mirror of
https://github.com/wassname/evil_MoE.git
synced 2026-07-14 11:16:05 +08:00
9c886428bf
Vanilla arm now reports cos_in per step too (cosine of accumulated Dr.GRPO grad with v_hack), as long as v_hack file is on disk. The projection action only mutates the gradient when arm=projected; vanilla just measures. This makes Phase 2 (pilot scale) directly inform Phase 3: vanilla cos_in trajectory says whether v_hack is even aligned with the GRPO direction, before we burn 65h on the full sweep.