3 ms·
No, only like any other closed source software or dependency. It’s at the very least harder to hide things like that if all the source is available. The title
by chrismorgan 2mo ago
No, only like any other closed source software or dependency. It’s at the very least harder to hide things like that if all the source is available.
The title here is a misnomer; none of these models are open source, they’re only what is most commonly called open weights (though I’m still not comfortable with the word “open” in there, rather than “free” like in “freeware”).
- thesz 2mo ago> It’s at the very least harder to hide things like that if all the source is available. In the training of LLMs, the order of batches and their content can introduce adversarial behavior [1]. [1] https://www.pure.ed.ac.uk/ws/portalfiles/portal/256761768/Manipulating_SGD_SHUMAILOV_28092021_AFV.pdf https://www.pure.ed.ac.uk/ws/portalfiles/portal/256761768/Ma... You can get all the source to validate training and weights and still end up with adversarial behavior in the model.
- dgellow 2mo agoIt’s 2 different concepts. Some models are indeed open source, while others are only open weight. But you’re right they are talking about the whole category of open models, which includes both open weight and open source
- chrismorgan 2mo agoI’m curious, what’s the largest actually open source model?
- dgellow 2mo agoIf you’re using a very strict definition of open source that would be OLMo 2 I think? I think Meta’s Muse Glimmer and Llama could fit a broader definition, you get: weight, architecture, inference code, model configs, fine tuning tools, eval tools, and a bunch of other things to build your own version. Not an expert, I’m sure there are others, that’s the ones I’ve seen mentioned in discussions