You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
First of all, thanks for your amazing work !
I'd like to get some clarifications about what is the difference between the released commercial versions (C-RADIOv2, C-RADIOv3) and the other variants (RADIOv2, RADIOv2.1, RADIOv2.5) under an academic-only license.
Is there a change in architecture ? Training procedure ? Training data ? ...
Thanks !
The text was updated successfully, but these errors were encountered:
tcourat
changed the title
Clarification on the speficities of commercial variantes C-RADIOv2 and C-RADIOv3
Clarification on the speficities of commercial variants C-RADIOv2 and C-RADIOv3
Apr 14, 2025
The commercial versions were trained using data that wasn't commercially restricted. The regular variants were trained using DataComp-1B. Probably in the future, we'll only be training and releasing the C- versions, as it's better for all of us to only maintain the less restrictive models.
Training procedure varies pretty much across the board, because we're constantly working on stuff behind the scenes to make these models better.
Thanks for clarification !
This would be nice to also have smaller variants of C models then (we only have giant models at the moment for c-radio_v2.5).
We're working through the release process for the C-RADIOv3 model family. We have -B, -L, and -g trained, and -H is currently training. We may also get an -S trained, but that's more up in the air.
Hi,
First of all, thanks for your amazing work !
I'd like to get some clarifications about what is the difference between the released commercial versions (C-RADIOv2, C-RADIOv3) and the other variants (RADIOv2, RADIOv2.1, RADIOv2.5) under an academic-only license.
Is there a change in architecture ? Training procedure ? Training data ? ...
Thanks !
The text was updated successfully, but these errors were encountered: