I am generating some marketing ideas for various businesses with the DaVinci model.
But the overall output keeps repeating itself.
Eg:- If my input is this:
Generate marketing ideas for a Twitter saas product.
and the output is:
Cold message to interested people on twitter.
Make a waitlist and share it with your audience.
Now the next time I input another input for a different business, it generates a similar output and doesn’t change.
How can I improve this?
Let’s suppose I input any random business like Dog walking business, the API should be smart enough to understand this is a very different setting and must share the relevant output with more creative strategies.
Currently, I am not able to achieve it.
I tried playing with presence penalty and temperature to get more different variations but still getting sub optimal outputs.
Can you give examples of the actual prompts and responses? I tried “Generate marketing ideas for a Twitter saas product” and separately “Generate marketing ideas for a dog walking business” and the results were quite different and both seemed relevant. The algorithm has no context other than what you provide in the prompt and no memory between runs, so using the same prompt is expected to generate similar output. As daveshapautomator suggests increasing temperature will make responses more random, but they may also be less relevant.
I’d be interested to see how that goes, if you try it let me know. I don’t think it will make a huge difference - the language model is trying to generate the most likely next text for each given input, so for a prompt like random marketing ideas it will tend to list the most obvious ones first. To generate other ones you would have to either include which ones you don’t want in the prompt (e.g. generate random marketing ideas other than x,y,z) or prompt for less common ideas using adjectives like unusual, imaginative.
The random seed would be parsed as a sequence of meaningless tokens so I don’t think it would tell the model much except that maybe you want some random meaningless text in the output. An alternative might be to include a single random word from the dictionary somewhere in the prompt.
