Source item

“The compute used in our post-training of Composer 1.5 even surpasses the amount used to pretrain the base model.”