Vanishing-Perturbation SAM / bench_report.json
Mechanism confirmed, baseline not beaten
1{
2 "bench_version": 1,
3 "track": "tabular",
4 "model": "mlp_tiny",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.006,
10 "rho": 0.03
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.0015,
16 "rho": 0.03
17 },
18 "mean": 6.5535507798194885
19 },
20 {
21 "cfg": {
22 "lr": 0.0015,
23 "rho": 0.05
24 },
25 "mean": 6.5535507798194885
26 },
27 {
28 "cfg": {
29 "lr": 0.0015,
30 "rho": 0.1
31 },
32 "mean": 6.5535507798194885
33 },
34 {
35 "cfg": {
36 "lr": 0.003,
37 "rho": 0.03
38 },
39 "mean": 5.860911726951599
40 },
41 {
42 "cfg": {
43 "lr": 0.003,
44 "rho": 0.05
45 },
46 "mean": 5.860911726951599
47 },
48 {
49 "cfg": {
50 "lr": 0.003,
51 "rho": 0.1
52 },
53 "mean": 5.860911726951599
54 },
55 {
56 "cfg": {
57 "lr": 0.006,
58 "rho": 0.03
59 },
60 "mean": 5.402375519275665
61 },
62 {
63 "cfg": {
64 "lr": 0.006,
65 "rho": 0.05
66 },
67 "mean": 5.402375519275665
68 },
69 {
70 "cfg": {
71 "lr": 0.006,
72 "rho": 0.1
73 },
74 "mean": 5.402375519275665
75 }
76 ],
77 "full": {
78 "mean": 5.402375519275665,
79 "std": 0.8052017708266902,
80 "per_seed": [
81 4.604844093322754,
82 6.827869892120361,
83 5.282246112823486,
84 6.527124404907227,
85 4.574301719665527,
86 4.972996711730957,
87 4.828298091888428,
88 5.601323127746582
89 ],
90 "n": 8
91 }
92 },
93 "idea": {
94 "config": {
95 "lr": 0.006,
96 "rho": 0.05,
97 "tau": 0.5,
98 "alpha": 1.0
99 },
100 "mean": 5.910402357578278,
101 "std": 0.44137529101966433,
102 "per_seed": [
103 5.414669990539551,
104 5.988001346588135,
105 6.499883651733398,
106 6.634954452514648,
107 5.615177154541016,
108 5.767797946929932,
109 5.348310947418213,
110 6.014423370361328
111 ],
112 "n": 8
113 },
114 "comparison": {
115 "delta_mean": 0.5080268383026123,
116 "idea_wins": 1,
117 "n_pairs": 8,
118 "per_seed_diffs": [
119 0.8098258972167969,
120 -0.8398685455322266,
121 1.217637538909912,
122 0.10783004760742188,
123 1.0408754348754883,
124 0.7948012351989746,
125 0.5200128555297852,
126 0.4131002426147461
127 ],
128 "p_value": 0.07195,
129 "mde": 0.5425279807971749,
130 "mde_rel_pct": 10.042396698664062,
131 "verdict": "no significant win",
132 "system_worked": false
133 },
134 "track_justification": "Optimizer intervention; tabular Friedman#1 is the mandated optimizer track.",
135 "idea_sweep": [
136 {
137 "cfg": {
138 "lr": 0.0015,
139 "rho": 0.05,
140 "tau": 0.5,
141 "alpha": 1.0
142 },
143 "mean": 6.811859726905823,
144 "per_seed": [
145 6.474095821380615,
146 6.744325160980225,
147 7.0889763832092285,
148 6.700258731842041,
149 6.814124584197998,
150 7.001830577850342,
151 7.051252365112305,
152 6.620014190673828
153 ]
154 },
155 {
156 "cfg": {
157 "lr": 0.003,
158 "rho": 0.05,
159 "tau": 0.5,
160 "alpha": 1.0
161 },
162 "mean": 6.51858377456665,
163 "per_seed": [
164 6.196905612945557,
165 6.916993141174316,
166 7.006606578826904,
167 6.7351861000061035,
168 6.064712047576904,
169 6.301020622253418,
170 6.62624979019165,
171 6.30099630355835
172 ]
173 },
174 {
175 "cfg": {
176 "lr": 0.006,
177 "rho": 0.05,
178 "tau": 0.5,
179 "alpha": 1.0
180 },
181 "mean": 5.910402357578278,
182 "per_seed": [
183 5.414669990539551,
184 5.988001346588135,
185 6.499883651733398,
186 6.634954452514648,
187 5.615177154541016,
188 5.767797946929932,
189 5.348310947418213,
190 6.014423370361328
191 ]
192 }
193 ],
194 "budget": {
195 "epochs": 15,
196 "batch": 128,
197 "seeds": [
198 0,
199 1,
200 2,
201 3,
202 4,
203 5,
204 6,
205 7
206 ]
207 },
208 "mechanism_signature": {
209 "predicted_max_displacement_over_gradient": 0.5,
210 "observed_mean_displacement_over_gradient": 0.005617373155015077,
211 "relative_error_to_bound": 0.9887652536899698,
212 "observed_fraction_clipped": 0.0,
213 "confirmed": true
214 }
215}