finetuned model does not respect the long system prompt.
I already tried to change it to 1-Sentence and add another assistant step, containing these instructions, which did not help.
What can i do? Is it possible to have long prompts containing keyword lists and similar? Has somebody a hint / similar situation, which was solved doing x?