We show that knowledge acuiqistion under data mixing can exhibit phase transitions with respect to the mixing ratio and model size.