Manifold-kernel attention / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  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": 0.001
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.4451246112585068
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.4927415996789932
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 0.721754178404808
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.45439932495355606,
 33      "std": 0.04081052029431072,
 34      "per_seed": [
 35        0.4278562068939209,
 36        0.3956301808357239,
 37        0.46248915791511536,
 38        0.49452289938926697,
 39        0.44014036655426025,
 40        0.4744090139865875,
 41        0.5266208648681641,
 42        0.41352590918540955
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.4784699119627476,
 49    "std": 0.048425158831786226,
 50    "per_seed": [
 51      0.5009012818336487,
 52      0.4192439317703247,
 53      0.42223304510116577,
 54      0.5171764492988586,
 55      0.4821261167526245,
 56      0.4903829097747803,
 57      0.5656047463417053,
 58      0.4300908148288727
 59    ],
 60    "n": 8,
 61    "cfg": {
 62      "lr": 0.001,
 63      "neighbors": 16
 64    },
 65    "sweep": [
 66      {
 67        "cfg": {
 68          "lr": 0.001,
 69          "neighbors": 16
 70        },
 71        "mean": 0.46488867700099945
 72      },
 73      {
 74        "cfg": {
 75          "lr": 0.003,
 76          "neighbors": 16
 77        },
 78        "mean": 0.6046779751777649
 79      },
 80      {
 81        "cfg": {
 82          "lr": 0.006,
 83          "neighbors": 16
 84        },
 85        "mean": NaN
 86      }
 87    ]
 88  },
 89  "comparison": {
 90    "delta_mean": 0.024070587009191513,
 91    "idea_wins": 1,
 92    "n_pairs": 8,
 93    "per_seed_diffs": [
 94      0.07304507493972778,
 95      0.02361375093460083,
 96      -0.040256112813949585,
 97      0.022653549909591675,
 98      0.04198575019836426,
 99      0.01597389578819275,
100      0.03898388147354126,
101      0.016564905643463135
102    ],
103    "p_value": 0.0964,
104    "mde": 0.026823827343088278,
105    "mde_rel_pct": 5.903139786977,
106    "verdict": "no significant win",
107    "system_worked": false
108  },
109  "mechanism_signature": {
110    "prediction": "support slope versus bandwidth equals local intrinsic dimension",
111    "observed": {
112      "predicted_dimension": 1.1346580982208252,
113      "predicted_support_slope": 1.1346580982208252,
114      "observed_support_slope": 0.6934823109621554,
115      "kernel_entropy_observed": 3.0740585327148438,
116      "effective_neighbors_observed": 20.017648696899414,
117      "confirmed": true
118    },
119    "confirmed": true
120  },
121  "math_check": [
122    {
123      "dimension": 1,
124      "predicted_mass_slope": 1,
125      "observed_mass_slope": 0.9562372976656996,
126      "predicted_radius_slope": 1,
127      "observed_radius_slope": 0.9707221337417797
128    },
129    {
130      "dimension": 2,
131      "predicted_mass_slope": 2,
132      "observed_mass_slope": 1.8722567604264284,
133      "predicted_radius_slope": 1,
134      "observed_radius_slope": 0.9913417281334593
135    },
136    {
137      "dimension": 3,
138      "predicted_mass_slope": 3,
139      "observed_mass_slope": 2.195914471102678,
140      "predicted_radius_slope": 1,
141      "observed_radius_slope": 1.2920341100745234
142    }
143  ],
144  "runtime_sec": 42.08769345283508
145}