Tensorized concentration mixing / bench_report.json

Unverified

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": 0.43008940666913986
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.45614606887102127
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.005
 27        },
 28        "mean": 0.5769104957580566
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.43690821528434753,
 33      "std": 0.04200321451882879,
 34      "per_seed": [
 35        0.40570154786109924,
 36        0.35749006271362305,
 37        0.4590928256511688,
 38        0.49807319045066833,
 39        0.43421313166618347,
 40        0.44106775522232056,
 41        0.48329558968544006,
 42        0.41633161902427673
 43      ],
 44      "n": 8
 45    },
 46    "sweep_union_lrs": [
 47      0.001,
 48      0.003,
 49      0.005
 50    ]
 51  },
 52  "idea": {
 53    "mean": 0.4732055962085724,
 54    "std": 0.05560245931804311,
 55    "per_seed": [
 56      0.4548509120941162,
 57      0.4054109454154968,
 58      0.4097138047218323,
 59      0.4698072671890259,
 60      0.5605213046073914,
 61      0.5118848085403442,
 62      0.5427255630493164,
 63      0.4307301640510559
 64    ],
 65    "n": 8,
 66    "best_cfg": {
 67      "lr": 0.001
 68    },
 69    "sweep": [
 70      {
 71        "cfg": {
 72          "lr": 0.001
 73        },
 74        "mean": 0.4732055962085724,
 75        "std": 0.05560245931804311,
 76        "per_seed": [
 77          0.4548509120941162,
 78          0.4054109454154968,
 79          0.4097138047218323,
 80          0.4698072671890259,
 81          0.5605213046073914,
 82          0.5118848085403442,
 83          0.5427255630493164,
 84          0.4307301640510559
 85        ]
 86      },
 87      {
 88        "cfg": {
 89          "lr": 0.003
 90        },
 91        "mean": 0.6192752122879028,
 92        "std": 0.09782140112106731,
 93        "per_seed": [
 94          0.6161494255065918,
 95          0.5007503628730774,
 96          0.5499135255813599,
 97          0.5674563646316528,
 98          0.7630049586296082,
 99          0.7055208683013916,
100          0.7415022253990173,
101          0.5099039673805237
102        ]
103      },
104      {
105        "cfg": {
106          "lr": 0.005
107        },
108        "mean": 0.7139862775802612,
109        "std": 0.13291049641408512,
110        "per_seed": [
111          0.7121991515159607,
112          0.5803670287132263,
113          0.6345881223678589,
114          0.5550000071525574,
115          0.9426005482673645,
116          0.8525124788284302,
117          0.8183250427246094,
118          0.6162978410720825
119        ]
120      }
121    ]
122  },
123  "comparison": {
124    "delta_mean": 0.036297380924224854,
125    "idea_wins": 2,
126    "n_pairs": 8,
127    "per_seed_diffs": [
128      0.04914936423301697,
129      0.04792088270187378,
130      -0.04937902092933655,
131      -0.028265923261642456,
132      0.12630817294120789,
133      0.07081705331802368,
134      0.05942997336387634,
135      0.014398545026779175
136    ],
137    "p_value": 0.1233,
138    "mde": 0.04700065274938183,
139    "mde_rel_pct": 10.757557561327378,
140    "verdict": "no significant win",
141    "system_worked": false
142  },
143  "mechanism_signature": {
144    "track_structure": "multi-token sequence forecast",
145    "implementation": "learned DCT-basis positive contraction",
146    "signature": {
147      "predicted": "axis mixing is non-expansive with spectral norm <=1",
148      "observed_mean_ratio": 0.7310447409751756,
149      "observed_max_ratio": 0.7322232178834247,
150      "predicted_max_singular_mean": 0.7532902434468269,
151      "confirmed": true
152    }
153  }
154}