Shell-Wise Balanced MoE Routing / results.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  1{
  2  "seed": 1234,
  3  "toy_predictions": [
  4    {
  5      "prediction": "per-shell count spread <= 1",
  6      "E": 3,
  7      "K": 2,
  8      "observed": 1,
  9      "predicted_max": 1
 10    },
 11    {
 12      "prediction": "per-shell count spread <= 1",
 13      "E": 3,
 14      "K": 4,
 15      "observed": 1,
 16      "predicted_max": 1
 17    },
 18    {
 19      "prediction": "per-shell count spread <= 1",
 20      "E": 3,
 21      "K": 7,
 22      "observed": 1,
 23      "predicted_max": 1
 24    },
 25    {
 26      "prediction": "per-shell count spread <= 1",
 27      "E": 5,
 28      "K": 2,
 29      "observed": 1,
 30      "predicted_max": 1
 31    },
 32    {
 33      "prediction": "per-shell count spread <= 1",
 34      "E": 5,
 35      "K": 4,
 36      "observed": 1,
 37      "predicted_max": 1
 38    },
 39    {
 40      "prediction": "per-shell count spread <= 1",
 41      "E": 5,
 42      "K": 7,
 43      "observed": 1,
 44      "predicted_max": 1
 45    },
 46    {
 47      "prediction": "per-shell count spread <= 1",
 48      "E": 8,
 49      "K": 2,
 50      "observed": 1,
 51      "predicted_max": 1
 52    },
 53    {
 54      "prediction": "per-shell count spread <= 1",
 55      "E": 8,
 56      "K": 4,
 57      "observed": 1,
 58      "predicted_max": 1
 59    },
 60    {
 61      "prediction": "per-shell count spread <= 1",
 62      "E": 8,
 63      "K": 7,
 64      "observed": 1,
 65      "predicted_max": 1
 66    },
 67    {
 68      "prediction": "Eq.66 marginal spread is zero for all P",
 69      "E": 4,
 70      "K": 3,
 71      "observed_max_spread": 0.0,
 72      "predicted": 0.0,
 73      "baseline_max_spread": 20.346579000000002
 74    },
 75    {
 76      "prediction": "Eq.66 marginal spread is zero for all P",
 77      "E": 8,
 78      "K": 5,
 79      "observed_max_spread": 0.0,
 80      "predicted": 0.0,
 81      "baseline_max_spread": 18.0978129981
 82    },
 83    {
 84      "prediction": "Eq.66 marginal spread is zero for all P",
 85      "E": 8,
 86      "K": 8,
 87      "observed_max_spread": 0.0,
 88      "predicted": 0.0,
 89      "baseline_max_spread": 17.530729260504636
 90    }
 91  ],
 92  "comparison": {
 93    "baseline_cv": 1.0947531323437112,
 94    "idea_cv": 0.0,
 95    "baseline_loads": [
 96      1627.0,
 97      1320.0,
 98      182.0,
 99      211.0,
100      199.0,
101      186.0,
102      199.0,
103      172.0
104    ],
105    "idea_loads": [
106      512.0,
107      512.0,
108      512.0,
109      512.0,
110      512.0,
111      512.0,
112      512.0,
113      512.0
114    ],
115    "reassigned_fraction": 0.47119140625,
116    "router_score_loss_per_token": 0.5094270727563867,
117    "shells": [
118      {
119        "shell": 0,
120        "baseline_cv": 0.3569893643373707,
121        "idea_cv": 0.0,
122        "size": 512
123      },
124      {
125        "shell": 1,
126        "baseline_cv": 0.4122898375612113,
127        "idea_cv": 0.0,
128        "size": 512
129      },
130      {
131        "shell": 2,
132        "baseline_cv": 0.667317390751957,
133        "idea_cv": 0.0,
134        "size": 512
135      },
136      {
137        "shell": 3,
138        "baseline_cv": 0.9123581867542484,
139        "idea_cv": 0.0,
140        "size": 512
141      },
142      {
143        "shell": 4,
144        "baseline_cv": 1.2966867333226635,
145        "idea_cv": 0.0,
146        "size": 512
147      },
148      {
149        "shell": 5,
150        "baseline_cv": 1.6385403422329887,
151        "idea_cv": 0.0,
152        "size": 512
153      },
154      {
155        "shell": 6,
156        "baseline_cv": 1.7581423301669863,
157        "idea_cv": 0.0,
158        "size": 512
159      },
160      {
161        "shell": 7,
162        "baseline_cv": 2.122370270807205,
163        "idea_cv": 0.0,
164        "size": 512
165      }
166    ]
167  }
168}