Co-Prime Virtual-Aperture Attention / results.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
 1{
 2  "seed": 3030,
 3  "device": "cuda",
 4  "math": {
 5    "coprime": {
 6      "gcd": 1,
 7      "Dt": [
 8        0,
 9        4,
10        8
11      ],
12      "Dr": [
13        0,
14        3,
15        6,
16        9
17      ],
18      "physical_union_count": 6,
19      "physical_formula": 6,
20      "pair_count": 12,
21      "virtual_offsets": [
22        0,
23        3,
24        4,
25        6,
26        7,
27        8,
28        9,
29        10,
30        11,
31        13,
32        14,
33        17
34      ],
35      "virtual_count": 12,
36      "graph_reachable_offsets": [
37        0,
38        3,
39        4,
40        6,
41        7,
42        8,
43        9,
44        10,
45        11,
46        13,
47        14,
48        17
49      ]
50    },
51    "non_coprime_control": {
52      "gcd": 2,
53      "Dt": [
54        0,
55        4
56      ],
57      "Dr": [
58        0,
59        2,
60        4,
61        6
62      ],
63      "virtual_offsets": [
64        0,
65        2,
66        4,
67        6,
68        8,
69        10
70      ],
71      "virtual_count": 6
72    },
73    "claim_holds": true
74  },
75  "training": {
76    "cpa": {
77      "final_loss": 0.020130235701799393,
78      "accuracy": 0.9983259439468384,
79      "seconds": 3.4058370100101456
80    },
81    "single_sparse": {
82      "final_loss": 2.9466536045074463,
83      "accuracy": 0.0479910746216774,
84      "seconds": 3.195070609974209
85    },
86    "dense": {
87      "final_loss": 2.7530713081359863,
88      "accuracy": 0.14453125,
89      "seconds": 0.7880164489906747
90    },
91    "cost_summary": {
92      "cpa_edges_per_token_interior": 7,
93      "single_sparse_edges_per_token": 7,
94      "dense_edges_per_token_at_L": 12.0,
95      "cpa_virtual_count": 12,
96      "cpa_fraction_virtual_offsets_gt_8": 0.5
97    }
98  }
99}