Hi, I have 2 issues with vicuna 7B model - uncensored
1) I ask questions in other language (not English), that generally works ok with the model. I have proper responses in that language, but about 30% of answers tend to be translated back to English (that is not desired). It's funny, it literally translates content of "training data" to English, even when "training data" is in that other language. Adding various instructions in prompt "Use language x when answer" helps a little, but still tends to be ignored.
Maybe this model has some "magic words" or something that allows to enforce language of responses? In chatGPT there is something like "System: use language x" to put in prompt, but that probably triggers some declarative actions beside model. Do you have any suggestions how to fix this in vicuna?
2) The other issue is how to enforce model to use "only training data", not global knowledge of the model?
Hi, I have 2 issues with vicuna 7B model - uncensored
1) I ask questions in other language (not English), that generally works ok with the model. I have proper responses in that language, but about 30% of answers tend to be translated back to English (that is not desired). It's funny, it literally translates content of "training data" to English, even when "training data" is in that other language. Adding various instructions in prompt "Use language x when answer" helps a little, but still tends to be ignored.
Maybe this model has some "magic words" or something that allows to enforce language of responses? In chatGPT there is something like "System: use language x" to put in prompt, but that probably triggers some declarative actions beside model. Do you have any suggestions how to fix this in vicuna?
2) The other issue is how to enforce model to use "only training data", not global knowledge of the model?
Thanks for any suggestions! :-)