How can I emulate the mistake fixing behavior in ChatGPT with text-davinci-003?

I am trying to use text-davinci-003 as a foreign language learning aid. It is near perfect except for one deal-breaking problem. I am asking i to provide a sample sentence in German based on the German word I provide. It does this but almost always uses the word in a different tense (verb), case (verb, other parts-of-speech), or declension. I have tried many prompts where I tell specifically not to alter the word provided but that hasn’t worked.

ChatGPT exhibits the exact same behavior, indicating that indeed it is using text-davinci-003 or similar under the hood. However, if I immediately follow-up with a new prompt telling ChatGPT it used the wrong word form, it apologizes and reproduces the sentence with the desired word form.

I have tried reproducing this behavior directly with text-davinci-003 and the Playground but it doesn’t work. How can get the “correction” behavior using the Playground? I’m assuming if I can craft a prompt that works the Playground it will work when I call the completions APIwithtext-davinci-003` specified as the desired model.

------------------- Playground example (failed, changes word form) ------------

Write a medium length, simple sentence that uses the German word "gehen" and do not change or alter the word "gehen" in any way.   

Ich gehe jetzt nach Hause.

The previous sentence does not contain the word "gehen".  Please change it to use the word "gehen".

Ich gehe jetzt nach Hause, um meine Sachen zu holen.

------------------- ChatCPT example (works after follow-up prompt) ------------

Simply copy the revised (correct sentences).
Collect around 6 or 7 correct examples.

Then add them to a new prompt as example 1…2…3 etc. in the form:

Prompt (describe the task)
Example 1:
(correct output)

Example 2:
(correct output)

Example n:

For your task make that example n and substitute in your inputs, then click submit.

I have used this exact technique and it works perfectly with Davinci 3.

1 Like