Post-Training Layer Scaling Prevents Forgetting and Enhances Model Merging | Dark Hacker News