Neighborhood-separator 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": 0.001
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.25371627509593964
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.2935705631971359
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 0.39029930531978607
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.26352048851549625,
 33      "std": 0.029814906699763923,
 34      "per_seed": [
 35        0.28988832235336304,
 36        0.23809197545051575,
 37        0.21863380074501038,
 38        0.2682510018348694,
 39        0.2573082745075226,
 40        0.26551553606987,
 41        0.3222951292991638,
 42        0.2481798678636551
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.2635671868920326,
 49    "std": 0.029872832584909817,
 50    "per_seed": [
 51      0.2901400327682495,
 52      0.23809197545051575,
 53      0.21863380074501038,
 54      0.2682510018348694,
 55      0.25730830430984497,
 56      0.26551553606987,
 57      0.3224169909954071,
 58      0.2481798529624939
 59    ],
 60    "n": 8,
 61    "selected_cfg": {
 62      "lr": 0.001
 63    },
 64    "sweep": [
 65      {
 66        "cfg": {
 67          "lr": 0.001
 68        },
 69        "mean": 0.2635671868920326,
 70        "per_seed": [
 71          0.2901400327682495,
 72          0.23809197545051575,
 73          0.21863380074501038,
 74          0.2682510018348694,
 75          0.25730830430984497,
 76          0.26551553606987,
 77          0.3224169909954071,
 78          0.2481798529624939
 79        ]
 80      },
 81      {
 82        "cfg": {
 83          "lr": 0.003
 84        },
 85        "mean": 0.307208513841033,
 86        "per_seed": [
 87          0.38006386160850525,
 88          0.22656194865703583,
 89          0.23053526878356934,
 90          0.3371211290359497,
 91          0.2561066746711731,
 92          0.34927764534950256,
 93          0.30410173535346985,
 94          0.3738998472690582
 95        ]
 96      },
 97      {
 98        "cfg": {
 99          "lr": 0.006
100        },
101        "mean": 0.4231664314866066,
102        "per_seed": [
103          0.28377190232276917,
104          0.34400641918182373,
105          0.41005638241767883,
106          0.5222185850143433,
107          0.3521057069301605,
108          0.5003604292869568,
109          0.3952743709087372,
110          0.5775376558303833
111        ]
112      }
113    ]
114  },
115  "comparison": {
116    "delta_mean": 4.6698376536369324e-05,
117    "idea_wins": 1,
118    "n_pairs": 8,
119    "per_seed_diffs": [
120      0.0002517104148864746,
121      0.0,
122      0.0,
123      0.0,
124      2.9802322387695312e-08,
125      0.0,
126      0.00012186169624328613,
127      -1.4901161193847656e-08
128    ],
129    "p_value": 0.25605,
130    "mde": 7.790313751068211e-05,
131    "mde_rel_pct": 0.029562459431347415,
132    "verdict": "no measurable effect",
133    "system_worked": false
134  },
135  "structural_check": {
136    "separator_size": 2,
137    "components": [
138      1,
139      1,
140      1,
141      1
142    ],
143    "anti_adjacent_components": true,
144    "mask_formula_exact": true
145  },
146  "custom_track": null,
147  "protocol_notes": "8 paired seeds; baseline sweep and idea sweep share lr union.",
148  "mechanism_signature": {
149    "predicted": {
150      "sparse_allowed_fraction": "< 1.0",
151      "anti_adjacent_components": true
152    },
153    "observed_mean_allowed_fraction": 1.0,
154    "observed_mean_components": 0.0,
155    "observed_mean_separator_size": 2.0,
156    "confirmed": false
157  }
158}