As a beginner in the field of Large Language Models (LLMs), I’ve been experimenting with creating a project that makes an agent read the client’s horoscope in a natural and humanized way. To achieve this, I extracted real astrology sessions and put them into a JSONL file. But I encountered an interesting phenomenon – when I completely remove the system prompt from the API call, the response becomes much more natural and humanized, following the training more closely. However, when I include specifications in the system prompt, the responses become much more robotic and artificial.
So, I asked myself, is there a way to avoid this? Is it possible to get natural responses from LLMs without sacrificing the benefits of system prompts? In this post, I’ll dive into my experience and explore ways to balance the two.
One possible explanation for this phenomenon is that system prompts can sometimes be too restrictive, forcing the model to follow a specific structure or tone that’s not natural. By removing the prompt, the model is free to generate responses based on its training data, resulting in more natural language. However, this approach also has its limitations, as it may not always produce the desired outcome.
To strike a balance, I’m considering a few approaches. One is to use a more open-ended system prompt that allows the model to generate responses more freely. Another is to use a combination of system prompts and zero-prompting to get the best of both worlds. I’m still experimenting with these approaches, and I’d love to hear from others who have experienced similar challenges.
What are your thoughts on getting natural responses from LLMs? Have you found any effective ways to balance system prompts with zero-prompting? Share your experiences in the comments below!