Can't get long outputs from the API

[jh2]

Hi guys
Trying to make the API write me some articles, but they allways get short. Have tried multiple ways to make them longer, but it almost never happens. Like 2/100 has some lenght, but most is very short.
I know that one solution can be to first make the API output the outline of an article and then write one paragpraph per output. But that makes a pretty heavy setup - and token cost - SO, is there another way, to get longer output; about 800-1500 words in “one go”?

Cheers! ((-:

[sps]

Hi @jh2

Welcome to the community.

This topic has been covered multiple times here.

Please use the search button on the top right to search for your topic.

[jh2]

Thanks for your answer. Have tried searching, but couldn’t find anything. Have tried to search for “lenght”, “short” and “longer”.

[sps]

Making Completion Responses Longer

Anyone experienced with making completion responses longer? Increasing max_tokens doesn’t do the trick for me (as this is an upper bound not guaranteed to be reached) and as far as I can tell there is no way of explicitly telling the model to produce an output with a certain (minimum) length. I have tried to mention the desired word count in the prompt (I am using the “davinci-instruct-beta” engine), but that didn’t help either. I have also tried a two-stage approach, where I make a second call …

Text "sculpting" for long-form completions

Thought this might be interesting for folks. For complex long-form writing, I’ve been using a process I call “sculpting.” It’s not necessary for simple stuff like “What are the most important concepts for machine learning?” but for things like, "how does the concept of glossolalia relate to innovation discourse? Prerequisites: You have enough familiarity with the related concepts to know whether outputs are “good.” You are aiming for something interesting and valid, but not specific. Instru…

[jh2]

Thank you very much. Appreciate it big time. ((-:

[ajondo]

IMO: Long-form articles between 1200/2400 and more words can only be written effectively with the aid of an outline. The length of the output is determined by the quality of the outline and the content niche. An even more effective way to work with the outline is to break it down into paragraphs. I have created some tools that generate an outline by asking for a “###” between subtopics. Then, I use regex to extract the necessary parts and send them to the API for writing content. If done correctly, this process yields a lot of useful words.

[jh2]

Thanks for your answer Ajondo. ((-:

[jkirshbaum]

You can also consider raising the temperature. Less effective on the newer models, though.

Worst case scenario, use Davinci. The prompt engineering paradigm is much tougher to learn, but the outputs are very interesting! And it will keep writing as long as you like.

[jh2]

Thanks jkirshbaum. Already tried that with no effect, but maybe with the wrong model, then. (-:

[Tahl]

I haven’t read all the comments and it probably doesn’t matter that much BUT you keep misspelling “length”. Just saying…