it was based on base model (pretrain) comparison

→ View original post on X — @01ai_yi