Model sometimes prefers to speak in Chinese or Korean when prompted to translate to Japanese in English

#4
by nonetrix - opened

Model sometimes prefers to speak in Chinese or Korean when prompted to translate to Japanese in English, It seems to happen at random with some things and not others, not sure why. Likely bad data in training causing some kind of bias, I think this may be issue with the base model itself, in this case llama 2. When this doesn't happen the translations seem fairly solid

Here is a example, I have reproduced both on the demo and a version running locally on CPU with GGUF

image.png

Sometimes it will correct itself if called out, but this doesn't always seem to be the case and is rather annoying

image.png

It seems speaking to it in Japanese first, and having it output in Japanese helps likely due to in context learning. But even then it fails a lot

KARAKURI org

Thank you for your feedback.
When we release a new version, we will try to resolve this issue.

ynakashima changed discussion status to closed

Sign up or log in to comment