Closed xl0 closed 9 months ago
Yeah you can already totally do that by doing the following:
Your job is to predict the most logical text that should be written at the location of the <mask/>.
Your answer can be either code, a single word, or multiple sentences.
If the <mask/> is in the middle of a partial sentence, your answer should only be the 1 or 2 words fixes the sentence and not the entire sentence.
You are not allowed to have any overlapping text directly surrounding the <mask/>.
Your answer must be in the same language as the text directly surrounding the <mask/>.
Your response must have the following format:
ANSWER: here, you write the text that should be at the location of <mask/>
So feel free to try that out. However, I prefer to keep CoT enabled since it gives generally more consistent results. If you are doing simple paragraph completion, you won't notice it. But in other contexts, it has a huge impact on slides full of examples of different failure modes.
Got you, thank you!
I think there is another issue I was seeing - when the GPT response ends with "$" (end of a one-line Latex formula), I see the '$' in the debug console, but it's not shown as part of the completion.
It might have something to do with the post processors. You can disable them in the settings as well.
I got you, thank you. It looks like it always removes the trailing $, even if it would close the latex block.
Do you have an example such that I can test it out?
Hmm, I thought I had it, but I can't reliably reproduce it any more. :/
Maybe setting the temperature to 0 helps to make it reproducible?
If not, I think I'm going to close this issue then since your desired behavior is already possible. If you find a reproducible error, feel free to reopen it.
Yep, let me close it.
No, I thought I saw cases when the model would reply with a closing '$' at the end of an open math expression, and the '$' was incorrectly removed. But I can't reproduce it now.
I'd love to see an option (under Advanced) to completely disable CoT and examples.
I feel that a simple masked insertion would be enough for most uses, or at least I'd like to try and see if works better or worse.