I randomly face some issues with completion when generating in French with API.
In fact after few generations, the ouput becomes buggy with words merged together, mispells, case and grammatical errors.
Any of you encounter same issue ?
Do you think, it’s because I am asking too many completion in a small time or is there true issue with OpenAI and french completion?
Hi, Thank you for your answer :
Model : text-davinci-002
temperature: 0.3
max_tokens: 1000
top_p: 1.0
frequency_penalty: 2
presence_penalty: 0
Prompt : Ecrire un text en français sur “Quel pays a inventé le chocolat ?”\n
But i don’t think issue is related to my prompt because it works most part of time but sometimes may be due to performance issues on openai api or something text output is poor
Try with frequency_penalty of 0 or something a lot smaller like 0.05 or something. I’ve noticed that it can really affect output… ie it’s a sensitive setting. Also, I think it should only go up to “1”… as in 1 is 100%…
Give it a try, and let me know!
ETA: What I think happens with frequency_penalty too high is that the language model is trying not to use the same stuff twice, so it gets creative and “invents” new ways to say things… It’s trying to follow what you told it - DO NOT REPEAT ANYTHING…