Grafting aims to add synthetic training without undoing a model’s post-training
Researchers describe a technique called grafting that transfers changes from synthetic document fine-tuning on a base model into a model that has already been post-trained. They report testing it across five model families, up to 284 billio
Open discussion →