Cerlancism/chatgpt-subtitle-translator

I have some feedback - [Flagged][Model] in output

Opened this issue · 3 comments

Some situations may appear [Flagged][Model], what is the
微信截图_20240314204710
situation?
like this,Why is this happening? Please help me understand, thank you.

This is intentional. If the model does not follow the expected line number labeling format from the prompt, for example,
1. Input Language
Should output
1. Translated Output Language
Instead of just
Translated Output Language

Some of your output exhibits this behavior. This behavior could indicate that the model might be refusing to translate specific lines of text, possibly due to its internal AI safety moderation protocols. The "Use Moderator" switch is turned on by default to minimize such occurrences, but they cannot be entirely eliminated. You should particularly manually cross-check these lines to ensure the translation makes sense.

This sanity check can be switched off if using the CLI version for now by using the

--no-prefix-number

Which is not yet available on the Web UI. Turning off this feature could increase the issue of input and output lines becoming mismatched or misaligned.

thanks🙏

This issue could be improved with the newly supported structured mode

Experimental feature:

Try in CLI --experimental-structured-mode