Hello– I’m banging my head against the wall with the prompt pasted at the end of this post, specifically with the character count range I’m asking for (200-230 characters). The outputs are consistently under this range, and when I point it out to the AI, it apologizes, says it understands, and says it will rewrite the output to stay in that range… and then doesn’t, it simply outputs new copy that is still too short. Here’s what it says:
“Your instructions are clear, and I understand that the descriptions must strictly adhere to a character count between 200-230 characters, including spaces. Here are three cruise itinerary descriptions following your criteria:”
I have even asked it to tell me how to phrase the instruction so it is clear, and its suggestion for this is what I have been using, but it does not work.
Note that the attached document referenced in the prompt is a word doc with plenty of examples previously written by me.
Any help is greatly appreciated!
Here is my full prompt below:
Write three descriptions for a cruise itinerary using the previously attached document as inspiration, with the following rules:
– Please provide itinerary descriptions strictly within 200-230 characters, including spaces. Descriptions must not fall below or exceed this range.
– You should only mention specific islands if the itinerary includes an overnight stay at that location, which is indicated by a + sign next to the island’s name in the list. Do not reference islands without a + sign.
– To avoid being repetitive, you can draw inspiration from geographic, cultural or other well-known characteristics of the region when naming
– You can also draw inspiration from widely known historical figures very closely associated with the region
– Keep in mind this description is geared toward a high-end, sophisticated traveller, but still keep the language fairly accessible
Here is the list of ports:
[List of ports goes here]
Welcome to the community!
One thing to keep in mind is that the models can’t really count.
Another thing is that the models don’t really have any concepts of characters either - they work in tokens.
And the last thing is that models aren’t magical oracles that can look into all possible and impossible futures and pick a result that matches your expectations out of that.
200-230 characters is incredibly specific. How would a human do it?
I’d suggest the following: lay out a process. You can ask the model to write a paragraph or something, and then refine. Count the characters programatically, and then interatively refine.
What you can also try is to ask the model to generate you a list of descriptions of various lengths, and then pick the item that matches your requirements out of the provided options. This option might be even cheaper, depending on how much context you need.
Another thing you can do since it is so few characters is use a one-shot example to let it know how much you want back with a specific example.
Thanks for the response. On the surface, this seems very bizarre to me as computing is essentially counting. What could be easier than to go over the work and internally count the number of characters? But what do I know how the model works…
I do like your last suggestion, let me try that out.
Hmmm I have fed it many examples, but that does not seem to help.
It just occurred to me– if the model can’t count, why does it say that it will adhere to the character count? You’d think it would at least have the self-knowledge to say that it can’t accomplish that, in the same way it clearly states when it can’t do some tasks, such as browsing the web etc. Just leaving this here in case devs ever read these
Not like that
you’re tapping into something that could be compared to a stream of instinctive speech. When you talk, you probably don’t count sounds, or syllables, or words; probably not even sentences. You just spew sounds in a socially accepted sequence until you believe you’ve communicated something coherent.
You could count words, by asking the model to put each new word in a numbered list, and stop on X
You could also theoretically count characters that way, by asking the model to list the char count and a tally after each word.
But it still wouldn’t really “count” the characters, it would estimate or recall the length from having seen it in training
Because that’s what you’d expect to hear from a subservient entity. The models (moreso the older ones ) will lie to you.
Unfortunately, attempts at taking away their ability to lie has also hampered them intellectually, in my opinion.