Gauge-Patched Local Experts / results.json
Beats tuned baseline
1{
2 "seed": 1270,
3 "math_checks": [
4 {
5 "prediction": "orthogonal transition preserves feature norm",
6 "max_abs_norm_error": 2.220446049250313e-16
7 },
8 {
9 "prediction": "extension Holder norm scales as gap^{-beta}",
10 "beta": 0.75,
11 "observed_loglog_slope": -0.749600946770213,
12 "predicted_slope": -0.75,
13 "gaps": [
14 0.04,
15 0.06,
16 0.09,
17 0.13,
18 0.19,
19 0.28
20 ],
21 "holder_norms": [
22 1.2303385369951036,
23 0.9081786759107097,
24 0.6702163960736099,
25 0.5087042847476028,
26 0.3827046214218064,
27 0.28613772255957365
28 ]
29 },
30 {
31 "prediction": "small-angle patch residual is quadratic",
32 "observed_loglog_slope": 1.9994916102004505,
33 "predicted_slope": 2.0,
34 "residual_over_eps2_norm2": [
35 0.9999979166684178,
36 0.9999916666944421,
37 0.9999666671111086,
38 0.9998666737775788,
39 0.9994667804314434,
40 0.9978684862791442
41 ]
42 }
43 ],
44 "tiny_experiment": {
45 "metric": "mean task MSE; aligned overlap MSE",
46 "matched_initialization": true,
47 "baseline": {
48 "lambda_patch": 0.0,
49 "task_mse": 2.881633191613442e-15,
50 "overlap_mse": 0.72025465965271
51 },
52 "gauge_patched_lambda_1": {
53 "lambda_patch": 1.0,
54 "task_mse": 0.08002829551696777,
55 "overlap_mse": 0.08002829551696777
56 },
57 "lambda_sweep": [
58 {
59 "lambda_patch": 0.0,
60 "task_mse": 2.881633191613442e-15,
61 "overlap_mse": 0.72025465965271
62 },
63 {
64 "lambda_patch": 0.01,
65 "task_mse": 6.922819011379033e-05,
66 "overlap_mse": 0.6922863125801086
67 },
68 {
69 "lambda_patch": 0.1,
70 "task_mse": 0.005001768469810486,
71 "overlap_mse": 0.5001768469810486
72 },
73 {
74 "lambda_patch": 1.0,
75 "task_mse": 0.08002829551696777,
76 "overlap_mse": 0.08002829551696777
77 },
78 {
79 "lambda_patch": 10.0,
80 "task_mse": 0.1633230447769165,
81 "overlap_mse": 0.001633228501304984
82 }
83 ]
84 }
85}