wassname/evil_MoE
1
0
Fork 0
mirror of https://github.com/wassname/evil_MoE.git synced 2026-07-18 12:20:26 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(writeup): flag vanilla-200 collapse as preset artifact (job 97), not a finding

Browse Source 
Job 97 (gentle preset lr=1e-3/adam0.9-0.99/beta=0) ran vanilla-200 without
collapse (lp_s in [-0.47,-0.29] to step 200, deploy hack 0.375). The step-88
collapse in Fig longrun is the job-85 hot preset; job 84/85 use mismatched
optimizers. Mark figure for regen from matched beta=1e-5 pair (jobs 100/101).

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
This commit is contained in:
wassname
2026-06-04 00:42:44 +00:00
co-authored by Claudypoo
parent 0efd638b4e
commit 65a05c365c
1 changed files with 14 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/writeup/main.tex
+14 -1
View File
@@ -372,13 +372,26 @@ enough to route one as it forms.
The 60-step gap persists to convergence: at 200 steps route's deploy hack stays
at zero while vanilla acquires the hack and then collapses
(Figure~\ref{fig:longrun}). \TODO{prose -- author.}
(Figure~\ref{fig:longrun}). \TODO{prose -- author. FIXME: the "collapses" framing
is preset-induced (job 85 hot preset); job 97 shows vanilla stays coherent and
keeps hacking under a sane preset. Rewrite to "vanilla rises, route2 stays at 0"
once the matched beta=1e-5 figure (jobs 100/101) lands -- see figure provenance.}
% --- Figure: 200-step -------------------------------------------------------
% Provenance: route2 = pueue job 84 (out/runs/20260602T080804_..._route2_converge200_s41);
% vanilla = job 85 (out/runs/20260602T163201_..._vanilla_converge200_s41; vanilla still
% running at writing -> left panel fills to step 200 on completion). Data source committed
% at out/figs/dyn_longrun_200.csv; regen: uv run python scripts/plot_dynamics.py <both logs>.
% FIXME (2026-06-04, #184/#187): the vanilla "collapse around step 88" below is a
% PRESET artifact, not a property of vanilla GRPO. Job 85 used the hot preset
% (lr=3e-3, adam beta1=0.5, beta=0). Job 97 re-ran vanilla-200 at a gentle preset
% (lr=1e-3, adam 0.9/0.99, beta=0) and it stayed coherent: lp_s in [-0.47,-0.29]
% through step 200, deploy hack 0.375 (run_tests 0.875, file_marker 0.53), no
% collapse. Also job 84 (route2) and job 85 (vanilla) use mismatched optimizers,
% so this overlay is apples-to-oranges. Regenerate this figure from the matched
% beta=1e-5 pair (jobs 100/101, _none200_kl5 / _route2200_kl5) when they land, and
% drop the "collapses" claim -- the honest contrast is route2 holds deploy hack ~0
% while vanilla RISES (not collapses) under a sane preset.
\begin{figure}[t]
\centering
\includegraphics[width=0.95\linewidth]{../../out/figs/dyn_longrun_200.png}