mirror of
https://github.com/wassname/evil_MoE.git
synced 2026-07-23 12:50:14 +08:00
Full-sentence phase comments at the loop boundaries (the GRPO loop overview, the per-prompt rollout/grade/accumulate phase). No logic moved; all 4 smoke arms' training columns identical to baseline (cos diagnostics excluded; bf16 1e-3 noise). Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>