Centered Heavy-Tail Clipping Optimizer / results.json
Failed on benchmark
1{
2 "operator_check": {
3 "inside_max_error": 0.0,
4 "max_output_residual": 0.7299999999998885,
5 "tau": 0.73,
6 "minimum_direction_cosine": 0.9999999999999996
7 },
8 "bias_scaling": {
9 "taus": [
10 2.0,
11 4.0,
12 8.0,
13 16.0,
14 32.0,
15 64.0
16 ],
17 "tail_bias": [
18 2.1078901113465762,
19 1.485287712794427,
20 1.047741566357388,
21 0.7380726599802617,
22 0.5182458643749303,
23 0.36457934163643035
24 ],
25 "loglog_slope": -0.5062864083363753,
26 "expected_slope": -0.5
27 },
28 "runs": {
29 "heavy_tail/sgd": {
30 "final_loss_mean": 1.5904582091610717,
31 "final_loss_std": 4.791403952086443,
32 "p99_update_mean": 1.4383751036429422,
33 "divergences": 0,
34 "final_losses": [
35 0.0909823080030608,
36 0.05312237551349112,
37 0.09223474792462609,
38 0.07947341840642966,
39 0.15228072810776913,
40 0.1532125708388954,
41 0.10080784091637439,
42 0.10629070802367911,
43 0.09703321835086269,
44 0.5679931664106993,
45 17.47586128718058,
46 0.11620614025639271
47 ]
48 },
49 "heavy_tail/global_clip": {
50 "final_loss_mean": 0.05626926714363623,
51 "final_loss_std": 0.01365064201168186,
52 "p99_update_mean": 0.14999999999999325,
53 "divergences": 0,
54 "final_losses": [
55 0.05423517478316375,
56 0.02972028068154693,
57 0.041147602941888496,
58 0.05373200563878424,
59 0.06955790364444688,
60 0.03964291055194818,
61 0.07311448424448197,
62 0.06882789460159842,
63 0.0468672212466314,
64 0.06692207337140618,
65 0.06285721333444437,
66 0.06860644068329401
67 ]
68 },
69 "heavy_tail/centered": {
70 "final_loss_mean": 0.0012210588447355575,
71 "final_loss_std": 0.0004985480793692171,
72 "p99_update_mean": 1.3289220351289603,
73 "divergences": 0,
74 "final_losses": [
75 0.0013646504546732253,
76 0.0006280569779644884,
77 0.001277164577435138,
78 0.0007120214502243642,
79 0.0009400869903349203,
80 0.001395007860398821,
81 0.0010673624658662617,
82 0.0010298607072653143,
83 0.0011537571390326617,
84 0.0015756784939763432,
85 0.0008945247718432248,
86 0.0026145342478119268
87 ]
88 },
89 "gaussian/sgd": {
90 "final_loss_mean": 0.0016830468711389003,
91 "final_loss_std": 0.0007888298476053901,
92 "p99_update_mean": 1.3292792381825755,
93 "divergences": 0,
94 "final_losses": [
95 0.0007722236026109946,
96 0.001467706388583531,
97 0.0013038306410087111,
98 0.0008065105614733637,
99 0.0024601787884927727,
100 0.002035386811184109,
101 0.0016890198835242236,
102 0.0008171132569252109,
103 0.0014557284199099977,
104 0.0014400223096305545,
105 0.0035888134542687105,
106 0.0023600283360546244
107 ]
108 },
109 "gaussian/global_clip": {
110 "final_loss_mean": 0.0016830755245458389,
111 "final_loss_std": 0.0007888650369945481,
112 "p99_update_mean": 0.14999999999999317,
113 "divergences": 0,
114 "final_losses": [
115 0.000772225305850193,
116 0.001467882370216722,
117 0.001303797848379775,
118 0.0008065040691951997,
119 0.0024603130507194304,
120 0.002035403423231237,
121 0.001688927591324899,
122 0.0008171641409295063,
123 0.0014557480466857592,
124 0.0014400554891856017,
125 0.0035890348798259565,
126 0.0023598500790057864
127 ]
128 },
129 "gaussian/centered": {
130 "final_loss_mean": 0.0007145656274612843,
131 "final_loss_std": 0.00021586353232532467,
132 "p99_update_mean": 1.3292060953351132,
133 "divergences": 0,
134 "final_losses": [
135 0.0006701332955378414,
136 0.00046670808528075574,
137 0.0005240023180628656,
138 0.0006419451792648101,
139 0.0012294410212866519,
140 0.00036388347280563916,
141 0.0008182256003758723,
142 0.0006522662877620825,
143 0.0007491919267796089,
144 0.0007148583521299482,
145 0.0008539547149137506,
146 0.0008901772753355874
147 ]
148 }
149 }
150}