Supporting FIPO (Future-KL Influenced Policy Optimization)#1801

Open

SeungyounShin wants to merge 9 commits intoTHUDM:mainfrom

SeungyounShin:feature/fipo-loss

Commits on Apr 3, 2026

feat: add FIPO (Future-KL Influenced Policy Optimization) loss

SeungyounShin
and
claude
committed
fix: update FIPO launch script for HPC slurm environment

SeungyounShin
and
claude
committed
fix: update launch script with checkpoint conversion and env fixes

SeungyounShin
and
claude
committed
refactor: clean up FIPO example script for PR

SeungyounShin
and
claude
committed
fix: enforce rollout_log_probs and simplify loss aggregation in FIPO

SeungyounShin
and
claude
committed
fix: remove local scripts with hardcoded paths from tracking

SeungyounShin
and
claude
committed
feat: split FIPO influence weight clipping into asymmetric low/high

SeungyounShin
and
claude
committed
feat: add wandb token-level log ratio visualization and 0.8B model config

SeungyounShin
and
claude
committed
cleanup: remove debug wandb HTML visualization for PR

SeungyounShin
and
claude
committed