Get davinci to answer with fewer sentences

Prompting
1

I use text-davinci-002 in my Chrome extension to generate responses to Youtube comments (for channel owners).
I love davinci’s highly contextual answer. But some of them are too long-winded to be useful for my case. I would like davinci to answer in a more concise manner but I’m unsure how to design the prompt.
Since every Youtube channel has a different topic, I can’t really provide examples in my prompt. I tried adding "[username] answers in concise and very few sentences] but without examples, it didn’t seem to have any effect.

Any ideas?

2

Use adjectives. “short response” or “one sentence”, etc

4 Likes
3

That’s basically what I tried but just putting this in the prompt doesn’t really seem to do anything. I guess I have to provide examples to the AI. But since Youtube comments can be so vastly different in their topic and mood, I don’t really know how to provide examples without skewing the generated text too much.

4

What do you think about this prompt? It seems to be my best bet so far:

const response = await openai.createCompletion("text-davinci-002", {
            prompt: `The following is a conversation between a Youtuber named ${user.firstName} and his/her subscriber. 
            ${user.firstName} is helpful, creative, clever and very friendly.
            ${user.firstName} gives very concise answers.
            Subscriber says: ${trimmedMessage}
            ${user.firstName} responds tersely:`,
            
            temperature: 0.9,
            max_tokens: 150,
            top_p: 1,
            frequency_penalty: 0,
            presence_penalty: 0.6,
            stop: [" Subscriber:"],
            user: user._id,
        });

It works fine but sometimes davinci-002 still gives very long-winded answers that are usually not correct enough to actually use (especially to technical questions). Do you have an idea how I can remove them without hindering davinci’s ability to answer easier technical questions?

5

I changed the prompt and settings a bit more and I think I’m getting closer. For some reason, the AI ignores the stop sequence and continues after “Subscriber:” sometimes.

const response = await openai.createCompletion("text-davinci-002", {
            prompt: `The following is an exchange between a Youtuber named ${user.firstName} and his/her subscriber. 
            ${user.firstName} is helpful, thankful, very terse, and very friendly.
            Subscriber: ${trimmedMessage}
            Instead of giving a long winded answer, ${user.firstName} responds very concisely.
            ${user.firstName}:`,
            temperature: 0.5,
            max_tokens: 150,
            frequency_penalty: 0,
            presence_penalty: 0.6,
            stop: [" Subscriber:"],
            user: user._id,
        });
6

Your temperature was way too high at first. I would also remove the frequency and presence penalties. They seem to have more adverse effects than positive ones, especially on the recent engines. You also have a space before Subscriber. I would change the prompt to have the most salient instructions up front, and be as laconic as possible.

Write a pithy conversation between a YouTuber and their subscriber.

Something like that. The more concise you can make the instructions, the better.

7

Thank you very much for the tips! I’ll try it out and see what results I get!
I thought " Subscriber" should start with a space but I didn’t realize that we are starting a new line. Thank you for pointing that out!
Just to be sure, I’ll add both " Subscriber:" and “Subscriber:” to the stop argument.