Isometric tensor-network token mixer / 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      "epochs": 6,
 11      "weight_decay": 0.0
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "epochs": 6,
 18          "weight_decay": 0.0
 19        },
 20        "mean": 0.46786613017320633
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.003,
 25          "epochs": 6,
 26          "weight_decay": 0.0
 27        },
 28        "mean": 0.5299046412110329
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.006,
 33          "epochs": 6,
 34          "weight_decay": 0.0
 35        },
 36        "mean": 0.7419369965791702
 37      }
 38    ],
 39    "full": {
 40      "mean": 0.4766220450401306,
 41      "std": 0.09934194575564281,
 42      "per_seed": [
 43        0.5715442299842834,
 44        0.39496564865112305,
 45        0.3930272161960602,
 46        0.5119274258613586,
 47        0.4745565354824066,
 48        0.4187285602092743,
 49        0.6776894927024841,
 50        0.37053725123405457
 51      ],
 52      "n": 8
 53    }
 54  },
 55  "idea": {
 56    "mean": 0.5983163416385651,
 57    "std": 0.16587939812424876,
 58    "per_seed": [
 59      0.3822963237762451,
 60      0.5229498744010925,
 61      0.8081433773040771,
 62      0.863442063331604,
 63      0.6520702838897705,
 64      0.45361945033073425,
 65      0.6645830869674683,
 66      0.4394262731075287
 67    ],
 68    "n": 8
 69  },
 70  "comparison": {
 71    "delta_mean": 0.12169429659843445,
 72    "idea_wins": 2,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      -0.18924790620803833,
 76      0.12798422574996948,
 77      0.41511616110801697,
 78      0.35151463747024536,
 79      0.1775137484073639,
 80      0.03489089012145996,
 81      -0.01310640573501587,
 82      0.06888902187347412
 83    ],
 84    "p_value": 0.1299,
 85    "mde": 0.16337849985483555,
 86    "mde_rel_pct": 34.27841862435872,
 87    "verdict": "no significant win",
 88    "system_worked": false
 89  },
 90  "mechanism_signature": {
 91    "idea_sweep": [
 92      {
 93        "cfg": {
 94          "lr": 0.001,
 95          "epochs": 6,
 96          "weight_decay": 0.0
 97        },
 98        "result": {
 99          "mean": 0.5983163416385651,
100          "std": 0.16587939812424876,
101          "per_seed": [
102            0.3822963237762451,
103            0.5229498744010925,
104            0.8081433773040771,
105            0.863442063331604,
106            0.6520702838897705,
107            0.45361945033073425,
108            0.6645830869674683,
109            0.4394262731075287
110          ],
111          "n": 8
112        }
113      },
114      {
115        "cfg": {
116          "lr": 0.003,
117          "epochs": 6,
118          "weight_decay": 0.0
119        },
120        "result": {
121          "mean": 2.1142318695783615,
122          "std": 0.7959891886633405,
123          "per_seed": [
124            2.495609998703003,
125            2.4600517749786377,
126            0.8293719291687012,
127            1.1857432126998901,
128            1.480914831161499,
129            3.2832841873168945,
130            2.778350830078125,
131            2.4005281925201416
132          ],
133          "n": 8
134        }
135      },
136      {
137        "cfg": {
138          "lr": 0.006,
139          "epochs": 6,
140          "weight_decay": 0.0
141        },
142        "result": {
143          "mean": 5.809824541211128,
144          "std": 2.711991843326257,
145          "per_seed": [
146            4.945932388305664,
147            11.078203201293945,
148            1.813473105430603,
149            5.887286186218262,
150            2.609509229660034,
151            6.428954124450684,
152            7.734877109527588,
153            5.980360984802246
154          ],
155          "n": 8
156        }
157      }
158    ],
159    "observed_best_cfg": {
160      "lr": 0.001,
161      "epochs": 6,
162      "weight_decay": 0.0
163    },
164    "math_check": [
165      {
166        "n": 2,
167        "norm_ratio": 1.0000001192092896,
168        "inverse_max_error": 2.384185791015625e-07
169      },
170      {
171        "n": 4,
172        "norm_ratio": 1.0,
173        "inverse_max_error": 4.76837158203125e-07
174      },
175      {
176        "n": 8,
177        "norm_ratio": 1.0000001192092896,
178        "inverse_max_error": 4.76837158203125e-07
179      },
180      {
181        "n": 16,
182        "norm_ratio": 1.0000001192092896,
183        "inverse_max_error": 7.152557373046875e-07
184      },
185      {
186        "n": 32,
187        "norm_ratio": 1.0,
188        "inverse_max_error": 9.5367431640625e-07
189      }
190    ],
191    "mechanism_signature": {
192      "prediction": "trained isometric token mixer preserves feature-token Euclidean norm and has exact adjoint inverse",
193      "predicted_norm_ratio": 1.0,
194      "observed_norm_ratio_seed0": 1.0,
195      "observed_inverse_max_error_seed0": 1.1920928955078125e-06,
196      "baseline_observed_norm_ratio_seed0": 0.5025865435600281,
197      "confirmed": true,
198      "trained_model_metric_seed0": 0.3822963237762451
199    }
200  }
201}