Spiderweb Hierarchical Attention / bench_report_rerun.json

✓✓ Beats tuned baseline

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": 0.16495772451162338
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.18329624831676483
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.43168069422245026
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.16532287374138832,
 33      "std": 0.008160238971414253,
 34      "per_seed": [
 35        0.17448444664478302,
 36        0.1530463844537735,
 37        0.17006324231624603,
 38        0.16223682463169098,
 39        0.16158440709114075,
 40        0.15676827728748322,
 41        0.17869338393211365,
 42        0.16570602357387543
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.12441098596900702,
 49    "std": 0.008560845180504158,
 50    "per_seed": [
 51      0.1358136236667633,
 52      0.10833562165498734,
 53      0.11411764472723007,
 54      0.12626340985298157,
 55      0.1262723207473755,
 56      0.12286490947008133,
 57      0.13141369819641113,
 58      0.1302066594362259
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": -0.040911887772381306,
 64    "idea_wins": 8,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      -0.038670822978019714,
 68      -0.04471076279878616,
 69      -0.05594559758901596,
 70      -0.03597341477870941,
 71      -0.03531208634376526,
 72      -0.033903367817401886,
 73      -0.047279685735702515,
 74      -0.035499364137649536
 75    ],
 76    "p_value": 0.0081,
 77    "mde": 0.006478177383445342,
 78    "mde_rel_pct": 3.9185003483420218,
 79    "verdict": "idea better (significant)",
 80    "system_worked": true
 81  },
 82  "mechanism_signature": {
 83    "prediction": "hierarchical bounded-neighbor attention preserves measurable long-range influence",
 84    "baseline_signature": {
 85      "far_first_token_effect": 0.027809560298919678,
 86      "local_last_token_effect": 1.0498290061950684,
 87      "far_to_local_ratio": 0.02648960913626757
 88    },
 89    "idea_signature": {
 90      "far_first_token_effect": 0.020777523517608643,
 91      "local_last_token_effect": 1.4758778810501099,
 92      "far_to_local_ratio": 0.014078077626615245
 93    },
 94    "confirmed": false,
 95    "confirmation_rule": "idea far-to-local ratio > baseline far-to-local ratio by at least 20%"
 96  },
 97  "extra": {
 98    "idea_sweep": [
 99      {
100        "cfg": {
101          "lr": 0.001
102        },
103        "result": {
104          "mean": 0.12441098596900702,
105          "std": 0.008560845180504158,
106          "per_seed": [
107            0.1358136236667633,
108            0.10833562165498734,
109            0.11411764472723007,
110            0.12626340985298157,
111            0.1262723207473755,
112            0.12286490947008133,
113            0.13141369819641113,
114            0.1302066594362259
115          ],
116          "n": 8
117        }
118      },
119      {
120        "cfg": {
121          "lr": 0.003
122        },
123        "result": {
124          "mean": 0.13378733582794666,
125          "std": 0.012523346697500463,
126          "per_seed": [
127            0.13517771661281586,
128            0.11398444324731827,
129            0.1407897174358368,
130            0.1486937254667282,
131            0.13302475214004517,
132            0.11315087229013443,
133            0.14389027655124664,
134            0.14158718287944794
135          ],
136          "n": 8
137        }
138      },
139      {
140        "cfg": {
141          "lr": 0.01
142        },
143        "result": {
144          "mean": 0.1740857046097517,
145          "std": 0.015146774808178592,
146          "per_seed": [
147            0.2027415633201599,
148            0.14727982878684998,
149            0.17759738862514496,
150            0.1726507693529129,
151            0.16134709119796753,
152            0.1725672036409378,
153            0.18487928807735443,
154            0.1736225038766861
155          ],
156          "n": 8
157        }
158      }
159    ],
160    "idea_best_cfg": {
161      "lr": 0.001
162    },
163    "params_baseline": 71169,
164    "params_idea": 71181,
165    "elapsed_sec": 418.1431157588959,
166    "n_train": 2000,
167    "n_test": 500,
168    "epochs": 15
169  }
170}