Doubly-Stochastic Hyper-Residual Blocks / bench_report.json
Mechanism confirmed, baseline not beaten
1{
2 "bench_version": 1,
3 "track": "sequence",
4 "model": "transformer_tiny",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.0015
10 },
11 "sweep": [
12 {
13 "cfg": {
14 "lr": 0.0007
15 },
16 "mean": 0.18085692822933197
17 },
18 {
19 "cfg": {
20 "lr": 0.001
21 },
22 "mean": 0.17446092143654823
23 },
24 {
25 "cfg": {
26 "lr": 0.0015
27 },
28 "mean": 0.17426595836877823
29 },
30 {
31 "cfg": {
32 "lr": 0.002
33 },
34 "mean": 0.1786409355700016
35 },
36 {
37 "cfg": {
38 "lr": 0.003
39 },
40 "mean": 0.18349025771021843
41 }
42 ],
43 "full": {
44 "mean": 0.16209838539361954,
45 "std": 0.024681871508716122,
46 "per_seed": [
47 0.1827893853187561,
48 0.13988947868347168,
49 0.18467898666858673,
50 0.1897059828042984,
51 0.11963248252868652,
52 0.15095271170139313,
53 0.18380902707576752,
54 0.14532902836799622
55 ],
56 "n": 8
57 }
58 },
59 "idea": {
60 "mean": 0.18732197023928165,
61 "std": 0.025151544057018065,
62 "per_seed": [
63 0.19494226574897766,
64 0.15557610988616943,
65 0.19609080255031586,
66 0.2065867781639099,
67 0.1564088761806488,
68 0.16889484226703644,
69 0.23473307490348816,
70 0.18534301221370697
71 ],
72 "n": 8
73 },
74 "comparison": {
75 "delta_mean": 0.025223584845662117,
76 "idea_wins": 0,
77 "n_pairs": 8,
78 "per_seed_diffs": [
79 0.012152880430221558,
80 0.015686631202697754,
81 0.011411815881729126,
82 0.01688079535961151,
83 0.03677639365196228,
84 0.01794213056564331,
85 0.05092404782772064,
86 0.040013983845710754
87 ],
88 "p_value": 0.0081,
89 "mde": 0.01259364706714858,
90 "mde_rel_pct": 7.769137882877572,
91 "verdict": "idea worse (significant)",
92 "system_worked": false
93 },
94 "mechanism_signature": {
95 "idea_lr_sweep": [
96 {
97 "lr": 0.001,
98 "result": {
99 "mean": 0.2357968259602785,
100 "std": 0.02821421648971268,
101 "per_seed": [
102 0.23772037029266357,
103 0.20389337837696075,
104 0.24708855152130127,
105 0.242384672164917,
106 0.21312211453914642,
107 0.2053222507238388,
108 0.2967439293861389,
109 0.24009934067726135
110 ],
111 "n": 8
112 }
113 },
114 {
115 "lr": 0.0015,
116 "result": {
117 "mean": 0.20389264076948166,
118 "std": 0.024621166913138676,
119 "per_seed": [
120 0.21074053645133972,
121 0.1622573882341385,
122 0.21052612364292145,
123 0.21601423621177673,
124 0.1940460503101349,
125 0.18469412624835968,
126 0.2528582811355591,
127 0.20000438392162323
128 ],
129 "n": 8
130 }
131 },
132 {
133 "lr": 0.002,
134 "result": {
135 "mean": 0.18732197023928165,
136 "std": 0.025151544057018065,
137 "per_seed": [
138 0.19494226574897766,
139 0.15557610988616943,
140 0.19609080255031586,
141 0.2065867781639099,
142 0.1564088761806488,
143 0.16889484226703644,
144 0.23473307490348816,
145 0.18534301221370697
146 ],
147 "n": 8
148 }
149 }
150 ],
151 "selected_lr": 0.002,
152 "trained_model": true,
153 "predicted": {
154 "max_frobenius_ratio": "<=1",
155 "row_col_error": "0"
156 },
157 "observed": {
158 "max_frobenius_ratio": 0.7491575479507446,
159 "mean_frobenius_ratio": 0.748988151550293,
160 "max_row_error": 2.384185791015625e-07,
161 "max_col_error": 1.1920928955078125e-07
162 },
163 "confirmed": true,
164 "metric_on_signature_model": 0.19494226574897766
165 },
166 "custom_track": null
167}