Dissipative Softmax Latent Layer / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "token_expert_sequence",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01,
 10      "epochs": 25
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 25
 17        },
 18        "mean": 0.0012499999720603228
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "epochs": 25
 24        },
 25        "mean": 0.0012499999720603228
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "epochs": 25
 31        },
 32        "mean": 0.0006249999860301614
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.0015624999650754035,
 37      "std": 0.0017399263244939971,
 38      "per_seed": [
 39        0.0,
 40        0.0024999999441206455,
 41        0.0,
 42        0.0,
 43        0.0024999999441206455,
 44        0.0024999999441206455,
 45        0.0,
 46        0.004999999888241291
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.4946874864399433,
 53    "std": 0.02705426452774727,
 54    "per_seed": [
 55      0.5399999618530273,
 56      0.5024999976158142,
 57      0.48499998450279236,
 58      0.5049999952316284,
 59      0.4449999928474426,
 60      0.5199999809265137,
 61      0.47999998927116394,
 62      0.47999998927116394
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": 0.4931249864748679,
 68    "idea_wins": 0,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      0.5399999618530273,
 72      0.49999999767169356,
 73      0.48499998450279236,
 74      0.5049999952316284,
 75      0.442499992903322,
 76      0.517499980982393,
 77      0.47999998927116394,
 78      0.47499998938292265
 79    ],
 80    "p_value": 0.0081,
 81    "mde": 0.024702187634040935,
 82    "mde_rel_pct": 1580.9400439153835,
 83    "verdict": "idea worse (significant)",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "predicted_effect": "occupation mismatch decreases as reset_rate increases",
 88    "mismatch_at_r": NaN,
 89    "mismatch_at_r_0.5": NaN,
 90    "observed_cycle_proxy": NaN,
 91    "confirmed": false
 92  },
 93  "idea_sweep": [
 94    {
 95      "cfg": {
 96        "lr": 0.01,
 97        "epochs": 25,
 98        "r": 0.5
 99      },
100      "result": {
101        "mean": 0.4946874864399433,
102        "std": 0.02705426452774727,
103        "per_seed": [
104          0.5399999618530273,
105          0.5024999976158142,
106          0.48499998450279236,
107          0.5049999952316284,
108          0.4449999928474426,
109          0.5199999809265137,
110          0.47999998927116394,
111          0.47999998927116394
112        ],
113        "n": 8
114      }
115    },
116    {
117      "cfg": {
118        "lr": 0.01,
119        "epochs": 25,
120        "r": 2.0
121      },
122      "result": {
123        "mean": 0.4946874864399433,
124        "std": 0.02705426452774727,
125        "per_seed": [
126          0.5399999618530273,
127          0.5024999976158142,
128          0.48499998450279236,
129          0.5049999952316284,
130          0.4449999928474426,
131          0.5199999809265137,
132          0.47999998927116394,
133          0.47999998927116394
134        ],
135        "n": 8
136      }
137    },
138    {
139      "cfg": {
140        "lr": 0.01,
141        "epochs": 25,
142        "r": 8.0
143      },
144      "result": {
145        "mean": 0.4946874864399433,
146        "std": 0.02705426452774727,
147        "per_seed": [
148          0.5399999618530273,
149          0.5024999976158142,
150          0.48499998450279236,
151          0.5049999952316284,
152          0.4449999928474426,
153          0.5199999809265137,
154          0.47999998927116394,
155          0.47999998927116394
156        ],
157        "n": 8
158      }
159    }
160  ],
161  "custom_track": {
162    "name": "token_expert_sequence",
163    "file": "/home/maxwelhelp/all/math2nn/bench/custom_tracks/token_expert_sequence.py",
164    "domain": "moe-routing"
165  }
166}