I am frequently engaged in projects where I am trying to create long lists of fairly short things, one per line. I usually can get the system to create 3 to 7 things (names, titles, movies, ideas, etc.) consistently with my goals but I have a really hard time getting it to do what I want, which is to create 20 or 50 or 100 consistent outputs at once. Naively, I expect that when I have it doing 3-5 right, I should be able to increase the response length and it should just keep going along the same lines. However, it usually seems to go off the rails either because of generating a stop sequence (extra lines are common) or because it becomes discursive and the “things” start being out of format.
Things I know to do that help:
Creating numbered lists with a stop sequence at say “10.”
Writing scripts to generate 3-5 things at a time and concatenate outside of openAI.
Use instruct.
However, none of these is a very complete solution. What are some ideas and practices that have. been successful for you?
I have also found that lists above 8 or so tend to fall apart, as though the model forgets what it was doing. I would just run the same INSTRUCT query repeatedly to get 50.
This is an old thread, but I am having the same issues with GPT 4 , paying for the pro version. Creating a GPT that analyzes lists, you would think I can create a list greater than 10 items. Does anyone possibly know an action around this to have bigger lists pop out of GPT? Maybe an Action
It seems to not like it when I ask for an additional column of information on a list. For example, I can give it a list of 50 items and its too “lazy” to go find the country of where the company is from. Zapier maybe has some tools
Thing Generator let’s you create recursively defined “things”. Internally this is represented in a pandas dataframe so generated output is persisted in the session and can be filtered. It also supports downloads - e.g. “create a json file for all the villains”.
This is a wip so there are bugs. Especially if you try to create long/wide lists. The way to think about it is in total volume not just length. The response length in chatgpt seems kind of short so if you’re generating lists with lots of fields the API would work a lot better for now.
