From d3df5628b570af8fe2e22644b9c0849f69b9f3a1 Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Wed, 8 Apr 2026 02:11:00 -0500 Subject: Extend perturbation audit to vanilla early-epoch checkpoints MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Cross-metric disambiguation confirmation. Vanilla DFA at ep 1 (meaningful regime, ||g||~6e-7) deep rho across 3 seeds: s42: deep rho -0.008 s123: deep rho +0.000 s456: deep rho -0.000 mean: -0.003 ± 0.005 Compare to penalized DFA 3-seed: deep rho +0.080 ± 0.011. The disambiguation (penalty CREATES alignment, not just reveals it) is now confirmed by TWO independent metrics: - cos: vanilla -0.008 ± 0.013, penalized +0.155 ± 0.025 - rho: vanilla -0.003 ± 0.005, penalized +0.080 ± 0.011 Both metrics agree on the vanilla→penalized transition. The l0 (embedding) rho is high (~0.25-0.29) at every vanilla checkpoint, mirroring the cos l0 +0.42 — the embedding layer is genuinely useful while the deep blocks are not, by BOTH metrics. The penalty restores some deep usefulness to ~+0.08 rho / +0.16 cos. Cross-metric agreement rules out single-metric artifacts on either side. --- experiments/perturbation_correlation_audit.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/experiments/perturbation_correlation_audit.py b/experiments/perturbation_correlation_audit.py index cba84ea..163d3a8 100644 --- a/experiments/perturbation_correlation_audit.py +++ b/experiments/perturbation_correlation_audit.py @@ -127,6 +127,12 @@ def main(): "results/dfa_pen_short/dfa_pen_lam0.01_s123.pt", 123), ("penalized DFA s456 lam=1e-2 30ep", "results/dfa_pen_short/dfa_pen_lam0.01_s456.pt", 456), + ("vanilla DFA s42 ep1 (meaningful regime)", + "results/vanilla_dfa_early_ckpts/vanilla_dfa_s42_ep1.pt", 42), + ("vanilla DFA s123 ep1 (meaningful regime)", + "results/vanilla_dfa_early_ckpts/vanilla_dfa_s123_ep1.pt", 123), + ("vanilla DFA s456 ep1 (meaningful regime)", + "results/vanilla_dfa_early_ckpts/vanilla_dfa_s456_ep1.pt", 456), ] print("=" * 76) -- cgit v1.2.3