Pre-train to Gain: Robust Learning Without Clean Labels