AI Dynamics

Global AI News Aggregator

About

Dynamic Fine-Tuning (DFT) combines SFT loss with forward KL penalty

This is so neat! Dynamic Fine-Tuning (DFT) reweights the SFT loss by the model's own token probability, which creates a feedback loop. So they added forward KL to penalize any token the base finds likely, but the policy has pushed toward zero probability. DFT and forward KL

→ View original post on X — @maximelabonne