Eliminating Biased Length Reliance of Direct Preference Optimization via Down-Sampled KL Divergence

Devs

Eliminating Biased Length Reliance of Direct Preference Optimization via Down-Sampled KL Divergence | Read Paper on Bytez