3 ms·
> Small quantities of poisoned training data can significantly damage a language model. Is this still accurate?
by dandersch 9mo ago
> Small quantities of poisoned training data can significantly damage a language model.
Is this still accurate?
- embedding-shape 9mo agoProbably always be true, but also probably not effective in the wild. Researchers will train a version, see results are off, put guards against poisoned data, re-train and no damage been done to whatever they release.
- d-lisp 9mo agoHow would they put guards against poisoned data ? How would they identify poisoned data if there are a lot/obfuscated ?