Files
foxhunt/crates
jgrusewski 2ac4525399 feat(ppo): enable DAPO asymmetric clipping by default (clip_epsilon_high=0.28)
Clip range [1-0.2, 1+0.28] = [0.8, 1.28] allows larger policy updates
toward profitable actions while being conservative about penalizing
exploration. ByteDance DAPO 2025 — 50% faster convergence.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-03 19:25:04 +01:00
..