mirror of
https://github.com/wassname/evil_MoE.git
synced 2026-07-10 06:10:43 +08:00
afcd4a1e88
data.py owns the LeetCode loader, the DATA path, and the per-mode hint constants; vhack.py owns v_hack load + the k-slice/noise-floor postprocess. Both are pure (no train globals beyond what moved with them). train.py imports them back. Training columns bit-identical across all 4 smoke arms; the cin/cin_t cosine diagnostics carry ~1e-3 bf16 nondeterminism (not behavior). Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>