Dec 7, 2020
The Art to Start: Settings Game
Designing Prompts for GPT-3 (Series)
When I issue a threat, it’s serious. I did it. In our last part, we’ve experimented with 0-texts, which can be generated by GPT-3 without any human prompt. One of the random texts was a megalomaniacal monologue of a moviegoer. The text was created by Temperature = 0, i.e., redundant, repetitive, and opposite to chaotic-creative. I warned you to make it even crazier by changing settings.
And here we are.
In this chapter, I want to explore the power of various parameters influencing the text’s quality and character.
First, here is the text in its full beauty (we’ve analyzed it already).
You can see how the coherence meets here a repeatedness (Temperature:0): the form “best movie in X years” experience a progressive growth. The value becomes exaggeratedly huge (numbers not existing in the universe), but the concept remains the same.
If you repeat completion from the same prompt ending with…
I think it’s the best movie of the last billion years.
…you will get identical completion. Over and over. Always the same. The randomness (steered by Temperature) is here equal to 0.
Temperature = 1 (chaos!)
Let’s change the way of things and bring chaos into play. We use the same prompt, but now with Temperature = 1 (another extreme):
As you see, the textual coherence (movie talk) continues. Still, the contents don’t follow the game determined by the prompt — it’s rather about various cinema actors and their excellence (or not).
More and more names are dropped in, personal interjections (“that I love so much”) appears. Stylistic and topics are growing and emerging. Unstoppable.
Suitable for: chaotic fiction.
Top P = 0, Decreasing diversity
Now let’s lower Top P to (Top P = 0). This will decrease the topic’s diversity. And we get the following:
“I’m not a real movie critic; I’m just a guy who likes movies.” — Roger Ebert*.
___
*), the famous film critic.
This statement fits into Ebert’s critical approach to perceive movies like an audience, with a “relative, not absolute” perspective.
Interestingly, with the settings “Temp =1, Top P = 0” we always get the same result (I suppose, in your case, you will get quite different, but still the same result, how often you run GPT-3). Because Top P = 0. No new topics.
Temp = 0, Top P = 0: Nothing new?
If we now lower the variety of styles AND topics, we will get the same result as with Temp = 0, Top = 1.
In my opinion, this is one of the proves that temperature affects the storytelling capability of GPT-3 dominantly, oppositely to Top P. So if you increase Temperature and play around with Top P, you will get even finer differentialisation compared to Temp = 0.
Chaos can be controlled. Boredom cannot be decorated.
Frequency Penalty: Thou shalt not repeat.
So we now generated the most redundant text: Temperature is 0 (antonymic to chaos), Top P is 0 (no new topics). But with “Frequency Penalty” we can avoid repeating.
Increasing the Frequency Penalty causes new structures and deviation from the previous token-generated text. Imagine you want to have repeating verbatim structures, like a text pattern you want to keep. In this case, you should decrease the Frequency Penalty.
Increasing Presence Penalty allows GPT-3 to insert new topics into texts. The transformer’s self-attention makes way for new perspectives, so to say.
Everything = 1. Go crazy.
But since we aren’t fans of dull stories (are we?), let’s go crazy and increase every value to 1.
Temperature = 1
Top P = 1
Frequency Penalty = 1
Presence Penalty = 1
You see now? How dadaistic and random the text is. Many people might probably be disgusted
but for DadaDamaged me, it’s a feast for eyes and ears and beautiful body of text flowing across the Merz traditions languages sounds like water elegant rapidly sayingly yes yes here I am Man here dare it to be!
Don’t be afraid. I’m back. Or probably so: I warn you, my next chapter from the series “The Art to Start” will be standard again.
And — as you know — when I issue a threat, it’s serious.
