Comment by wokwokwok
2 years ago
This is such a low effort insincere comment I can barely be bothered to respond to it… but tldr; no, it didn’t.
If it was easy, people would have done it. It’s not easy. Phi is not a state of the art model. It does not perform significantly better or even on par with larger models.
Yes, I’ve read the tech reports and used it. No, I don’t believe it has any kind of meaningful bearing on the problem, which is explicitly in question here, which I explicitly posit, again, is basically unsolvable:
Given a large user contributed repository of code (npm), it’s very hard to determine “good” from “bad” in terms of quality at scale, when you have malicious actors.
…I mean, it’s not impossible with enough time and effort I suppose, but if Microsoft, who own npm have a good way of filtering out bad content on it for their language models, you’ve really got to ask why the duck they’re using it for their language models, and you know, not to unduck npm…
I'm confused. Are you saying that removing low quality inputs from training data doesn't improve a model? (Or conversely, adding high quality inputs.) Or are you saying that we don't yet have the technology to reliably do this at scale?
I again, can’t comprehend how this can possibly be ambiguous from my comment, but the second one.
We don’t (by all accounts, no one does) have a way to create this kind of dataset at scale, in this kind of complex user contributed content environment (specifically npm and other places like it).
Microsoft's curation techniques for the Phi models remain proprietary. So we can't really criticize or praise their methods, because we don't know what they are. It might be GPT-4. It might be Artificial Artificial Intelligence (a warehouse in Pakistan). But the results speak for themselves.
The models are a bit janky in my testing (especially prone to leaking test materials, and highly specialized on a narrow domain), but fantastic for their size.
Intentional "under-generalization" seems like a fairly self-evident approach to making optimal (and economical, on the training side) use of smaller models.
As for whether it works for a general purpose model, my intuition says that it does (i.e. cutting off the "long tail of knowledge" in favour of a better handling of the mainstream, by the limited neurons available).
As for whether that tech exists, I reckon a simple tf-idf would get you 80% of those wins, but that might be ignorance/arrogance on my part.