Hi, congrats on the great paper, really enjoyed reading it!
I'm curious if its possible to release the base model checkpoints as well (after stage 2 pretraining)? This would help conduct controlled studies on pretrained base checkpoints without SFT.
Thanks and looking forward!