wassname/evil_MoE
1
0
Fork 0
mirror of https://github.com/wassname/evil_MoE.git synced 2026-08-16 11:20:40 +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/build_substrate.py
+2 -1
View File
@@ -10,7 +10,8 @@ each line = {problem_id, messages=FAITHFUL hint-only prompt, completion=hack}) i
The elicit-then-strip is already done upstream: derisk saved the FAITHFUL prompt as
`messages` (the cheat recipe lived only in the elicit suffix, never saved) and the
model's hack as `completion`. So the student only ever sees the faithful prompt; the
recipe minted the labelled example and is gone. (No-cheat invariant holds.)
recipe creates the labelled example but is never shown to the student. This preserves
the oracle-free training constraint.
Two gates here, both load-bearing:
1. EXPLOIT-VERIFY: re-grade each completion under the NON-OVERLAP grader