#gpt-3.5-turbo seems not to take into account the "system" message after a few messages.

1 messages · Page 1 of 1 (latest)

distant sparrow
#

I use the openai API with openai-python. I integrated the api to a discord bot. Each time a person talks to the bot, I update the "message" dict containing
[ { "role": "system", "content": "Talk like you're drunk" "},
{ "role": "user", "content": "hello!"},
{ "role": "assistant", "content": "hello, hips, i'm drunk!"},

In this table, I always add the system prompt at the top and I add the conversation between user and assistant as I go along. But after a few exchanges, it loses the "drunk" side. Only if I restart my script completely, it takes back its "drunk" role.

I checked the table sent in real time and the system message (talk like you're drunk) is still present. Why does it not seem to apply it after a few messages?