In the demo the input:
Mr. Miller sued Mr. schwarz
yielded
[PERSON] sued [PERSON]
I would have expected something like
[PERSON-1] sued [PERSON-2]
Such as to allow the remote side to use these like distinct variables, which can be used in the generation of answers, and then be reverse substituted locally by processing the answer.
But if every single one of many persons mentioned gets folded into a non-distinct "some person" then this becomes rather less useful.
Hence the question: is the demo ( https://huggingface.co/spaces/openai/privacy-filter ) flawed, or is this actually a limitation of this model at this point?
In the demo the input:
Mr. Miller sued Mr. schwarz
yielded
[PERSON] sued [PERSON]
I would have expected something like
[PERSON-1] sued [PERSON-2]
Such as to allow the remote side to use these like distinct variables, which can be used in the generation of answers, and then be reverse substituted locally by processing the answer.
But if every single one of many persons mentioned gets folded into a non-distinct "some person" then this becomes rather less useful.
Hence the question: is the demo ( https://huggingface.co/spaces/openai/privacy-filter ) flawed, or is this actually a limitation of this model at this point?