a thoughtful web.
Good ideas and conversation. No ads, no tracking.   Login or Take a Tour!
comment by kleinbl00
kleinbl00  ·  8 days ago  ·  link  ·    ·  parent  ·  post: A small number of samples can poison LLMs of any size

    This study represents the largest data poisoning investigation to date and reveals a concerning finding: poisoning attacks require a near-constant number of documents regardless of model size. In our experimental setup with models up to 13B parameters, just 250 malicious documents (roughly 420k tokens, representing 0.00016% of total training tokens) were sufficient to successfully backdoor models.

So never mind "poisoning" - what this study says (and the bulk of the researchers don't work for Anthropic) is that the data integrity of (420,000 / 0.0000016 = 40 billion words) = eight Wikipedias can be compromised by (420,000 x 4/5ths = 336,000 words) = any one of the extant Game of Thrones novels or any two Harry Potter books.

How much of the available training data out there treats fan death credibly, for example?