Tail-Weighted Optimal Batch Scheduling / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "tabular",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.006,
 10      "batch": 32
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "batch": 32
 17        },
 18        "mean": 12.079814672470093
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.0015,
 23          "batch": 64
 24        },
 25        "mean": 16.354594945907593
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "batch": 32
 31        },
 32        "mean": 8.110420823097229
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "batch": 64
 38        },
 39        "mean": 13.231624126434326
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.006,
 44          "batch": 32
 45        },
 46        "mean": 6.854706883430481
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.006,
 51          "batch": 64
 52        },
 53        "mean": 9.873589992523193
 54      }
 55    ],
 56    "full": {
 57      "mean": 6.988916456699371,
 58      "std": 0.40272099288824403,
 59      "per_seed": [
 60        6.72859001159668,
 61        6.984129428863525,
 62        6.906972408294678,
 63        6.799135684967041,
 64        6.242666721343994,
 65        7.195944786071777,
 66        7.462729454040527,
 67        7.591163158416748
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "best_cfg": {
 74      "lr": 0.006,
 75      "batch_budget": 8000
 76    },
 77    "sweep": [
 78      {
 79        "cfg": {
 80          "lr": 0.006,
 81          "batch_budget": 8000
 82        },
 83        "mean": 9.922411799430847,
 84        "per_seed": [
 85          8.689536094665527,
 86          11.649765014648438,
 87          9.706146240234375,
 88          9.317113876342773,
 89          8.459336280822754,
 90          9.973560333251953,
 91          11.035902976989746,
 92          10.547933578491211
 93        ]
 94      },
 95      {
 96        "cfg": {
 97          "lr": 0.0015,
 98          "batch_budget": 8000
 99        },
100        "mean": 16.83699667453766,
101        "per_seed": [
102          15.854043006896973,
103          16.610029220581055,
104          18.131574630737305,
105          14.799296379089355,
106          15.572802543640137,
107          16.681474685668945,
108          17.97016716003418,
109          19.07658576965332
110        ]
111      },
112      {
113        "cfg": {
114          "lr": 0.003,
115          "batch_budget": 8000
116        },
117        "mean": 13.589162349700928,
118        "per_seed": [
119          12.75110912322998,
120          14.062518119812012,
121          13.924327850341797,
122          12.308281898498535,
123          12.10305404663086,
124          13.396442413330078,
125          14.798388481140137,
126          15.369176864624023
127        ]
128      }
129    ],
130    "mean": 9.922411799430847,
131    "per_seed": [
132      8.689536094665527,
133      11.649765014648438,
134      9.706146240234375,
135      9.317113876342773,
136      8.459336280822754,
137      9.973560333251953,
138      11.035902976989746,
139      10.547933578491211
140    ]
141  },
142  "comparison": {
143    "delta_mean": 2.933495342731476,
144    "idea_wins": 0,
145    "n_pairs": 8,
146    "per_seed_diffs": [
147      1.9609460830688477,
148      4.665635585784912,
149      2.7991738319396973,
150      2.5179781913757324,
151      2.2166695594787598,
152      2.777615547180176,
153      3.5731735229492188,
154      2.956770420074463
155    ],
156    "p_value": 0.0081,
157    "mde": 0.712641190380764,
158    "mde_rel_pct": 10.196733568014638,
159    "verdict": "idea worse (significant)",
160    "system_worked": false
161  },
162  "mechanism_signature": {
163    "predicted_vs_observed": {
164      "predicted_relative_weight_mean": 1.0,
165      "observed_batch_mean": 62.5,
166      "batch_min": 41,
167      "batch_max": 128,
168      "quantitative_match_tested": false
169    },
170    "confirmed": false,
171    "note": "Observed batch range is not a quantitative test of sqrt(c_t s_t); per-step trained-model noise estimates and correlation were not recorded, so the mechanism prediction is unconfirmed."
172  }
173}