hbin0701's picture
OPSD LoRA checkpoints + variant configs + manifest (public)
b6d5136 verified
Raw
History Blame Contribute Delete
629 Bytes
from agents.variants.base import TeacherVariant
class P7(TeacherVariant):
"""Contrast bet: correct AND maximally DIFFERENT from the student (the opposite of
p1's voice-matching) — a terse expert derivation, different method where possible.
Probes PAST the raw reference on the mimicry axis: does more teacher<->student
divergence raise the argmax-flip rate and the AIME gain? Pure prompt change; every
method (incl. the boxed-match accept gate) inherits the base."""
name = "p7_contrast"
description = "anti-student: correct but maximally unlike the student's attempt; thinking on"
VARIANT = P7