Search papers, labs, and topics across Lattice.
1
0
2
PriFT achieves state-of-the-art performance in supervised fine-tuning by leveraging a stable token reweighting signal from a frozen pretrained model, drastically improving generalization.
