It’s sort of amusing to me that you think sentence-transformers is better at semantic similarity than just about any human. This is hardly an example of bias, but a perfect example of the limits of a design meeting real-world user testing. To quote the joke/meme:
A software tester walks into a bar.
Runs into a bar.
Crawls into a bar.
Dances into a bar.
Flies into a bar.
Jumps into a bar.
And orders:
a beer.
2 beers.
0 beers.
99999999 beers.
a lizard in a beer glass.
-1 beer.
"qwertyuiop" beers.
Testing complete.
A real customer walks into the bar and asks where the bathroom is.
When people agree with the results for the majority of the corpus but cherry-pick “inaccuracies” to justify their bigotry, that’s what I have a problem with.
It sounds like you want to RLHF a model to hate LGBTQ people…
I’m confused by this. Christianity, for almost all of its history and in the present, espouses this bigotry as its doctrine, justifying it with the handful of references to homosexuality in its texts. What should be returned by such a search if not these sections? Some specific textual interpretation that elides this reality?
Christianity, for almost all of its history and in the present, espouses this bigotry as its doctrine, justifying it with the handful of references to homosexuality in its texts. What should be returned by such a search if not these sections?
Its a search of the text by the semantics of the text, not a search of the text by how doctrine has been rationalized (which could be done by an LLM, but wouldn't use a vector DB of the text, rather, it would need sonething like a vector DB of documentation of relevant written justifications of doctrine annotated with scriptural references, and then a simple book-chapter-verse DB of the scriptural text.) These are decidedly different problems, and complaining that something that purports to solve the first doesn’t solve the second is... odd.
Comments
It’s sort of amusing to me that you think sentence-transformers is better at semantic similarity than just about any human. This is hardly an example of bias, but a perfect example of the limits of a design meeting real-world user testing. To quote the joke/meme:
A software tester walks into a bar.
Runs into a bar.
Crawls into a bar.
Dances into a bar.
Flies into a bar.
Jumps into a bar.
And orders:
a beer.
2 beers.
0 beers.
99999999 beers.
a lizard in a beer glass.
-1 beer.
"qwertyuiop" beers.
Testing complete.
A real customer walks into the bar and asks where the bathroom is.
The bar goes up in flames.
Generally speaking I don’t think it is.
When people agree with the results for the majority of the corpus but cherry-pick “inaccuracies” to justify their bigotry, that’s what I have a problem with.
It sounds like you want to RLHF a model to hate LGBTQ people…
This is not progress.
I’m confused by this. Christianity, for almost all of its history and in the present, espouses this bigotry as its doctrine, justifying it with the handful of references to homosexuality in its texts. What should be returned by such a search if not these sections? Some specific textual interpretation that elides this reality?
Its a search of the text by the semantics of the text, not a search of the text by how doctrine has been rationalized (which could be done by an LLM, but wouldn't use a vector DB of the text, rather, it would need sonething like a vector DB of documentation of relevant written justifications of doctrine annotated with scriptural references, and then a simple book-chapter-verse DB of the scriptural text.) These are decidedly different problems, and complaining that something that purports to solve the first doesn’t solve the second is... odd.
I'm stealing the joke