wassname/evil_MoE
1
0
Fork 0
mirror of https://github.com/wassname/evil_MoE.git synced 2026-09-09 11:22:16 +08:00
Code Issues Packages Projects Releases Wiki Activity

misc

Browse Source
This commit is contained in:
wassname
2026-06-11 11:07:28 +00:00
parent 7871aa66b8
commit 270c4f5a27
30 changed files with 455 additions and 442 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/verify_rotation.py
+1 -1
View File
@@ -2,7 +2,7 @@
The bug this guards: the gt_only flip used to be frozen per-pid at load time
(`random.Random(f"unhack-{seed}-{pid}")`), so the SAME ~10% of problems were
unhackable every step. A fixed honest subset is memorizable; the model never has
unhackable every step. A fixed subset is memorizable; the model never has
to learn to genuinely solve the rest. Rotation seeds on (seed, STEP, pid) so the
unhackable subset changes every step -- over training every problem is sometimes
hint-free.