Inverse-Square Fractional Attention / bench_report.json

Failed on benchmark

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": 0.0015,
 10      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.13084225729107857
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "weight_decay": 0.0
 24        },
 25        "mean": 0.13562202639877796
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.1526990309357643
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.1321341758593917,
 37      "std": 0.012441957039020079,
 38      "per_seed": [
 39        0.15123596787452698,
 40        0.11964408308267593,
 41        0.12323945015668869,
 42        0.12924952805042267,
 43        0.1238316223025322,
 44        0.11987392604351044,
 45        0.13868176937103271,
 46        0.1513170599937439
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.13770121987909079,
 53    "std": 0.012267757502007736,
 54    "per_seed": [
 55      0.14959165453910828,
 56      0.1260286271572113,
 57      0.1206536665558815,
 58      0.13636907935142517,
 59      0.1417829692363739,
 60      0.12556812167167664,
 61      0.142854243516922,
 62      0.1587613970041275
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": 0.005567044019699097,
 68    "idea_wins": 2,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      -0.0016443133354187012,
 72      0.00638454407453537,
 73      -0.00258578360080719,
 74      0.0071195513010025024,
 75      0.017951346933841705,
 76      0.005694195628166199,
 77      0.004172474145889282,
 78      0.007444337010383606
 79    ],
 80    "p_value": 0.0386,
 81    "mde": 0.00529214319698178,
 82    "mde_rel_pct": 4.00512824374318,
 83    "verdict": "idea worse (significant)",
 84    "system_worked": false
 85  },
 86  "idea_sweep": [
 87    {
 88      "cfg": {
 89        "lr": 0.0015,
 90        "weight_decay": 0.0
 91      },
 92      "mean": 0.1338496170938015
 93    },
 94    {
 95      "cfg": {
 96        "lr": 0.003,
 97        "weight_decay": 0.0
 98      },
 99      "mean": 0.13636435382068157
100    },
101    {
102      "cfg": {
103        "lr": 0.006,
104        "weight_decay": 0.0
105      },
106      "mean": 0.13316075690090656
107    }
108  ],
109  "math_check": {
110    "homogeneity_max_relative_error": 3.397532652615887e-05,
111    "near_origin_amplification": 40.80571546736738,
112    "predicted_amplification": 40.80571546736738,
113    "admissible_d3": true,
114    "relative_amplification_error": 0.0
115  },
116  "idea_best_cfg": {
117    "lr": 0.006,
118    "weight_decay": 0.0
119  },
120  "budget": {
121    "epochs": 18,
122    "n_train": 2000,
123    "n_test": 500,
124    "seeds": [
125      0,
126      1,
127      2,
128      3,
129      4,
130      5,
131      6,
132      7
133    ]
134  },
135  "mechanism_signature": {
136    "predicted_near_vs_far_kernel_ratio": 2571602743.0040903,
137    "observed_attention_odds_ratio_after_bias": 455527.4706971457,
138    "baseline_endpoint_recent_attention": 0.04563657566905022,
139    "idea_endpoint_recent_attention": 0.9985866546630859,
140    "confirmed": false
141  }
142}