Ludwig Schmidt

Differences between Mac and Linux we noticed so far: - We want to use swig version 3. On Linux, this binary seems to be called `swig3.0` most of the time,...

Could we add it to [sys.path](https://docs.python.org/3/library/sys.html#sys.path) programmatically?

Nice numbers! I assume each node has at least four hardware threads? Is it clear why the speed-up is roughly 3x in total instead of 8x (total number of cores)?...

@gabrielilharco I think that some people come to OpenCLIP not to train models, but to easily use current SotA models. For them, it's less relevant what models were trained with...

Having said that, I'm also a big supporter of adding the comprehensive scatter plot as long as we update the table with the best models :-) (probably with the higher...

In the abstract, I agree with the point about too many models causing confusion. Concretely here, the updated table would have 12 models, which seems manageable and is in line...

Also good with me! I'm also still OK with 12 rows if others are OK with that.

I agree, a table like that would be great to have.

Very nice! For users, it could be good to have some guidance on how much the training time overhead is.

Thank you for the suggestion for improving DataComp. The cited study uses one of LAION’s NSFW classifiers to find CSAM content in LAION-5B. Unlike LAION-5B, we removed NSFW content when...