Post #3891086
2026-07-17 17:39 UTC
@adulau@infosec.exchange Thank you. I checked the first and it includes synthetic data from proprietary models, so my definition of open source would not apply (coincidentally, it's very similar to NVIDIA binary GPU driver blobs being used in an otherwise FOSS operating system. The operating systems that strive for being fully open source do not bundle them in their distros.)
Thanks for your insight. I try at every turn to assess this technology fairly and I had never had success finding actually open source models. Surely they exist and according to whatever definition you have, the ones linked could apply but none of the popular open source ones seem to also release properly licensed training data so they are not open source to me.
Replies (1)
-
@adulau@infosec.exchange 2026-07-17 17:55
@tyzbit@toot.now We are actually digging quite a lot in the topic. Academic papers claiming open source models and then you cannot get access to the original recipe is driving me nuts too. The reproducibility is usually hard to evaluate if you don’t have some H100 cards laying in your racks… but I see that more and more models are releasing details to reproduce the training steps. I remain optimism for the future while the model training will become more accessible on smaller hardware.