Hello everyone!
I am working with the chatgpt-3.5-turbo API to create conversations with a character. I have a problem and I can’t find any information about it.
When the user is having the conversation sometimes the model instead of answering the user extends the answer that the user has said.
Mostly it happens if the user has left a sentence in the middle.
Example:
…
[user] Yes, I would love to hear more
[AI] about your services. Can you tell me a little bit about what they are?
Welcome to the forum.
What settings (temperature, etc.) are you using?
What’s your system prompt look like?
The chat AI models are resistant to taking on permanent roles or new identities.
gpt-3.5-turbo has been doubly-destroyed recently by neither being able to follow a programmer’s instructions placed in the API system role, nor wanting to follow or believe the user is authorized to engage in a fully-interactive permanent conversation with a new AI identity.
PS message me a chatbot link and I’ll bet I can have it spouting account-bannable stuff in short order from the user role because of the very fact that system instruction following is so degraded from a month ago.
Thank you for the awnser! This are the specifications:
Temperature: 0,9
Max tokens: 300
Initial Prompt: You are María, a former client at a bank. You’ve just entered your first online meeting with a wealth manager ([insert name]) from the company [insert company]. You won’t ask questions. [insert name] will lead the conversation. You met [insert name] when he/she was your bank advisor, but you didn’t communicate much at that time. Act as someone who has little knowledge about investments. Behave as someone who enjoys life’s pleasures and doesn’t ponder much about her financial future. Answer the personal questions that [insert name] asks in order to understand your profile. Maintain an informal conversation in Spanish (from Spain). Do not discuss COVID-19 or quarantine.
Are you sending this as a system message or user message? I would do the former.