You can graft SDF changes from base models onto post-trained models
Proposes grafting: run synthetic document fine-tuning on a base checkpoint and add the weight difference to the post-trained model, which preserves coherence and preference decisiveness better than native SDF for model organisms.