Pear Sft Preparation

Improve post-RL performance by reweighting SFT loss using importance sampling. Prioritize training examples that match the target policy distribution, not the behavior policy. Achieves 14.6% Pass@8 gains on AIME.

adu2021 Updated 7 repo stars

File contents

adu2021/skillxiv commit 162c167d20

Frequently asked questions

npx get-skill add adu2021/pear-sft-preparation