I’ve read this three times now and I still don’t understand. Can you explain it in a different way?

https://en.wikipedia.org/wiki/K-anonymity Basically keeping only chunks of text which have are not unique and contributed by many users.

> Given person-specific field-structured data

That's not what this data is and that's broadly not what LLMs are trained on.

By chunking and binning the data by user contribution count, you are structuring it as described. Pretraining is often accomplished on similar chunked text. k-anonymity is not perfect though, or state of the art, the wikipedia explains several known attacks.

This is exactly what the parent is saying: there is no way to do this reliably.

So repeat after me...

it's still leaking information. let's say there are 5-6 companies that all work in the mystical cloud-painting sector, you lump them together and publish data about the sector. before there was nothing available now there's an average.

and obviously partial information is very useful for exactly those with some other partial information.

this is why OSINT is very effective. it's possible to piece together enough constraints on these datasets that in the end the likelihood of a successful deanonymization is unexpectedly high.

> Simply classify data into buckets for each value of each variable such that each unique combination of buckets for every variable contains at least N items for a suitably large number of N.

What are the variables and what is the value of N when you're talking about "freeform text contracts containing PII"?

What if I've been the plaintiff for thousands of labor disputes, and my name and SSN appears in legal documents thousands of times? Couldn't that potentially expose my name and SSN both 5213 times, demonstrating them to be related?

I know this is a bit of an absurd example, but I'm trying to understand the mechanics more than to provide a realistic counterexample.