{
  "metadata": {
    "created_at_unix": 1786240974.1639438,
    "duration_seconds": 111.84773588180542,
    "python": "3.14.3 (main, Mar 24 2026, 22:50:36) [Clang 22.1.1 ]",
    "platform": "Linux-7.0.0-28-generic-x86_64-with-glibc2.39",
    "torch": "2.13.0+cu130",
    "device": "cuda",
    "device_name": "NVIDIA GeForce RTX 3060",
    "seeds": 20,
    "epochs": 600,
    "command": "research/ensemble_specialization/experiment.py --seeds 20 --epochs 600 --device cuda --output research/ensemble_specialization/results",
    "pid": 3798532,
    "source_sha256": "129f7b7e6b4ad4ab77fa904716c6bbd3199a7f265cbc7ad07404fa88b1f9cf8e",
    "deterministic_algorithms": true
  },
  "scenarios": [
    {
      "name": "homogeneous",
      "heterogeneous": false,
      "train_size": 256,
      "noise": 0.5,
      "validation_size": 512,
      "test_size": 4096,
      "dimensions": 5
    },
    {
      "name": "heterogeneous",
      "heterogeneous": true,
      "train_size": 256,
      "noise": 0.5,
      "validation_size": 512,
      "test_size": 4096,
      "dimensions": 5
    },
    {
      "name": "heterogeneous_scarce",
      "heterogeneous": true,
      "train_size": 64,
      "noise": 0.5,
      "validation_size": 512,
      "test_size": 4096,
      "dimensions": 5
    },
    {
      "name": "heterogeneous_noisy",
      "heterogeneous": true,
      "train_size": 256,
      "noise": 1.5,
      "validation_size": 512,
      "test_size": 4096,
      "dimensions": 5
    }
  ],
  "summary": [
    {
      "scenario": "homogeneous",
      "model": "linear",
      "runs": 20,
      "mse_median": 0.25267330846682917,
      "mse_iqr": [
        0.24728459613450054,
        0.2569883440436709
      ],
      "mse_mean": 0.252556665040095,
      "mse_std": 0.007074367396551557,
      "paired_delta_vs_uniform_median": -0.0008418530617365222,
      "paired_delta_vs_uniform_95ci": [
        -0.0019624143387046294,
        -0.00019467037981210633
      ],
      "win_rate_vs_uniform": 0.75,
      "gate_accuracy_median": null
    },
    {
      "scenario": "homogeneous",
      "model": "mlp",
      "runs": 20,
      "mse_median": 0.30529205962064454,
      "mse_iqr": [
        0.2986946011280682,
        0.3098614529912088
      ],
      "mse_mean": 0.30514284187764057,
      "mse_std": 0.011807245194370453,
      "paired_delta_vs_uniform_median": 0.04896607019068788,
      "paired_delta_vs_uniform_95ci": [
        0.04400279411134862,
        0.05886404276232886
      ],
      "win_rate_vs_uniform": 0.0,
      "gate_accuracy_median": null
    },
    {
      "scenario": "homogeneous",
      "model": "uniform_ensemble",
      "runs": 20,
      "mse_median": 0.2539185125810701,
      "mse_iqr": [
        0.2483402377479372,
        0.2577708196412841
      ],
      "mse_mean": 0.25355812004804035,
      "mse_std": 0.006859690832453257,
      "paired_delta_vs_uniform_median": 0.0,
      "paired_delta_vs_uniform_95ci": [
        0.0,
        0.0
      ],
      "win_rate_vs_uniform": 0.0,
      "gate_accuracy_median": null
    },
    {
      "scenario": "homogeneous",
      "model": "aggregate_moe",
      "runs": 20,
      "mse_median": 0.33346249988585636,
      "mse_iqr": [
        0.31943414034040135,
        0.3647351794112058
      ],
      "mse_mean": 0.3503581235722296,
      "mse_std": 0.04848606930847931,
      "paired_delta_vs_uniform_median": 0.08116294474160338,
      "paired_delta_vs_uniform_95ci": [
        0.07059452835045348,
        0.10654816440907333
      ],
      "win_rate_vs_uniform": 0.0,
      "gate_accuracy_median": null
    },
    {
      "scenario": "homogeneous",
      "model": "sample_weight_moe",
      "runs": 20,
      "mse_median": 0.2627656750467243,
      "mse_iqr": [
        0.2577415560341591,
        0.26846439455070603
      ],
      "mse_mean": 0.2645087804166679,
      "mse_std": 0.00867999468663524,
      "paired_delta_vs_uniform_median": 0.010755819000084418,
      "paired_delta_vs_uniform_95ci": [
        0.0058287370306655956,
        0.015410730502668127
      ],
      "win_rate_vs_uniform": 0.0,
      "gate_accuracy_median": null
    },
    {
      "scenario": "homogeneous",
      "model": "sample_weight_warmstart",
      "runs": 20,
      "mse_median": 0.2628334921823474,
      "mse_iqr": [
        0.25863484325695874,
        0.26715983939681376
      ],
      "mse_mean": 0.26355070993769003,
      "mse_std": 0.0064050411677100275,
      "paired_delta_vs_uniform_median": 0.00932157409666666,
      "paired_delta_vs_uniform_95ci": [
        0.005129984591935546,
        0.013053909861410848
      ],
      "win_rate_vs_uniform": 0.05,
      "gate_accuracy_median": null
    },
    {
      "scenario": "homogeneous",
      "model": "oracle_partition",
      "runs": 20,
      "mse_median": 0.2539315675526641,
      "mse_iqr": [
        0.24895849307844384,
        0.25886050033187
      ],
      "mse_mean": 0.2537846670439187,
      "mse_std": 0.006976745973425599,
      "paired_delta_vs_uniform_median": 0.00011265593447537525,
      "paired_delta_vs_uniform_95ci": [
        -1.865780161473818e-05,
        0.0005100113439216319
      ],
      "win_rate_vs_uniform": 0.3,
      "gate_accuracy_median": null
    },
    {
      "scenario": "heterogeneous",
      "model": "linear",
      "runs": 20,
      "mse_median": 2.700204654606786,
      "mse_iqr": [
        2.6618036292133618,
        2.752796520652117
      ],
      "mse_mean": 2.7073898760246675,
      "mse_std": 0.07752222732231476,
      "paired_delta_vs_uniform_median": -0.0016044759235993578,
      "paired_delta_vs_uniform_95ci": [
        -0.006226693561569796,
        0.004853554763833756
      ],
      "win_rate_vs_uniform": 0.6,
      "gate_accuracy_median": null
    },
    {
      "scenario": "heterogeneous",
      "model": "mlp",
      "runs": 20,
      "mse_median": 0.7734887145038714,
      "mse_iqr": [
        0.7330459190702122,
        0.8242638437059381
      ],
      "mse_mean": 0.7701740078423661,
      "mse_std": 0.09596337383020165,
      "paired_delta_vs_uniform_median": -1.9219246788034954,
      "paired_delta_vs_uniform_95ci": [
        -1.967647636471224,
        -1.8802489167956824
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": null
    },
    {
      "scenario": "heterogeneous",
      "model": "uniform_ensemble",
      "runs": 20,
      "mse_median": 2.695570925057713,
      "mse_iqr": [
        2.6684454591353663,
        2.7517741353214733
      ],
      "mse_mean": 2.709561283216562,
      "mse_std": 0.0788356574748816,
      "paired_delta_vs_uniform_median": 0.0,
      "paired_delta_vs_uniform_95ci": [
        0.0,
        0.0
      ],
      "win_rate_vs_uniform": 0.0,
      "gate_accuracy_median": null
    },
    {
      "scenario": "heterogeneous",
      "model": "aggregate_moe",
      "runs": 20,
      "mse_median": 0.5656031486597977,
      "mse_iqr": [
        0.45600225276512396,
        1.6006589166598757
      ],
      "mse_mean": 0.9455393350533914,
      "mse_std": 0.5752222727657305,
      "paired_delta_vs_uniform_median": -2.0622190100384503,
      "paired_delta_vs_uniform_95ci": [
        -2.215259895854966,
        -1.1363967033782898
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": 0.9715576171875
    },
    {
      "scenario": "heterogeneous",
      "model": "sample_weight_moe",
      "runs": 20,
      "mse_median": 1.1321607438383388,
      "mse_iqr": [
        0.39668321869608025,
        1.800438089927308
      ],
      "mse_mean": 1.1162959841464568,
      "mse_std": 0.7218366250820388,
      "paired_delta_vs_uniform_median": -1.5389845246821232,
      "paired_delta_vs_uniform_95ci": [
        -2.2138949968095947,
        -0.9136508463980856
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": 0.7880859375
    },
    {
      "scenario": "heterogeneous",
      "model": "sample_weight_warmstart",
      "runs": 20,
      "mse_median": 0.5583063061081834,
      "mse_iqr": [
        0.4673343795621252,
        1.8187211534316154
      ],
      "mse_mean": 1.0184072550269376,
      "mse_std": 0.6884093933372045,
      "paired_delta_vs_uniform_median": -2.084861773316393,
      "paired_delta_vs_uniform_95ci": [
        -2.2260093050223055,
        -0.8963356783280951
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": 0.9739990234375
    },
    {
      "scenario": "heterogeneous",
      "model": "oracle_partition",
      "runs": 20,
      "mse_median": 0.2587370345033757,
      "mse_iqr": [
        0.25659963728288054,
        0.26386724221610763
      ],
      "mse_mean": 0.2602206927579226,
      "mse_std": 0.00670012162527671,
      "paired_delta_vs_uniform_median": -2.4320294351386407,
      "paired_delta_vs_uniform_95ci": [
        -2.489519615658107,
        -2.4099463811587807
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "model": "linear",
      "runs": 20,
      "mse_median": 2.893541265172675,
      "mse_iqr": [
        2.7655164434867885,
        3.1233946707264515
      ],
      "mse_mean": 2.956298157758451,
      "mse_std": 0.24368802748562662,
      "paired_delta_vs_uniform_median": -0.004721651948155214,
      "paired_delta_vs_uniform_95ci": [
        -0.011875053916083234,
        0.00436959253096636
      ],
      "win_rate_vs_uniform": 0.65,
      "gate_accuracy_median": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "model": "mlp",
      "runs": 20,
      "mse_median": 1.804527684418809,
      "mse_iqr": [
        1.4571878621289844,
        1.9290939181158517
      ],
      "mse_mean": 1.763559111335247,
      "mse_std": 0.36642137976571687,
      "paired_delta_vs_uniform_median": -1.2354529104604648,
      "paired_delta_vs_uniform_95ci": [
        -1.416784516711981,
        -1.0663868006366513
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "model": "uniform_ensemble",
      "runs": 20,
      "mse_median": 2.9504699313161415,
      "mse_iqr": [
        2.7700656714260545,
        3.130050628253755
      ],
      "mse_mean": 2.9681829965222915,
      "mse_std": 0.23968801304290624,
      "paired_delta_vs_uniform_median": 0.0,
      "paired_delta_vs_uniform_95ci": [
        0.0,
        0.0
      ],
      "win_rate_vs_uniform": 0.0,
      "gate_accuracy_median": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "model": "aggregate_moe",
      "runs": 20,
      "mse_median": 1.4764996365233367,
      "mse_iqr": [
        0.8549470329591891,
        2.2051260326140127
      ],
      "mse_mean": 1.568430441370675,
      "mse_std": 0.745527257223867,
      "paired_delta_vs_uniform_median": -1.35253747248272,
      "paired_delta_vs_uniform_95ci": [
        -2.0044223999586253,
        -0.8793951859047147
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": 0.8739013671875
    },
    {
      "scenario": "heterogeneous_scarce",
      "model": "sample_weight_moe",
      "runs": 20,
      "mse_median": 1.8043296082814688,
      "mse_iqr": [
        0.7398951621585609,
        2.163001371748746
      ],
      "mse_mean": 1.5502101222692635,
      "mse_std": 0.8078957839270903,
      "paired_delta_vs_uniform_median": -1.16140008488287,
      "paired_delta_vs_uniform_95ci": [
        -2.1626997843044,
        -0.6972743755531303
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": 0.7698974609375
    },
    {
      "scenario": "heterogeneous_scarce",
      "model": "sample_weight_warmstart",
      "runs": 20,
      "mse_median": 0.9362131613053616,
      "mse_iqr": [
        0.7517568862512596,
        2.2878054927136233
      ],
      "mse_mean": 1.4943347538164717,
      "mse_std": 0.8924624398709305,
      "paired_delta_vs_uniform_median": -1.9219133427596986,
      "paired_delta_vs_uniform_95ci": [
        -2.268962321481067,
        -0.741466951567297
      ],
      "win_rate_vs_uniform": 0.95,
      "gate_accuracy_median": 0.9395751953125
    },
    {
      "scenario": "heterogeneous_scarce",
      "model": "oracle_partition",
      "runs": 20,
      "mse_median": 0.3061572492643092,
      "mse_iqr": [
        0.29010960144645614,
        0.32424241453643127
      ],
      "mse_mean": 0.3069982293795641,
      "mse_std": 0.022479946971485858,
      "paired_delta_vs_uniform_median": -2.6199469187382913,
      "paired_delta_vs_uniform_95ci": [
        -2.794771794718021,
        -2.484595814624477
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "model": "linear",
      "runs": 20,
      "mse_median": 4.699335711131225,
      "mse_iqr": [
        4.66663020252743,
        4.802093529437657
      ],
      "mse_mean": 4.721803667505236,
      "mse_std": 0.13540903298324372,
      "paired_delta_vs_uniform_median": 0.0028067087338001784,
      "paired_delta_vs_uniform_95ci": [
        -0.016893904216885858,
        0.008455441213310433
      ],
      "win_rate_vs_uniform": 0.45,
      "gate_accuracy_median": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "model": "mlp",
      "runs": 20,
      "mse_median": 3.3212522117181673,
      "mse_iqr": [
        3.266034538326395,
        3.4583649537386676
      ],
      "mse_mean": 3.348312158528639,
      "mse_std": 0.147159302585422,
      "paired_delta_vs_uniform_median": -1.3588401533037435,
      "paired_delta_vs_uniform_95ci": [
        -1.4976270978696085,
        -1.2501957337495861
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "model": "uniform_ensemble",
      "runs": 20,
      "mse_median": 4.722727444843237,
      "mse_iqr": [
        4.678233165316912,
        4.796197934139116
      ],
      "mse_mean": 4.728697121192171,
      "mse_std": 0.14312299538265616,
      "paired_delta_vs_uniform_median": 0.0,
      "paired_delta_vs_uniform_95ci": [
        0.0,
        0.0
      ],
      "win_rate_vs_uniform": 0.0,
      "gate_accuracy_median": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "model": "aggregate_moe",
      "runs": 20,
      "mse_median": 3.1859811243625638,
      "mse_iqr": [
        2.631626780202231,
        3.8670226395572183
      ],
      "mse_mean": 3.3348222173778117,
      "mse_std": 0.6498964104617184,
      "paired_delta_vs_uniform_median": -1.4975601923352195,
      "paired_delta_vs_uniform_95ci": [
        -1.8817890845009992,
        -0.8370638363474974
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": 0.9376220703125
    },
    {
      "scenario": "heterogeneous_noisy",
      "model": "sample_weight_moe",
      "runs": 20,
      "mse_median": 3.1528623908925058,
      "mse_iqr": [
        2.6512571133154963,
        3.91603745681983
      ],
      "mse_mean": 3.291353425039844,
      "mse_std": 0.6734043679180759,
      "paired_delta_vs_uniform_median": -1.6511525007381853,
      "paired_delta_vs_uniform_95ci": [
        -1.972735528819095,
        -0.7974085519229694
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": 0.9224853515625
    },
    {
      "scenario": "heterogeneous_noisy",
      "model": "sample_weight_warmstart",
      "runs": 20,
      "mse_median": 2.9725131245490335,
      "mse_iqr": [
        2.672242939046767,
        3.950390298268107
      ],
      "mse_mean": 3.2520984543072955,
      "mse_std": 0.6219381904663753,
      "paired_delta_vs_uniform_median": -1.7260988413531921,
      "paired_delta_vs_uniform_95ci": [
        -1.9495059840889584,
        -0.827008194678577
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": 0.9451904296875
    },
    {
      "scenario": "heterogeneous_noisy",
      "model": "oracle_partition",
      "runs": 20,
      "mse_median": 2.328633237918191,
      "mse_iqr": [
        2.309396762527834,
        2.3748051493911326
      ],
      "mse_mean": 2.3419861541251517,
      "mse_std": 0.06030105549047026,
      "paired_delta_vs_uniform_median": -2.3452127864424805,
      "paired_delta_vs_uniform_95ci": [
        -2.4684788660906194,
        -2.316582389943795
      ],
      "win_rate_vs_uniform": 1.0,
      "gate_accuracy_median": null
    }
  ],
  "runs": [
    {
      "scenario": "homogeneous",
      "seed": 0,
      "model": "linear",
      "best_epoch": 127,
      "test_mse": 0.25706718629298686,
      "test_mse_normalized": 0.09893317520618439,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 0,
      "model": "mlp",
      "best_epoch": 60,
      "test_mse": 0.30610319193657937,
      "test_mse_normalized": 0.11780484765768051,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 0,
      "model": "uniform_ensemble",
      "best_epoch": 107,
      "test_mse": 0.2575642615896531,
      "test_mse_normalized": 0.099124476313591,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 0,
      "model": "aggregate_moe",
      "best_epoch": 59,
      "test_mse": 0.32445964677067,
      "test_mse_normalized": 0.12486939132213593,
      "gate_accuracy": null,
      "gate_entropy": 0.3274915814399719,
      "largest_expert_share": 0.6422233581542969,
      "utilization": [
        0.35777658224105835,
        0.6422233581542969
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 0,
      "model": "sample_weight_moe",
      "best_epoch": 150,
      "test_mse": 0.26306095491246695,
      "test_mse_normalized": 0.1012398973107338,
      "gate_accuracy": null,
      "gate_entropy": 0.05923081189393997,
      "largest_expert_share": 0.5137909650802612,
      "utilization": [
        0.486208975315094,
        0.5137909650802612
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 0,
      "model": "sample_weight_warmstart",
      "best_epoch": 133,
      "test_mse": 0.26577094066484036,
      "test_mse_normalized": 0.10228284448385239,
      "gate_accuracy": null,
      "gate_entropy": 0.11750295758247375,
      "largest_expert_share": 0.5208873748779297,
      "utilization": [
        0.4791126251220703,
        0.5208873748779297
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 0,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.25731026455381184,
      "test_mse_normalized": 0.09902672469615936,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 1,
      "model": "linear",
      "best_epoch": 86,
      "test_mse": 0.25755796867431285,
      "test_mse_normalized": 0.09218843281269073,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 1,
      "model": "mlp",
      "best_epoch": 84,
      "test_mse": 0.30170649789291903,
      "test_mse_normalized": 0.10799063742160797,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 1,
      "model": "uniform_ensemble",
      "best_epoch": 79,
      "test_mse": 0.2570035038659471,
      "test_mse_normalized": 0.09198997169733047,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 1,
      "model": "aggregate_moe",
      "best_epoch": 53,
      "test_mse": 0.3051516025433444,
      "test_mse_normalized": 0.10922375321388245,
      "gate_accuracy": null,
      "gate_entropy": 0.30300450325012207,
      "largest_expert_share": 0.7852144241333008,
      "utilization": [
        0.21478557586669922,
        0.7852144241333008
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 1,
      "model": "sample_weight_moe",
      "best_epoch": 198,
      "test_mse": 0.2680201744479338,
      "test_mse_normalized": 0.09593319892883301,
      "gate_accuracy": null,
      "gate_entropy": 0.0502493791282177,
      "largest_expert_share": 0.5100886225700378,
      "utilization": [
        0.48991137742996216,
        0.5100886225700378
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 1,
      "model": "sample_weight_warmstart",
      "best_epoch": 156,
      "test_mse": 0.26440675495573024,
      "test_mse_normalized": 0.0946398377418518,
      "gate_accuracy": null,
      "gate_entropy": 0.07605429738759995,
      "largest_expert_share": 0.5027656555175781,
      "utilization": [
        0.5027656555175781,
        0.49723437428474426
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 1,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.25906726530400026,
      "test_mse_normalized": 0.09272865951061249,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 2,
      "model": "linear",
      "best_epoch": 106,
      "test_mse": 0.2541860437572965,
      "test_mse_normalized": 0.09336239099502563,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 2,
      "model": "mlp",
      "best_epoch": 55,
      "test_mse": 0.30448092730470977,
      "test_mse_normalized": 0.11183567345142365,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 2,
      "model": "uniform_ensemble",
      "best_epoch": 43,
      "test_mse": 0.2584500015333138,
      "test_mse_normalized": 0.09492854028940201,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 2,
      "model": "aggregate_moe",
      "best_epoch": 43,
      "test_mse": 0.39801829222076823,
      "test_mse_normalized": 0.14619189500808716,
      "gate_accuracy": null,
      "gate_entropy": 0.4416944682598114,
      "largest_expert_share": 0.5203496217727661,
      "utilization": [
        0.5203496217727661,
        0.4796503782272339
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 2,
      "model": "sample_weight_moe",
      "best_epoch": 57,
      "test_mse": 0.2753307954515444,
      "test_mse_normalized": 0.10112884640693665,
      "gate_accuracy": null,
      "gate_entropy": 0.05773526430130005,
      "largest_expert_share": 0.5830481052398682,
      "utilization": [
        0.4169519245624542,
        0.5830481052398682
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 2,
      "model": "sample_weight_warmstart",
      "best_epoch": 131,
      "test_mse": 0.274300086274717,
      "test_mse_normalized": 0.10075026750564575,
      "gate_accuracy": null,
      "gate_entropy": 0.06886638700962067,
      "largest_expert_share": 0.530440092086792,
      "utilization": [
        0.469559907913208,
        0.530440092086792
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 2,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.25451392660257394,
      "test_mse_normalized": 0.09348282217979431,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 3,
      "model": "linear",
      "best_epoch": 75,
      "test_mse": 0.2563245868436006,
      "test_mse_normalized": 0.0904638022184372,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 3,
      "model": "mlp",
      "best_epoch": 76,
      "test_mse": 0.3118674853620913,
      "test_mse_normalized": 0.11006637662649155,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 3,
      "model": "uniform_ensemble",
      "best_epoch": 82,
      "test_mse": 0.25839049379617723,
      "test_mse_normalized": 0.09119291603565216,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 3,
      "model": "aggregate_moe",
      "best_epoch": 51,
      "test_mse": 0.32945946627550227,
      "test_mse_normalized": 0.11627505719661713,
      "gate_accuracy": null,
      "gate_entropy": 0.2904491424560547,
      "largest_expert_share": 0.6903976798057556,
      "utilization": [
        0.6903976798057556,
        0.3096023201942444
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 3,
      "model": "sample_weight_moe",
      "best_epoch": 197,
      "test_mse": 0.2740289407665003,
      "test_mse_normalized": 0.0967121422290802,
      "gate_accuracy": null,
      "gate_entropy": 0.040309660136699677,
      "largest_expert_share": 0.5616533756256104,
      "utilization": [
        0.5616533756256104,
        0.43834662437438965
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 3,
      "model": "sample_weight_warmstart",
      "best_epoch": 158,
      "test_mse": 0.27190713230554003,
      "test_mse_normalized": 0.09596329927444458,
      "gate_accuracy": null,
      "gate_entropy": 0.06902264058589935,
      "largest_expert_share": 0.5742666721343994,
      "utilization": [
        0.5742666721343994,
        0.4257332980632782
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 3,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.25879157867449326,
      "test_mse_normalized": 0.09133446961641312,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 4,
      "model": "linear",
      "best_epoch": 76,
      "test_mse": 0.24443465978588408,
      "test_mse_normalized": 0.10323944687843323,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 4,
      "model": "mlp",
      "best_epoch": 56,
      "test_mse": 0.30856037169275885,
      "test_mse_normalized": 0.130323588848114,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 4,
      "model": "uniform_ensemble",
      "best_epoch": 121,
      "test_mse": 0.24560742563547602,
      "test_mse_normalized": 0.10373477637767792,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 4,
      "model": "aggregate_moe",
      "best_epoch": 39,
      "test_mse": 0.3199906138620918,
      "test_mse_normalized": 0.13515126705169678,
      "gate_accuracy": null,
      "gate_entropy": 0.2555925250053406,
      "largest_expert_share": 0.8730391263961792,
      "utilization": [
        0.8730391263961792,
        0.1269608736038208
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 4,
      "model": "sample_weight_moe",
      "best_epoch": 227,
      "test_mse": 0.2551045675241062,
      "test_mse_normalized": 0.10774599015712738,
      "gate_accuracy": null,
      "gate_entropy": 0.03303280100226402,
      "largest_expert_share": 0.514039933681488,
      "utilization": [
        0.514039933681488,
        0.48596006631851196
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 4,
      "model": "sample_weight_warmstart",
      "best_epoch": 182,
      "test_mse": 0.2511759201183284,
      "test_mse_normalized": 0.10608668625354767,
      "gate_accuracy": null,
      "gate_entropy": 0.06413625925779343,
      "largest_expert_share": 0.5227094888687134,
      "utilization": [
        0.5227094888687134,
        0.4772905111312866
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 4,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.24565882961979826,
      "test_mse_normalized": 0.10375648736953735,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 5,
      "model": "linear",
      "best_epoch": 61,
      "test_mse": 0.2517068906438955,
      "test_mse_normalized": 0.0787118673324585,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 5,
      "model": "mlp",
      "best_epoch": 89,
      "test_mse": 0.2953240089051359,
      "test_mse_normalized": 0.09235148131847382,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 5,
      "model": "uniform_ensemble",
      "best_epoch": 74,
      "test_mse": 0.254099225529993,
      "test_mse_normalized": 0.07945998013019562,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 5,
      "model": "aggregate_moe",
      "best_epoch": 34,
      "test_mse": 0.33369852813901957,
      "test_mse_normalized": 0.10435166954994202,
      "gate_accuracy": null,
      "gate_entropy": 0.3924069404602051,
      "largest_expert_share": 0.784867525100708,
      "utilization": [
        0.784867525100708,
        0.215132474899292
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 5,
      "model": "sample_weight_moe",
      "best_epoch": 70,
      "test_mse": 0.25832851908273835,
      "test_mse_normalized": 0.08078253269195557,
      "gate_accuracy": null,
      "gate_entropy": 0.08201771974563599,
      "largest_expert_share": 0.5750735998153687,
      "utilization": [
        0.5750735998153687,
        0.42492642998695374
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 5,
      "model": "sample_weight_warmstart",
      "best_epoch": 138,
      "test_mse": 0.25847271199908795,
      "test_mse_normalized": 0.08082762360572815,
      "gate_accuracy": null,
      "gate_entropy": 0.08965857326984406,
      "largest_expert_share": 0.5483258962631226,
      "utilization": [
        0.5483258962631226,
        0.45167410373687744
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 5,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.25389308588817927,
      "test_mse_normalized": 0.07939551770687103,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 6,
      "model": "linear",
      "best_epoch": 62,
      "test_mse": 0.27010932470321203,
      "test_mse_normalized": 0.08913219720125198,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 6,
      "model": "mlp",
      "best_epoch": 64,
      "test_mse": 0.309192775534248,
      "test_mse_normalized": 0.10202917456626892,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 6,
      "model": "uniform_ensemble",
      "best_epoch": 69,
      "test_mse": 0.2717741955404827,
      "test_mse_normalized": 0.08968158066272736,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 6,
      "model": "aggregate_moe",
      "best_epoch": 36,
      "test_mse": 0.3431693535526336,
      "test_mse_normalized": 0.11324095726013184,
      "gate_accuracy": null,
      "gate_entropy": 0.29614630341529846,
      "largest_expert_share": 0.9016537666320801,
      "utilization": [
        0.9016537666320801,
        0.09834626317024231
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 6,
      "model": "sample_weight_moe",
      "best_epoch": 98,
      "test_mse": 0.2736443952546285,
      "test_mse_normalized": 0.09029871970415115,
      "gate_accuracy": null,
      "gate_entropy": 0.06960035860538483,
      "largest_expert_share": 0.5110636353492737,
      "utilization": [
        0.5110636353492737,
        0.4889363944530487
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 6,
      "model": "sample_weight_warmstart",
      "best_epoch": 144,
      "test_mse": 0.2757954940843674,
      "test_mse_normalized": 0.09100855141878128,
      "gate_accuracy": null,
      "gate_entropy": 0.0831640362739563,
      "largest_expert_share": 0.503659188747406,
      "utilization": [
        0.503659188747406,
        0.496340811252594
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 6,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.27180490230016496,
      "test_mse_normalized": 0.08969171345233917,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 7,
      "model": "linear",
      "best_epoch": 152,
      "test_mse": 0.26003377522592813,
      "test_mse_normalized": 0.0905403196811676,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 7,
      "model": "mlp",
      "best_epoch": 65,
      "test_mse": 0.2940017109666331,
      "test_mse_normalized": 0.10236750543117523,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 7,
      "model": "uniform_ensemble",
      "best_epoch": 86,
      "test_mse": 0.25900903538129577,
      "test_mse_normalized": 0.09018351882696152,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 7,
      "model": "aggregate_moe",
      "best_epoch": 65,
      "test_mse": 0.3013264535433569,
      "test_mse_normalized": 0.10491788387298584,
      "gate_accuracy": null,
      "gate_entropy": 0.19547119736671448,
      "largest_expert_share": 0.8947020173072815,
      "utilization": [
        0.1052979826927185,
        0.8947020173072815
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 7,
      "model": "sample_weight_moe",
      "best_epoch": 98,
      "test_mse": 0.26113758153104033,
      "test_mse_normalized": 0.09092465043067932,
      "gate_accuracy": null,
      "gate_entropy": 0.05741932988166809,
      "largest_expert_share": 0.5134855508804321,
      "utilization": [
        0.48651444911956787,
        0.5134855508804321
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 7,
      "model": "sample_weight_warmstart",
      "best_epoch": 122,
      "test_mse": 0.25748946146397195,
      "test_mse_normalized": 0.08965442329645157,
      "gate_accuracy": null,
      "gate_entropy": 0.1267564445734024,
      "largest_expert_share": 0.5218469500541687,
      "utilization": [
        0.4781530499458313,
        0.5218469500541687
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 7,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2601405095980183,
      "test_mse_normalized": 0.09057748317718506,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 8,
      "model": "linear",
      "best_epoch": 80,
      "test_mse": 0.2473760751190366,
      "test_mse_normalized": 0.08561702817678452,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 8,
      "model": "mlp",
      "best_epoch": 57,
      "test_mse": 0.3204193267164309,
      "test_mse_normalized": 0.11089734733104706,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 8,
      "model": "uniform_ensemble",
      "best_epoch": 64,
      "test_mse": 0.24857210514600028,
      "test_mse_normalized": 0.086030974984169,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 8,
      "model": "aggregate_moe",
      "best_epoch": 55,
      "test_mse": 0.46613138569355894,
      "test_mse_normalized": 0.16132839024066925,
      "gate_accuracy": null,
      "gate_entropy": 0.7183681726455688,
      "largest_expert_share": 0.5282888412475586,
      "utilization": [
        0.4717111587524414,
        0.5282888412475586
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 8,
      "model": "sample_weight_moe",
      "best_epoch": 236,
      "test_mse": 0.26659356242587584,
      "test_mse_normalized": 0.0922682136297226,
      "gate_accuracy": null,
      "gate_entropy": 0.03844185918569565,
      "largest_expert_share": 0.5054364204406738,
      "utilization": [
        0.5054364204406738,
        0.49456357955932617
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 8,
      "model": "sample_weight_warmstart",
      "best_epoch": 177,
      "test_mse": 0.2613051890493949,
      "test_mse_normalized": 0.09043790400028229,
      "gate_accuracy": null,
      "gate_entropy": 0.07866530120372772,
      "largest_expert_share": 0.5427978038787842,
      "utilization": [
        0.4572021961212158,
        0.5427978038787842
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 8,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.24879482561719246,
      "test_mse_normalized": 0.08610805869102478,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 9,
      "model": "linear",
      "best_epoch": 52,
      "test_mse": 0.25898674636599095,
      "test_mse_normalized": 0.08588449656963348,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 9,
      "model": "mlp",
      "best_epoch": 56,
      "test_mse": 0.33648955959380006,
      "test_mse_normalized": 0.11158577352762222,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 9,
      "model": "uniform_ensemble",
      "best_epoch": 97,
      "test_mse": 0.2592780137832307,
      "test_mse_normalized": 0.085981085896492,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 9,
      "model": "aggregate_moe",
      "best_epoch": 100,
      "test_mse": 0.4096190478215673,
      "test_mse_normalized": 0.13583678007125854,
      "gate_accuracy": null,
      "gate_entropy": 0.7786237001419067,
      "largest_expert_share": 0.7417071461677551,
      "utilization": [
        0.7417071461677551,
        0.2582928538322449
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 9,
      "model": "sample_weight_moe",
      "best_epoch": 25,
      "test_mse": 0.2888342873247566,
      "test_mse_normalized": 0.09578245878219604,
      "gate_accuracy": null,
      "gate_entropy": 0.03752297908067703,
      "largest_expert_share": 0.9866876602172852,
      "utilization": [
        0.9866876602172852,
        0.013312385417521
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 9,
      "model": "sample_weight_warmstart",
      "best_epoch": 161,
      "test_mse": 0.26220637020317733,
      "test_mse_normalized": 0.08695217967033386,
      "gate_accuracy": null,
      "gate_entropy": 0.17051377892494202,
      "largest_expert_share": 0.5098092555999756,
      "utilization": [
        0.4901907444000244,
        0.5098092555999756
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 9,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.25935085310494393,
      "test_mse_normalized": 0.08600524067878723,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 10,
      "model": "linear",
      "best_epoch": 74,
      "test_mse": 0.25363972628976283,
      "test_mse_normalized": 0.09139228612184525,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 10,
      "model": "mlp",
      "best_epoch": 81,
      "test_mse": 0.3005589189606317,
      "test_mse_normalized": 0.10829836130142212,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 10,
      "model": "uniform_ensemble",
      "best_epoch": 72,
      "test_mse": 0.2537377996321473,
      "test_mse_normalized": 0.09142762422561646,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 10,
      "model": "aggregate_moe",
      "best_epoch": 46,
      "test_mse": 0.2883392038164331,
      "test_mse_normalized": 0.10389531403779984,
      "gate_accuracy": null,
      "gate_entropy": 0.37845924496650696,
      "largest_expert_share": 0.5623881220817566,
      "utilization": [
        0.5623881220817566,
        0.437611848115921
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 10,
      "model": "sample_weight_moe",
      "best_epoch": 95,
      "test_mse": 0.2573658309431525,
      "test_mse_normalized": 0.09273488819599152,
      "gate_accuracy": null,
      "gate_entropy": 0.06127461418509483,
      "largest_expert_share": 0.5324598550796509,
      "utilization": [
        0.4675402045249939,
        0.5324598550796509
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 10,
      "model": "sample_weight_warmstart",
      "best_epoch": 155,
      "test_mse": 0.25842927433316604,
      "test_mse_normalized": 0.09311807155609131,
      "gate_accuracy": null,
      "gate_entropy": 0.09765024483203888,
      "largest_expert_share": 0.5073678493499756,
      "utilization": [
        0.5073678493499756,
        0.49263209104537964
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 10,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2539700492171489,
      "test_mse_normalized": 0.0915113091468811,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 11,
      "model": "linear",
      "best_epoch": 83,
      "test_mse": 0.2516043380812391,
      "test_mse_normalized": 0.08424253016710281,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 11,
      "model": "mlp",
      "best_epoch": 54,
      "test_mse": 0.3179392514951581,
      "test_mse_normalized": 0.10645288228988647,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 11,
      "model": "uniform_ensemble",
      "best_epoch": 74,
      "test_mse": 0.250499884758513,
      "test_mse_normalized": 0.0838727355003357,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 11,
      "model": "aggregate_moe",
      "best_epoch": 35,
      "test_mse": 0.33322647163269314,
      "test_mse_normalized": 0.11157137155532837,
      "gate_accuracy": null,
      "gate_entropy": 0.780225396156311,
      "largest_expert_share": 0.5649292469024658,
      "utilization": [
        0.43507078289985657,
        0.5649292469024658
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 11,
      "model": "sample_weight_moe",
      "best_epoch": 77,
      "test_mse": 0.2666573962234273,
      "test_mse_normalized": 0.08928261697292328,
      "gate_accuracy": null,
      "gate_entropy": 0.06108737736940384,
      "largest_expert_share": 0.5261574387550354,
      "utilization": [
        0.5261574387550354,
        0.4738425612449646
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 11,
      "model": "sample_weight_warmstart",
      "best_epoch": 137,
      "test_mse": 0.2630030627443966,
      "test_mse_normalized": 0.08805906772613525,
      "gate_accuracy": null,
      "gate_entropy": 0.13377957046031952,
      "largest_expert_share": 0.5022639036178589,
      "utilization": [
        0.5022639036178589,
        0.4977360963821411
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 11,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.25197755531833466,
      "test_mse_normalized": 0.08436749130487442,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 12,
      "model": "linear",
      "best_epoch": 78,
      "test_mse": 0.24832032448567617,
      "test_mse_normalized": 0.0832996815443039,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 12,
      "model": "mlp",
      "best_epoch": 63,
      "test_mse": 0.2909472297494733,
      "test_mse_normalized": 0.09759898483753204,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 12,
      "model": "uniform_ensemble",
      "best_epoch": 63,
      "test_mse": 0.24764463555374797,
      "test_mse_normalized": 0.08307301998138428,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 12,
      "model": "aggregate_moe",
      "best_epoch": 54,
      "test_mse": 0.3177647197753299,
      "test_mse_normalized": 0.106594979763031,
      "gate_accuracy": null,
      "gate_entropy": 0.40337371826171875,
      "largest_expert_share": 0.5136244893074036,
      "utilization": [
        0.48637551069259644,
        0.5136244893074036
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 12,
      "model": "sample_weight_moe",
      "best_epoch": 106,
      "test_mse": 0.25380541629226533,
      "test_mse_normalized": 0.0851396694779396,
      "gate_accuracy": null,
      "gate_entropy": 0.05509260296821594,
      "largest_expert_share": 0.5343931317329407,
      "utilization": [
        0.5343931317329407,
        0.4656068682670593
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 12,
      "model": "sample_weight_warmstart",
      "best_epoch": 171,
      "test_mse": 0.2589752739587304,
      "test_mse_normalized": 0.08687391132116318,
      "gate_accuracy": null,
      "gate_entropy": 0.08172735571861267,
      "largest_expert_share": 0.5402061343193054,
      "utilization": [
        0.5402061343193054,
        0.4597938358783722
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 12,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.24901304889886097,
      "test_mse_normalized": 0.08353205770254135,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 13,
      "model": "linear",
      "best_epoch": 50,
      "test_mse": 0.2398586256878845,
      "test_mse_normalized": 0.08335313200950623,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 13,
      "model": "mlp",
      "best_epoch": 98,
      "test_mse": 0.31242942993873285,
      "test_mse_normalized": 0.10857217013835907,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 13,
      "model": "uniform_ensemble",
      "best_epoch": 38,
      "test_mse": 0.24209082344298585,
      "test_mse_normalized": 0.08412884175777435,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 13,
      "model": "aggregate_moe",
      "best_epoch": 40,
      "test_mse": 0.3319304261440561,
      "test_mse_normalized": 0.1153489500284195,
      "gate_accuracy": null,
      "gate_entropy": 0.2462974488735199,
      "largest_expert_share": 0.630407452583313,
      "utilization": [
        0.369592547416687,
        0.630407452583313
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 13,
      "model": "sample_weight_moe",
      "best_epoch": 85,
      "test_mse": 0.25727383747799903,
      "test_mse_normalized": 0.08940508216619492,
      "gate_accuracy": null,
      "gate_entropy": 0.07146309316158295,
      "largest_expert_share": 0.506476879119873,
      "utilization": [
        0.506476879119873,
        0.49352315068244934
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 13,
      "model": "sample_weight_warmstart",
      "best_epoch": 120,
      "test_mse": 0.2716075297899914,
      "test_mse_normalized": 0.09438617527484894,
      "gate_accuracy": null,
      "gate_entropy": 0.10984793305397034,
      "largest_expert_share": 0.5443969368934631,
      "utilization": [
        0.45560306310653687,
        0.5443969368934631
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 13,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.24211084834552793,
      "test_mse_normalized": 0.08413580060005188,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 14,
      "model": "linear",
      "best_epoch": 118,
      "test_mse": 0.2569620632938989,
      "test_mse_normalized": 0.0915868729352951,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 14,
      "model": "mlp",
      "best_epoch": 66,
      "test_mse": 0.2871284662015776,
      "test_mse_normalized": 0.10233883559703827,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 14,
      "model": "uniform_ensemble",
      "best_epoch": 111,
      "test_mse": 0.2568955891096084,
      "test_mse_normalized": 0.09156318008899689,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 14,
      "model": "aggregate_moe",
      "best_epoch": 63,
      "test_mse": 0.3430040316334334,
      "test_mse_normalized": 0.12225410342216492,
      "gate_accuracy": null,
      "gate_entropy": 0.8401464223861694,
      "largest_expert_share": 0.5536712408065796,
      "utilization": [
        0.5536712408065796,
        0.4463287591934204
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 14,
      "model": "sample_weight_moe",
      "best_epoch": 114,
      "test_mse": 0.25783354403071423,
      "test_mse_normalized": 0.09189748764038086,
      "gate_accuracy": null,
      "gate_entropy": 0.05506444722414017,
      "largest_expert_share": 0.5176159143447876,
      "utilization": [
        0.5176159143447876,
        0.4823840260505676
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 14,
      "model": "sample_weight_warmstart",
      "best_epoch": 138,
      "test_mse": 0.2586888870095823,
      "test_mse_normalized": 0.09220235049724579,
      "gate_accuracy": null,
      "gate_entropy": 0.1183461844921112,
      "largest_expert_share": 0.5139997005462646,
      "utilization": [
        0.48600026965141296,
        0.5139997005462646
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 14,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2570480616568459,
      "test_mse_normalized": 0.09161752462387085,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 15,
      "model": "linear",
      "best_epoch": 121,
      "test_mse": 0.24971981310526595,
      "test_mse_normalized": 0.0906970202922821,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 15,
      "model": "mlp",
      "best_epoch": 52,
      "test_mse": 0.2871944383265883,
      "test_mse_normalized": 0.104307621717453,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 15,
      "model": "uniform_ensemble",
      "best_epoch": 56,
      "test_mse": 0.25168538138821905,
      "test_mse_normalized": 0.09141090512275696,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 15,
      "model": "aggregate_moe",
      "best_epoch": 60,
      "test_mse": 0.423972725792392,
      "test_mse_normalized": 0.15398482978343964,
      "gate_accuracy": null,
      "gate_entropy": 0.35299649834632874,
      "largest_expert_share": 0.6018936038017273,
      "utilization": [
        0.6018936038017273,
        0.3981063961982727
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 15,
      "model": "sample_weight_moe",
      "best_epoch": 151,
      "test_mse": 0.2654066956867553,
      "test_mse_normalized": 0.09639441967010498,
      "gate_accuracy": null,
      "gate_entropy": 0.0574059896171093,
      "largest_expert_share": 0.5668840408325195,
      "utilization": [
        0.5668840408325195,
        0.43311601877212524
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 15,
      "model": "sample_weight_warmstart",
      "best_epoch": 212,
      "test_mse": 0.2672730993512752,
      "test_mse_normalized": 0.0970722883939743,
      "gate_accuracy": null,
      "gate_entropy": 0.08354894816875458,
      "largest_expert_share": 0.5479897260665894,
      "utilization": [
        0.5479897260665894,
        0.45201027393341064
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 15,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.24982392159482913,
      "test_mse_normalized": 0.09073483198881149,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 16,
      "model": "linear",
      "best_epoch": 60,
      "test_mse": 0.2428116260427828,
      "test_mse_normalized": 0.09061133116483688,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 16,
      "model": "mlp",
      "best_epoch": 51,
      "test_mse": 0.3079587262224692,
      "test_mse_normalized": 0.11492262780666351,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 16,
      "model": "uniform_ensemble",
      "best_epoch": 65,
      "test_mse": 0.24477088643723896,
      "test_mse_normalized": 0.09134247899055481,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 16,
      "model": "aggregate_moe",
      "best_epoch": 49,
      "test_mse": 0.35174750623547413,
      "test_mse_normalized": 0.13126352429389954,
      "gate_accuracy": null,
      "gate_entropy": 0.4443909525871277,
      "largest_expert_share": 0.6419990062713623,
      "utilization": [
        0.6419990062713623,
        0.3580009937286377
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 16,
      "model": "sample_weight_moe",
      "best_epoch": 176,
      "test_mse": 0.26247039518098175,
      "test_mse_normalized": 0.09794750064611435,
      "gate_accuracy": null,
      "gate_entropy": 0.03342048451304436,
      "largest_expert_share": 0.5090597867965698,
      "utilization": [
        0.4909402132034302,
        0.5090597867965698
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 16,
      "model": "sample_weight_warmstart",
      "best_epoch": 172,
      "test_mse": 0.2640999875979659,
      "test_mse_normalized": 0.09855562448501587,
      "gate_accuracy": null,
      "gate_entropy": 0.06117255240678787,
      "largest_expert_share": 0.5076421499252319,
      "utilization": [
        0.49235785007476807,
        0.5076421499252319
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 16,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2447135459314674,
      "test_mse_normalized": 0.09132108092308044,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 17,
      "model": "linear",
      "best_epoch": 69,
      "test_mse": 0.2569040231892346,
      "test_mse_normalized": 0.09960294514894485,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 17,
      "model": "mlp",
      "best_epoch": 72,
      "test_mse": 0.3066720636155818,
      "test_mse_normalized": 0.11889825761318207,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 17,
      "model": "uniform_ensemble",
      "best_epoch": 65,
      "test_mse": 0.2572788535811323,
      "test_mse_normalized": 0.09974826872348785,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 17,
      "model": "aggregate_moe",
      "best_epoch": 53,
      "test_mse": 0.3065044516864287,
      "test_mse_normalized": 0.1188332736492157,
      "gate_accuracy": null,
      "gate_entropy": 0.23966848850250244,
      "largest_expert_share": 0.8415083885192871,
      "utilization": [
        0.15849164128303528,
        0.8415083885192871
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 17,
      "model": "sample_weight_moe",
      "best_epoch": 168,
      "test_mse": 0.2697970548590226,
      "test_mse_normalized": 0.10460163652896881,
      "gate_accuracy": null,
      "gate_entropy": 0.03909089416265488,
      "largest_expert_share": 0.5322542786598206,
      "utilization": [
        0.46774566173553467,
        0.5322542786598206
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 17,
      "model": "sample_weight_warmstart",
      "best_epoch": 161,
      "test_mse": 0.26712208607865995,
      "test_mse_normalized": 0.1035645380616188,
      "gate_accuracy": null,
      "gate_entropy": 0.07348774373531342,
      "largest_expert_share": 0.5327597856521606,
      "utilization": [
        0.46724027395248413,
        0.5327597856521606
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 17,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2604515271715312,
      "test_mse_normalized": 0.10097832977771759,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 18,
      "model": "linear",
      "best_epoch": 69,
      "test_mse": 0.2470101591808924,
      "test_mse_normalized": 0.1130896732211113,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 18,
      "model": "mlp",
      "best_epoch": 71,
      "test_mse": 0.29981813186904566,
      "test_mse_normalized": 0.13726696372032166,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 18,
      "model": "uniform_ensemble",
      "best_epoch": 108,
      "test_mse": 0.2475210994547735,
      "test_mse_normalized": 0.11332359910011292,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 18,
      "model": "aggregate_moe",
      "best_epoch": 20,
      "test_mse": 0.353640808474685,
      "test_mse_normalized": 0.161908820271492,
      "gate_accuracy": null,
      "gate_entropy": 0.1635296642780304,
      "largest_expert_share": 0.9575292468070984,
      "utilization": [
        0.9575292468070984,
        0.042470742017030716
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 18,
      "model": "sample_weight_moe",
      "best_epoch": 204,
      "test_mse": 0.25801606687295564,
      "test_mse_normalized": 0.11812855303287506,
      "gate_accuracy": null,
      "gate_entropy": 0.04799140244722366,
      "largest_expert_share": 0.523415207862854,
      "utilization": [
        0.523415207862854,
        0.4765848219394684
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 18,
      "model": "sample_weight_warmstart",
      "best_epoch": 130,
      "test_mse": 0.25632101515057915,
      "test_mse_normalized": 0.11735250055789948,
      "gate_accuracy": null,
      "gate_entropy": 0.16865891218185425,
      "largest_expert_share": 0.5463103652000427,
      "utilization": [
        0.5463103652000427,
        0.4536896049976349
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 18,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.24735061787025336,
      "test_mse_normalized": 0.11324554681777954,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 19,
      "model": "linear",
      "best_epoch": 40,
      "test_mse": 0.24651934403311868,
      "test_mse_normalized": 0.08790396898984909,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 19,
      "model": "mlp",
      "best_epoch": 56,
      "test_mse": 0.304064325268246,
      "test_mse_normalized": 0.1084233820438385,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 19,
      "model": "uniform_ensemble",
      "best_epoch": 41,
      "test_mse": 0.24928918580087112,
      "test_mse_normalized": 0.0888916403055191,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "homogeneous",
      "seed": 19,
      "model": "aggregate_moe",
      "best_epoch": 37,
      "test_mse": 0.4260077358311535,
      "test_mse_normalized": 0.15190601348876953,
      "gate_accuracy": null,
      "gate_entropy": 0.5865135192871094,
      "largest_expert_share": 0.6403005123138428,
      "utilization": [
        0.6403005123138428,
        0.3596994876861572
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 19,
      "model": "sample_weight_moe",
      "best_epoch": 81,
      "test_mse": 0.25746559204449376,
      "test_mse_normalized": 0.09180718660354614,
      "gate_accuracy": null,
      "gate_entropy": 0.0645471066236496,
      "largest_expert_share": 0.5743846297264099,
      "utilization": [
        0.5743846297264099,
        0.4256153702735901
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 19,
      "model": "sample_weight_warmstart",
      "best_epoch": 145,
      "test_mse": 0.26266392162029817,
      "test_mse_normalized": 0.09366080909967422,
      "gate_accuracy": null,
      "gate_entropy": 0.05416885018348694,
      "largest_expert_share": 0.518014132976532,
      "utilization": [
        0.481985867023468,
        0.518014132976532
      ]
    },
    {
      "scenario": "homogeneous",
      "seed": 19,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.24990812361039835,
      "test_mse_normalized": 0.08911234140396118,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 0,
      "model": "linear",
      "best_epoch": 73,
      "test_mse": 2.831816551898585,
      "test_mse_normalized": 0.9297353029251099,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 0,
      "model": "mlp",
      "best_epoch": 70,
      "test_mse": 0.9011898073042084,
      "test_mse_normalized": 0.29587650299072266,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 0,
      "model": "uniform_ensemble",
      "best_epoch": 76,
      "test_mse": 2.83919238942972,
      "test_mse_normalized": 0.9321569204330444,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 0,
      "model": "aggregate_moe",
      "best_epoch": 101,
      "test_mse": 0.5782478346560956,
      "test_mse_normalized": 0.18984895944595337,
      "gate_accuracy": 0.9775390625,
      "gate_entropy": 0.05035644769668579,
      "largest_expert_share": 0.5159906148910522,
      "utilization": [
        0.48400941491127014,
        0.5159906148910522
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 0,
      "model": "sample_weight_moe",
      "best_epoch": 179,
      "test_mse": 0.45396003421420733,
      "test_mse_normalized": 0.14904308319091797,
      "gate_accuracy": 0.983154296875,
      "gate_entropy": 0.01596970297396183,
      "largest_expert_share": 0.5083574652671814,
      "utilization": [
        0.4916425347328186,
        0.5083574652671814
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 0,
      "model": "sample_weight_warmstart",
      "best_epoch": 109,
      "test_mse": 0.5794004681227979,
      "test_mse_normalized": 0.19022738933563232,
      "gate_accuracy": 0.976806640625,
      "gate_entropy": 0.04278107360005379,
      "largest_expert_share": 0.5182414650917053,
      "utilization": [
        0.4817585349082947,
        0.5182414650917053
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 0,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2588625005600353,
      "test_mse_normalized": 0.08498912304639816,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 1,
      "model": "linear",
      "best_epoch": 57,
      "test_mse": 2.6875714392568053,
      "test_mse_normalized": 0.7764409184455872,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 1,
      "model": "mlp",
      "best_epoch": 79,
      "test_mse": 0.776754172235258,
      "test_mse_normalized": 0.2244047224521637,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 1,
      "model": "uniform_ensemble",
      "best_epoch": 28,
      "test_mse": 2.6833215477411185,
      "test_mse_normalized": 0.7752131223678589,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 1,
      "model": "aggregate_moe",
      "best_epoch": 81,
      "test_mse": 0.44894639599102926,
      "test_mse_normalized": 0.12970086932182312,
      "gate_accuracy": 0.975341796875,
      "gate_entropy": 0.048135560005903244,
      "largest_expert_share": 0.5061987638473511,
      "utilization": [
        0.5061987638473511,
        0.4938012361526489
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 1,
      "model": "sample_weight_moe",
      "best_epoch": 116,
      "test_mse": 0.483354548453912,
      "test_mse_normalized": 0.1396414041519165,
      "gate_accuracy": 0.97314453125,
      "gate_entropy": 0.0189647413790226,
      "largest_expert_share": 0.5099669694900513,
      "utilization": [
        0.5099669694900513,
        0.49003297090530396
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 1,
      "model": "sample_weight_warmstart",
      "best_epoch": 115,
      "test_mse": 0.48525801459905143,
      "test_mse_normalized": 0.14019131660461426,
      "gate_accuracy": 0.97314453125,
      "gate_entropy": 0.03201736882328987,
      "largest_expert_share": 0.5065553188323975,
      "utilization": [
        0.5065553188323975,
        0.4934447109699249
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 1,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.26544929897318614,
      "test_mse_normalized": 0.07668845355510712,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 2,
      "model": "linear",
      "best_epoch": 55,
      "test_mse": 2.7735300830212566,
      "test_mse_normalized": 1.0456645488739014,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 2,
      "model": "mlp",
      "best_epoch": 63,
      "test_mse": 0.5922884152698265,
      "test_mse_normalized": 0.22330206632614136,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 2,
      "model": "uniform_ensemble",
      "best_epoch": 42,
      "test_mse": 2.7474698714637618,
      "test_mse_normalized": 1.0358394384384155,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 2,
      "model": "aggregate_moe",
      "best_epoch": 60,
      "test_mse": 1.5998637665621303,
      "test_mse_normalized": 0.6031738519668579,
      "gate_accuracy": 0.50634765625,
      "gate_entropy": 0.32605844736099243,
      "largest_expert_share": 0.5858860015869141,
      "utilization": [
        0.41411396861076355,
        0.5858860015869141
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 2,
      "model": "sample_weight_moe",
      "best_epoch": 41,
      "test_mse": 1.8311881028729964,
      "test_mse_normalized": 0.6903867721557617,
      "gate_accuracy": 0.51904296875,
      "gate_entropy": 0.05846356600522995,
      "largest_expert_share": 0.5298351049423218,
      "utilization": [
        0.47016486525535583,
        0.5298351049423218
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 2,
      "model": "sample_weight_warmstart",
      "best_epoch": 136,
      "test_mse": 1.8534356737655926,
      "test_mse_normalized": 0.6987744569778442,
      "gate_accuracy": 0.57275390625,
      "gate_entropy": 0.056186284869909286,
      "largest_expert_share": 0.5324712991714478,
      "utilization": [
        0.46752870082855225,
        0.5324712991714478
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 2,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2587892496799261,
      "test_mse_normalized": 0.09756762534379959,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 3,
      "model": "linear",
      "best_epoch": 70,
      "test_mse": 2.712785861267537,
      "test_mse_normalized": 0.7815321087837219,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 3,
      "model": "mlp",
      "best_epoch": 67,
      "test_mse": 0.854394224565166,
      "test_mse_normalized": 0.24614420533180237,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 3,
      "model": "uniform_ensemble",
      "best_epoch": 34,
      "test_mse": 2.704281672751508,
      "test_mse_normalized": 0.7790821194648743,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 3,
      "model": "aggregate_moe",
      "best_epoch": 260,
      "test_mse": 0.7165768297350184,
      "test_mse_normalized": 0.20644010603427887,
      "gate_accuracy": 0.97119140625,
      "gate_entropy": 0.03809821605682373,
      "largest_expert_share": 0.5050573348999023,
      "utilization": [
        0.49494266510009766,
        0.5050573348999023
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 3,
      "model": "sample_weight_moe",
      "best_epoch": 129,
      "test_mse": 1.8812868843051966,
      "test_mse_normalized": 0.5419838428497314,
      "gate_accuracy": 0.60302734375,
      "gate_entropy": 0.02836861088871956,
      "largest_expert_share": 0.5324530601501465,
      "utilization": [
        0.4675469994544983,
        0.5324530601501465
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 3,
      "model": "sample_weight_warmstart",
      "best_epoch": 205,
      "test_mse": 0.6206794526225979,
      "test_mse_normalized": 0.17881283164024353,
      "gate_accuracy": 0.9775390625,
      "gate_entropy": 0.02500120736658573,
      "largest_expert_share": 0.505691647529602,
      "utilization": [
        0.4943082928657532,
        0.505691647529602
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 3,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.276315963698478,
      "test_mse_normalized": 0.07960443943738937,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 4,
      "model": "linear",
      "best_epoch": 34,
      "test_mse": 2.67695089367222,
      "test_mse_normalized": 0.854108452796936,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 4,
      "model": "mlp",
      "best_epoch": 79,
      "test_mse": 0.5339877909528978,
      "test_mse_normalized": 0.17037424445152283,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 4,
      "model": "uniform_ensemble",
      "best_epoch": 27,
      "test_mse": 2.693601551462947,
      "test_mse_normalized": 0.8594210147857666,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 4,
      "model": "aggregate_moe",
      "best_epoch": 47,
      "test_mse": 1.5684142496079991,
      "test_mse_normalized": 0.5004185438156128,
      "gate_accuracy": 0.514404296875,
      "gate_entropy": 0.19802583754062653,
      "largest_expert_share": 0.522135853767395,
      "utilization": [
        0.47786420583724976,
        0.522135853767395
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 4,
      "model": "sample_weight_moe",
      "best_epoch": 216,
      "test_mse": 1.75729394631178,
      "test_mse_normalized": 0.5606825351715088,
      "gate_accuracy": 0.538818359375,
      "gate_entropy": 0.03338921070098877,
      "largest_expert_share": 0.5400288105010986,
      "utilization": [
        0.459971159696579,
        0.5400288105010986
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 4,
      "model": "sample_weight_warmstart",
      "best_epoch": 100,
      "test_mse": 1.812652666433419,
      "test_mse_normalized": 0.5783452987670898,
      "gate_accuracy": 0.57421875,
      "gate_entropy": 0.3105604946613312,
      "largest_expert_share": 0.5418288707733154,
      "utilization": [
        0.5418288707733154,
        0.45817118883132935
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 4,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2586848193268253,
      "test_mse_normalized": 0.08253602683544159,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 5,
      "model": "linear",
      "best_epoch": 76,
      "test_mse": 2.715103011586327,
      "test_mse_normalized": 0.7560945153236389,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 5,
      "model": "mlp",
      "best_epoch": 79,
      "test_mse": 0.7365494579521515,
      "test_mse_normalized": 0.2051122933626175,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 5,
      "model": "uniform_ensemble",
      "best_epoch": 21,
      "test_mse": 2.6892607386888923,
      "test_mse_normalized": 0.7488980293273926,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 5,
      "model": "aggregate_moe",
      "best_epoch": 86,
      "test_mse": 0.5284301428376488,
      "test_mse_normalized": 0.1471557915210724,
      "gate_accuracy": 0.972900390625,
      "gate_entropy": 0.06390704214572906,
      "largest_expert_share": 0.5088528394699097,
      "utilization": [
        0.4911471903324127,
        0.5088528394699097
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 5,
      "model": "sample_weight_moe",
      "best_epoch": 55,
      "test_mse": 1.8341483466165258,
      "test_mse_normalized": 0.5107686519622803,
      "gate_accuracy": 0.50048828125,
      "gate_entropy": 0.046004679054021835,
      "largest_expert_share": 0.5461182594299316,
      "utilization": [
        0.45388177037239075,
        0.5461182594299316
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 5,
      "model": "sample_weight_warmstart",
      "best_epoch": 121,
      "test_mse": 0.41330810595178025,
      "test_mse_normalized": 0.11509691923856735,
      "gate_accuracy": 0.983642578125,
      "gate_entropy": 0.03539521247148514,
      "largest_expert_share": 0.5038039684295654,
      "utilization": [
        0.49619606137275696,
        0.5038039684295654
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 5,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2608238576915746,
      "test_mse_normalized": 0.0726335197687149,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 6,
      "model": "linear",
      "best_epoch": 46,
      "test_mse": 2.5732283075858815,
      "test_mse_normalized": 0.727455735206604,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 6,
      "model": "mlp",
      "best_epoch": 74,
      "test_mse": 0.7468038353547082,
      "test_mse_normalized": 0.21112263202667236,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 6,
      "model": "uniform_ensemble",
      "best_epoch": 42,
      "test_mse": 2.5805461228390514,
      "test_mse_normalized": 0.7295244932174683,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 6,
      "model": "aggregate_moe",
      "best_epoch": 94,
      "test_mse": 0.5327148333736202,
      "test_mse_normalized": 0.15059933066368103,
      "gate_accuracy": 0.978515625,
      "gate_entropy": 0.052518170326948166,
      "largest_expert_share": 0.5084831118583679,
      "utilization": [
        0.5084831118583679,
        0.4915168881416321
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 6,
      "model": "sample_weight_moe",
      "best_epoch": 128,
      "test_mse": 0.3901111540674943,
      "test_mse_normalized": 0.11028504371643066,
      "gate_accuracy": 0.98828125,
      "gate_entropy": 0.01705126278102398,
      "largest_expert_share": 0.5098537802696228,
      "utilization": [
        0.5098537802696228,
        0.4901461899280548
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 6,
      "model": "sample_weight_warmstart",
      "best_epoch": 106,
      "test_mse": 0.49292305331759256,
      "test_mse_normalized": 0.13935013115406036,
      "gate_accuracy": 0.979248046875,
      "gate_entropy": 0.04336098954081535,
      "largest_expert_share": 0.5099048614501953,
      "utilization": [
        0.5099048614501953,
        0.4900951385498047
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 6,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2725913327383656,
      "test_mse_normalized": 0.07706200331449509,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 7,
      "model": "linear",
      "best_epoch": 71,
      "test_mse": 2.880773812525198,
      "test_mse_normalized": 0.8548643589019775,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 7,
      "model": "mlp",
      "best_epoch": 46,
      "test_mse": 0.8292228209665531,
      "test_mse_normalized": 0.24607035517692566,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 7,
      "model": "uniform_ensemble",
      "best_epoch": 70,
      "test_mse": 2.8859093843951675,
      "test_mse_normalized": 0.8563883304595947,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 7,
      "model": "aggregate_moe",
      "best_epoch": 149,
      "test_mse": 0.4116399976837459,
      "test_mse_normalized": 0.12215341627597809,
      "gate_accuracy": 0.98095703125,
      "gate_entropy": 0.024955224245786667,
      "largest_expert_share": 0.5117238759994507,
      "utilization": [
        0.4882761240005493,
        0.5117238759994507
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 7,
      "model": "sample_weight_moe",
      "best_epoch": 599,
      "test_mse": 0.3325699317883302,
      "test_mse_normalized": 0.09868951886892319,
      "gate_accuracy": 0.989990234375,
      "gate_entropy": 0.01390939299017191,
      "largest_expert_share": 0.5093435645103455,
      "utilization": [
        0.49065643548965454,
        0.5093435645103455
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 7,
      "model": "sample_weight_warmstart",
      "best_epoch": 599,
      "test_mse": 0.3791697781124933,
      "test_mse_normalized": 0.11251793801784515,
      "gate_accuracy": 0.98828125,
      "gate_entropy": 0.01271553710103035,
      "largest_expert_share": 0.5085070133209229,
      "utilization": [
        0.49149301648139954,
        0.5085070133209229
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 7,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2637699551212649,
      "test_mse_normalized": 0.07827325165271759,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 8,
      "model": "linear",
      "best_epoch": 32,
      "test_mse": 2.7304150088434245,
      "test_mse_normalized": 0.8209166526794434,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 8,
      "model": "mlp",
      "best_epoch": 59,
      "test_mse": 0.7431284449477483,
      "test_mse_normalized": 0.22342629730701447,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 8,
      "model": "uniform_ensemble",
      "best_epoch": 97,
      "test_mse": 2.7821784602743613,
      "test_mse_normalized": 0.836479663848877,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 8,
      "model": "aggregate_moe",
      "best_epoch": 91,
      "test_mse": 1.4538371300451975,
      "test_mse_normalized": 0.4371053874492645,
      "gate_accuracy": 0.51220703125,
      "gate_entropy": 0.3740537762641907,
      "largest_expert_share": 0.515407919883728,
      "utilization": [
        0.48459213972091675,
        0.515407919883728
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 8,
      "model": "sample_weight_moe",
      "best_epoch": 46,
      "test_mse": 1.7902873720694028,
      "test_mse_normalized": 0.5382612943649292,
      "gate_accuracy": 0.51953125,
      "gate_entropy": 0.06271661072969437,
      "largest_expert_share": 0.5238140821456909,
      "utilization": [
        0.5238140821456909,
        0.4761858880519867
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 8,
      "model": "sample_weight_warmstart",
      "best_epoch": 100,
      "test_mse": 1.6777274789009884,
      "test_mse_normalized": 0.5044194459915161,
      "gate_accuracy": 0.512939453125,
      "gate_entropy": 0.357629656791687,
      "largest_expert_share": 0.5013622045516968,
      "utilization": [
        0.5013622045516968,
        0.49863776564598083
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 8,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.253225239625484,
      "test_mse_normalized": 0.07613378018140793,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 9,
      "model": "linear",
      "best_epoch": 62,
      "test_mse": 2.6569555279412165,
      "test_mse_normalized": 0.7589794993400574,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 9,
      "model": "mlp",
      "best_epoch": 92,
      "test_mse": 0.7702232567724847,
      "test_mse_normalized": 0.2200201153755188,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 9,
      "model": "uniform_ensemble",
      "best_epoch": 53,
      "test_mse": 2.6639255285965984,
      "test_mse_normalized": 0.7609705328941345,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 9,
      "model": "aggregate_moe",
      "best_epoch": 64,
      "test_mse": 0.5529584626634996,
      "test_mse_normalized": 0.1579567790031433,
      "gate_accuracy": 0.970947265625,
      "gate_entropy": 0.09867739677429199,
      "largest_expert_share": 0.5082385540008545,
      "utilization": [
        0.4917614758014679,
        0.5082385540008545
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 9,
      "model": "sample_weight_moe",
      "best_epoch": 107,
      "test_mse": 0.3437383083511783,
      "test_mse_normalized": 0.09819145500659943,
      "gate_accuracy": 0.988525390625,
      "gate_entropy": 0.02569868415594101,
      "largest_expert_share": 0.5064408779144287,
      "utilization": [
        0.4935591220855713,
        0.5064408779144287
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 9,
      "model": "sample_weight_warmstart",
      "best_epoch": 111,
      "test_mse": 0.4054467784841817,
      "test_mse_normalized": 0.11581894755363464,
      "gate_accuracy": 0.982666015625,
      "gate_entropy": 0.04535454511642456,
      "largest_expert_share": 0.5021197199821472,
      "utilization": [
        0.5021197199821472,
        0.4978802800178528
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 9,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2619050150469696,
      "test_mse_normalized": 0.07481515407562256,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 10,
      "model": "linear",
      "best_epoch": 41,
      "test_mse": 2.7080204204644636,
      "test_mse_normalized": 0.884597897529602,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 10,
      "model": "mlp",
      "best_epoch": 76,
      "test_mse": 0.8403156645514478,
      "test_mse_normalized": 0.27449625730514526,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 10,
      "model": "uniform_ensemble",
      "best_epoch": 45,
      "test_mse": 2.7086484744879376,
      "test_mse_normalized": 0.884803056716919,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 10,
      "model": "aggregate_moe",
      "best_epoch": 88,
      "test_mse": 1.6854869786144193,
      "test_mse_normalized": 0.5505786538124084,
      "gate_accuracy": 0.51806640625,
      "gate_entropy": 0.3526279926300049,
      "largest_expert_share": 0.5574295520782471,
      "utilization": [
        0.5574295520782471,
        0.44257044792175293
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 10,
      "model": "sample_weight_moe",
      "best_epoch": 46,
      "test_mse": 1.830890243501024,
      "test_mse_normalized": 0.5980758666992188,
      "gate_accuracy": 0.52685546875,
      "gate_entropy": 0.0550498366355896,
      "largest_expert_share": 0.5088452100753784,
      "utilization": [
        0.4911547899246216,
        0.5088452100753784
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 10,
      "model": "sample_weight_warmstart",
      "best_epoch": 121,
      "test_mse": 1.885106888473368,
      "test_mse_normalized": 0.6157861948013306,
      "gate_accuracy": 0.5595703125,
      "gate_entropy": 0.059202469885349274,
      "largest_expert_share": 0.5085684657096863,
      "utilization": [
        0.5085684657096863,
        0.4914315342903137
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 10,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.25833760106336595,
      "test_mse_normalized": 0.08438817411661148,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 11,
      "model": "linear",
      "best_epoch": 68,
      "test_mse": 2.663419662970744,
      "test_mse_normalized": 0.7498977184295654,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 11,
      "model": "mlp",
      "best_epoch": 63,
      "test_mse": 0.757439580623903,
      "test_mse_normalized": 0.2132605016231537,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 11,
      "model": "uniform_ensemble",
      "best_epoch": 21,
      "test_mse": 2.6886185525932196,
      "test_mse_normalized": 0.7569925785064697,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 11,
      "model": "aggregate_moe",
      "best_epoch": 93,
      "test_mse": 1.6030443669531123,
      "test_mse_normalized": 0.4513443112373352,
      "gate_accuracy": 0.547119140625,
      "gate_entropy": 0.3438452482223511,
      "largest_expert_share": 0.5743182897567749,
      "utilization": [
        0.5743182897567749,
        0.4256817102432251
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 11,
      "model": "sample_weight_moe",
      "best_epoch": 26,
      "test_mse": 1.7775986283878138,
      "test_mse_normalized": 0.5004908442497253,
      "gate_accuracy": 0.54736328125,
      "gate_entropy": 0.09033910930156708,
      "largest_expert_share": 0.5197957754135132,
      "utilization": [
        0.5197957754135132,
        0.4802042245864868
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 11,
      "model": "sample_weight_warmstart",
      "best_epoch": 100,
      "test_mse": 1.8803240299266322,
      "test_mse_normalized": 0.529413640499115,
      "gate_accuracy": 0.580322265625,
      "gate_entropy": 0.3182218074798584,
      "largest_expert_share": 0.5724090337753296,
      "utilization": [
        0.5724090337753296,
        0.4275910258293152
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 11,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.25794087746778066,
      "test_mse_normalized": 0.07262440770864487,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 12,
      "model": "linear",
      "best_epoch": 77,
      "test_mse": 2.749693351179195,
      "test_mse_normalized": 0.7694194912910461,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 12,
      "model": "mlp",
      "best_epoch": 70,
      "test_mse": 0.6769630802825286,
      "test_mse_normalized": 0.18942788243293762,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 12,
      "model": "uniform_ensemble",
      "best_epoch": 30,
      "test_mse": 2.743831086887724,
      "test_mse_normalized": 0.7677791118621826,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 12,
      "model": "aggregate_moe",
      "best_epoch": 71,
      "test_mse": 0.45835420502315555,
      "test_mse_normalized": 0.12825672328472137,
      "gate_accuracy": 0.9814453125,
      "gate_entropy": 0.042229488492012024,
      "largest_expert_share": 0.502122163772583,
      "utilization": [
        0.502122163772583,
        0.4978778660297394
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 12,
      "model": "sample_weight_moe",
      "best_epoch": 140,
      "test_mse": 0.3791028183317991,
      "test_mse_normalized": 0.10608059167861938,
      "gate_accuracy": 0.986328125,
      "gate_entropy": 0.0139324851334095,
      "largest_expert_share": 0.5003893375396729,
      "utilization": [
        0.4996106028556824,
        0.5003893375396729
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 12,
      "model": "sample_weight_warmstart",
      "best_epoch": 599,
      "test_mse": 0.504945903629357,
      "test_mse_normalized": 0.14129401743412018,
      "gate_accuracy": 0.970703125,
      "gate_entropy": 0.020453935489058495,
      "largest_expert_share": 0.5158008933067322,
      "utilization": [
        0.5158008933067322,
        0.4841991066932678
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 12,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2534724773553453,
      "test_mse_normalized": 0.07092669606208801,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 13,
      "model": "linear",
      "best_epoch": 50,
      "test_mse": 2.6287906886094254,
      "test_mse_normalized": 0.7132083177566528,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 13,
      "model": "mlp",
      "best_epoch": 111,
      "test_mse": 0.7106630849597708,
      "test_mse_normalized": 0.19280759990215302,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 13,
      "model": "uniform_ensemble",
      "best_epoch": 65,
      "test_mse": 2.6233334705974447,
      "test_mse_normalized": 0.7117277383804321,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 13,
      "model": "aggregate_moe",
      "best_epoch": 107,
      "test_mse": 0.5467267399859752,
      "test_mse_normalized": 0.14833058416843414,
      "gate_accuracy": 0.971923828125,
      "gate_entropy": 0.03869147598743439,
      "largest_expert_share": 0.5070367455482483,
      "utilization": [
        0.5070367455482483,
        0.4929632544517517
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 13,
      "model": "sample_weight_moe",
      "best_epoch": 141,
      "test_mse": 0.49911827981907614,
      "test_mse_normalized": 0.13541409373283386,
      "gate_accuracy": 0.97802734375,
      "gate_entropy": 0.017131056636571884,
      "largest_expert_share": 0.5056635737419128,
      "utilization": [
        0.5056635737419128,
        0.49433642625808716
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 13,
      "model": "sample_weight_warmstart",
      "best_epoch": 170,
      "test_mse": 0.5372121440935691,
      "test_mse_normalized": 0.14574921131134033,
      "gate_accuracy": 0.974853515625,
      "gate_entropy": 0.01909870095551014,
      "largest_expert_share": 0.5049862861633301,
      "utilization": [
        0.5049862861633301,
        0.4950137436389923
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 13,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.24995272308459698,
      "test_mse_normalized": 0.06781382858753204,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 14,
      "model": "linear",
      "best_epoch": 43,
      "test_mse": 2.6262028184699284,
      "test_mse_normalized": 0.8455849289894104,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 14,
      "model": "mlp",
      "best_epoch": 68,
      "test_mse": 0.9388349160043228,
      "test_mse_normalized": 0.3022861182689667,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 14,
      "model": "uniform_ensemble",
      "best_epoch": 48,
      "test_mse": 2.626696901237591,
      "test_mse_normalized": 0.8457440137863159,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 14,
      "model": "aggregate_moe",
      "best_epoch": 90,
      "test_mse": 0.3947103015969216,
      "test_mse_normalized": 0.12708884477615356,
      "gate_accuracy": 0.978759765625,
      "gate_entropy": 0.0521075502038002,
      "largest_expert_share": 0.5123590230941772,
      "utilization": [
        0.5123590230941772,
        0.48764097690582275
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 14,
      "model": "sample_weight_moe",
      "best_epoch": 277,
      "test_mse": 0.3988739069056089,
      "test_mse_normalized": 0.12842944264411926,
      "gate_accuracy": 0.986083984375,
      "gate_entropy": 0.018527096137404442,
      "largest_expert_share": 0.500607967376709,
      "utilization": [
        0.500607967376709,
        0.49939200282096863
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 14,
      "model": "sample_weight_warmstart",
      "best_epoch": 101,
      "test_mse": 0.41356347445134645,
      "test_mse_normalized": 0.13315919041633606,
      "gate_accuracy": 0.977294921875,
      "gate_entropy": 0.0494232177734375,
      "largest_expert_share": 0.5090207457542419,
      "utilization": [
        0.5090207457542419,
        0.49097925424575806
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 14,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2566069252520152,
      "test_mse_normalized": 0.08262231200933456,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 15,
      "model": "linear",
      "best_epoch": 62,
      "test_mse": 2.8020591865435476,
      "test_mse_normalized": 0.818220317363739,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 15,
      "model": "mlp",
      "best_epoch": 58,
      "test_mse": 0.8225410879589308,
      "test_mse_normalized": 0.24018758535385132,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 15,
      "model": "uniform_ensemble",
      "best_epoch": 20,
      "test_mse": 2.8051250801646384,
      "test_mse_normalized": 0.8191155791282654,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 15,
      "model": "aggregate_moe",
      "best_epoch": 71,
      "test_mse": 1.833662394036542,
      "test_mse_normalized": 0.5354418754577637,
      "gate_accuracy": 0.550537109375,
      "gate_entropy": 0.3172553777694702,
      "largest_expert_share": 0.5358485579490662,
      "utilization": [
        0.5358485579490662,
        0.4641515016555786
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 15,
      "model": "sample_weight_moe",
      "best_epoch": 38,
      "test_mse": 1.9235543133586224,
      "test_mse_normalized": 0.5616909265518188,
      "gate_accuracy": 0.540283203125,
      "gate_entropy": 0.0665663629770279,
      "largest_expert_share": 0.5606364011764526,
      "utilization": [
        0.5606364011764526,
        0.439363569021225
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 15,
      "model": "sample_weight_warmstart",
      "best_epoch": 154,
      "test_mse": 1.9330214336080356,
      "test_mse_normalized": 0.5644553899765015,
      "gate_accuracy": 0.576171875,
      "gate_entropy": 0.0507880300283432,
      "largest_expert_share": 0.5011688470840454,
      "utilization": [
        0.5011688470840454,
        0.49883121252059937
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 15,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2571630030798847,
      "test_mse_normalized": 0.0750933438539505,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 16,
      "model": "linear",
      "best_epoch": 58,
      "test_mse": 2.7621060290708837,
      "test_mse_normalized": 0.8316883444786072,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 16,
      "model": "mlp",
      "best_epoch": 89,
      "test_mse": 0.8194216972743841,
      "test_mse_normalized": 0.24673327803611755,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 16,
      "model": "uniform_ensemble",
      "best_epoch": 64,
      "test_mse": 2.7646869268946084,
      "test_mse_normalized": 0.8324654698371887,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 16,
      "model": "aggregate_moe",
      "best_epoch": 96,
      "test_mse": 0.37334979304023447,
      "test_mse_normalized": 0.11241808533668518,
      "gate_accuracy": 0.989990234375,
      "gate_entropy": 0.041472770273685455,
      "largest_expert_share": 0.5073099732398987,
      "utilization": [
        0.5073099732398987,
        0.4926900565624237
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 16,
      "model": "sample_weight_moe",
      "best_epoch": 244,
      "test_mse": 0.36233307291194355,
      "test_mse_normalized": 0.10910087823867798,
      "gate_accuracy": 0.989501953125,
      "gate_entropy": 0.013230562210083008,
      "largest_expert_share": 0.5044481754302979,
      "utilization": [
        0.5044481754302979,
        0.49555182456970215
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 16,
      "model": "sample_weight_warmstart",
      "best_epoch": 237,
      "test_mse": 0.34464537129419726,
      "test_mse_normalized": 0.10377499461174011,
      "gate_accuracy": 0.991943359375,
      "gate_entropy": 0.014791266992688179,
      "largest_expert_share": 0.5035885572433472,
      "utilization": [
        0.5035885572433472,
        0.49641141295433044
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 16,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2517873132247051,
      "test_mse_normalized": 0.07581482082605362,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 17,
      "model": "linear",
      "best_epoch": 109,
      "test_mse": 2.6797011033688105,
      "test_mse_normalized": 0.7598170042037964,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 17,
      "model": "mlp",
      "best_epoch": 63,
      "test_mse": 0.8226108512857331,
      "test_mse_normalized": 0.2332475483417511,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 17,
      "model": "uniform_ensemble",
      "best_epoch": 28,
      "test_mse": 2.669952102648289,
      "test_mse_normalized": 0.7570527195930481,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 17,
      "model": "aggregate_moe",
      "best_epoch": 89,
      "test_mse": 1.6178155266666572,
      "test_mse_normalized": 0.4587242007255554,
      "gate_accuracy": 0.5556640625,
      "gate_entropy": 0.3026999235153198,
      "largest_expert_share": 0.5605466365814209,
      "utilization": [
        0.4394533932209015,
        0.5605466365814209
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 17,
      "model": "sample_weight_moe",
      "best_epoch": 46,
      "test_mse": 1.7741539555832566,
      "test_mse_normalized": 0.5030532479286194,
      "gate_accuracy": 0.51220703125,
      "gate_entropy": 0.06372132897377014,
      "largest_expert_share": 0.528558611869812,
      "utilization": [
        0.528558611869812,
        0.4714413285255432
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 17,
      "model": "sample_weight_warmstart",
      "best_epoch": 139,
      "test_mse": 1.836926614426204,
      "test_mse_normalized": 0.5208521485328674,
      "gate_accuracy": 0.66552734375,
      "gate_entropy": 0.05739796161651611,
      "largest_expert_share": 0.5264997482299805,
      "utilization": [
        0.4735002815723419,
        0.5264997482299805
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 17,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2679988252925363,
      "test_mse_normalized": 0.07598984241485596,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 18,
      "model": "linear",
      "best_epoch": 66,
      "test_mse": 2.59628487346879,
      "test_mse_normalized": 0.8733946084976196,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 18,
      "model": "mlp",
      "best_epoch": 94,
      "test_mse": 0.7225353024243942,
      "test_mse_normalized": 0.2430620938539505,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 18,
      "model": "uniform_ensemble",
      "best_epoch": 71,
      "test_mse": 2.593105502524185,
      "test_mse_normalized": 0.87232506275177,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 18,
      "model": "aggregate_moe",
      "best_epoch": 218,
      "test_mse": 0.39457231045492236,
      "test_mse_normalized": 0.13273479044437408,
      "gate_accuracy": 0.984619140625,
      "gate_entropy": 0.049668245017528534,
      "largest_expert_share": 0.5018083453178406,
      "utilization": [
        0.4981917142868042,
        0.5018083453178406
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 18,
      "model": "sample_weight_moe",
      "best_epoch": 69,
      "test_mse": 0.5070275413648975,
      "test_mse_normalized": 0.1705649197101593,
      "gate_accuracy": 0.973876953125,
      "gate_entropy": 0.03187664598226547,
      "largest_expert_share": 0.5155689716339111,
      "utilization": [
        0.5155689716339111,
        0.4844309985637665
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 18,
      "model": "sample_weight_warmstart",
      "best_epoch": 133,
      "test_mse": 0.5134946306310935,
      "test_mse_normalized": 0.17274045944213867,
      "gate_accuracy": 0.977294921875,
      "gate_entropy": 0.03342878818511963,
      "largest_expert_share": 0.5101516246795654,
      "utilization": [
        0.5101516246795654,
        0.48984837532043457
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 18,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.25657777337547655,
      "test_mse_normalized": 0.08631319552659988,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 19,
      "model": "linear",
      "best_epoch": 45,
      "test_mse": 2.6923888887491083,
      "test_mse_normalized": 0.8607102036476135,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 19,
      "model": "mlp",
      "best_epoch": 48,
      "test_mse": 0.8076126651609046,
      "test_mse_normalized": 0.25817981362342834,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 19,
      "model": "uniform_ensemble",
      "best_epoch": 42,
      "test_mse": 2.697540298652479,
      "test_mse_normalized": 0.8623570203781128,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous",
      "seed": 19,
      "model": "aggregate_moe",
      "best_epoch": 108,
      "test_mse": 1.6114344415399016,
      "test_mse_normalized": 0.5151477456092834,
      "gate_accuracy": 0.53759765625,
      "gate_entropy": 0.34168359637260437,
      "largest_expert_share": 0.5579012632369995,
      "utilization": [
        0.4420987069606781,
        0.5579012632369995
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 19,
      "model": "sample_weight_moe",
      "best_epoch": 53,
      "test_mse": 1.7753282937140722,
      "test_mse_normalized": 0.567541778087616,
      "gate_accuracy": 0.5458984375,
      "gate_entropy": 0.050699345767498016,
      "largest_expert_share": 0.5112892389297485,
      "utilization": [
        0.48871082067489624,
        0.5112892389297485
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 19,
      "model": "sample_weight_warmstart",
      "best_epoch": 119,
      "test_mse": 1.7989031396944577,
      "test_mse_normalized": 0.5750782489776611,
      "gate_accuracy": 0.5966796875,
      "gate_entropy": 0.06392669677734375,
      "largest_expert_share": 0.5042808651924133,
      "utilization": [
        0.49571913480758667,
        0.5042808651924133
      ]
    },
    {
      "scenario": "heterogeneous",
      "seed": 19,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2641591035006358,
      "test_mse_normalized": 0.08444710075855255,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 0,
      "model": "linear",
      "best_epoch": 23,
      "test_mse": 3.2965310705549973,
      "test_mse_normalized": 1.3691353797912598,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 0,
      "model": "mlp",
      "best_epoch": 52,
      "test_mse": 1.8058136895095382,
      "test_mse_normalized": 0.7500015497207642,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 0,
      "model": "uniform_ensemble",
      "best_epoch": 81,
      "test_mse": 3.3555019486723627,
      "test_mse_normalized": 1.3936275243759155,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 0,
      "model": "aggregate_moe",
      "best_epoch": 79,
      "test_mse": 2.552180270998704,
      "test_mse_normalized": 1.0599870681762695,
      "gate_accuracy": 0.5478515625,
      "gate_entropy": 0.2450094223022461,
      "largest_expert_share": 0.608180582523346,
      "utilization": [
        0.39181941747665405,
        0.608180582523346
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 0,
      "model": "sample_weight_moe",
      "best_epoch": 62,
      "test_mse": 2.7978169188916304,
      "test_mse_normalized": 1.1620063781738281,
      "gate_accuracy": 0.531005859375,
      "gate_entropy": 0.04719267413020134,
      "largest_expert_share": 0.5863513946533203,
      "utilization": [
        0.4136486053466797,
        0.5863513946533203
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 0,
      "model": "sample_weight_warmstart",
      "best_epoch": 130,
      "test_mse": 2.4395393065532773,
      "test_mse_normalized": 1.0132043361663818,
      "gate_accuracy": 0.529296875,
      "gate_entropy": 0.06874515116214752,
      "largest_expert_share": 0.6862789988517761,
      "utilization": [
        0.31372103095054626,
        0.6862789988517761
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 0,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2716146321289269,
      "test_mse_normalized": 0.11280864477157593,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 1,
      "model": "linear",
      "best_epoch": 54,
      "test_mse": 3.1147734115964782,
      "test_mse_normalized": 0.8674249649047852,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 1,
      "model": "mlp",
      "best_epoch": 29,
      "test_mse": 1.8032416793280797,
      "test_mse_normalized": 0.5021799802780151,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 1,
      "model": "uniform_ensemble",
      "best_epoch": 54,
      "test_mse": 3.120214482183663,
      "test_mse_normalized": 0.8689402341842651,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 1,
      "model": "aggregate_moe",
      "best_epoch": 93,
      "test_mse": 0.9702131731461016,
      "test_mse_normalized": 0.27019208669662476,
      "gate_accuracy": 0.9423828125,
      "gate_entropy": 0.11503582447767258,
      "largest_expert_share": 0.5030280351638794,
      "utilization": [
        0.5030280351638794,
        0.4969719648361206
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 1,
      "model": "sample_weight_moe",
      "best_epoch": 134,
      "test_mse": 1.7875560629279927,
      "test_mse_normalized": 0.4978117346763611,
      "gate_accuracy": 0.84228515625,
      "gate_entropy": 0.030023518949747086,
      "largest_expert_share": 0.579285204410553,
      "utilization": [
        0.420714795589447,
        0.579285204410553
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 1,
      "model": "sample_weight_warmstart",
      "best_epoch": 109,
      "test_mse": 0.7590327713922277,
      "test_mse_normalized": 0.21138101816177368,
      "gate_accuracy": 0.9619140625,
      "gate_entropy": 0.05918310210108757,
      "largest_expert_share": 0.519914984703064,
      "utilization": [
        0.519914984703064,
        0.4800850749015808
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 1,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.3249979959819574,
      "test_mse_normalized": 0.09050782769918442,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 2,
      "model": "linear",
      "best_epoch": 59,
      "test_mse": 2.806349988177237,
      "test_mse_normalized": 1.0480142831802368,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 2,
      "model": "mlp",
      "best_epoch": 36,
      "test_mse": 1.2971132242686318,
      "test_mse_normalized": 0.4843990206718445,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 2,
      "model": "uniform_ensemble",
      "best_epoch": 64,
      "test_mse": 2.8184811560336525,
      "test_mse_normalized": 1.0525445938110352,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 2,
      "model": "aggregate_moe",
      "best_epoch": 24,
      "test_mse": 1.7843836087256393,
      "test_mse_normalized": 0.6663671731948853,
      "gate_accuracy": 0.509033203125,
      "gate_entropy": 0.23042024672031403,
      "largest_expert_share": 0.5210139751434326,
      "utilization": [
        0.4789860248565674,
        0.5210139751434326
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 2,
      "model": "sample_weight_moe",
      "best_epoch": 91,
      "test_mse": 2.2199820110320347,
      "test_mse_normalized": 0.8290387392044067,
      "gate_accuracy": 0.697509765625,
      "gate_entropy": 0.0402216836810112,
      "largest_expert_share": 0.5046820044517517,
      "utilization": [
        0.5046820044517517,
        0.4953179657459259
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 2,
      "model": "sample_weight_warmstart",
      "best_epoch": 139,
      "test_mse": 2.363266926956299,
      "test_mse_normalized": 0.8825476169586182,
      "gate_accuracy": 0.50146484375,
      "gate_entropy": 0.05488498508930206,
      "largest_expert_share": 0.6648955345153809,
      "utilization": [
        0.6648955345153809,
        0.33510446548461914
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 2,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.27797234585340175,
      "test_mse_normalized": 0.10380707681179047,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 3,
      "model": "linear",
      "best_epoch": 37,
      "test_mse": 2.9391040123761467,
      "test_mse_normalized": 0.8630812168121338,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 3,
      "model": "mlp",
      "best_epoch": 57,
      "test_mse": 1.3432361556009962,
      "test_mse_normalized": 0.394447386264801,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 3,
      "model": "uniform_ensemble",
      "best_epoch": 24,
      "test_mse": 2.9725868334810817,
      "test_mse_normalized": 0.8729135990142822,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 3,
      "model": "aggregate_moe",
      "best_epoch": 96,
      "test_mse": 0.6247268036112915,
      "test_mse_normalized": 0.18345385789871216,
      "gate_accuracy": 0.968994140625,
      "gate_entropy": 0.06761060655117035,
      "largest_expert_share": 0.5109689235687256,
      "utilization": [
        0.4890311360359192,
        0.5109689235687256
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 3,
      "model": "sample_weight_moe",
      "best_epoch": 156,
      "test_mse": 0.63755642161381,
      "test_mse_normalized": 0.18722133338451385,
      "gate_accuracy": 0.9677734375,
      "gate_entropy": 0.027191197499632835,
      "largest_expert_share": 0.5106452703475952,
      "utilization": [
        0.4893547594547272,
        0.5106452703475952
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 3,
      "model": "sample_weight_warmstart",
      "best_epoch": 102,
      "test_mse": 0.5800438509722571,
      "test_mse_normalized": 0.17033250629901886,
      "gate_accuracy": 0.97802734375,
      "gate_entropy": 0.06303401291370392,
      "largest_expert_share": 0.5153622627258301,
      "utilization": [
        0.4846377372741699,
        0.5153622627258301
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 3,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.3151801167294201,
      "test_mse_normalized": 0.09255407005548477,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 4,
      "model": "linear",
      "best_epoch": 42,
      "test_mse": 2.8479785179692034,
      "test_mse_normalized": 0.8666418194770813,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 4,
      "model": "mlp",
      "best_epoch": 42,
      "test_mse": 2.1410014502082295,
      "test_mse_normalized": 0.6515082120895386,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 4,
      "model": "uniform_ensemble",
      "best_epoch": 28,
      "test_mse": 2.821654596388314,
      "test_mse_normalized": 0.858631432056427,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 4,
      "model": "aggregate_moe",
      "best_epoch": 106,
      "test_mse": 1.1686156643210344,
      "test_mse_normalized": 0.35561054944992065,
      "gate_accuracy": 0.930419921875,
      "gate_entropy": 0.1212732195854187,
      "largest_expert_share": 0.5050070881843567,
      "utilization": [
        0.5050070881843567,
        0.4949929118156433
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 4,
      "model": "sample_weight_moe",
      "best_epoch": 38,
      "test_mse": 2.1190882232085535,
      "test_mse_normalized": 0.6448400020599365,
      "gate_accuracy": 0.59326171875,
      "gate_entropy": 0.0897412896156311,
      "largest_expert_share": 0.5226916670799255,
      "utilization": [
        0.5226916670799255,
        0.47730833292007446
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 4,
      "model": "sample_weight_warmstart",
      "best_epoch": 127,
      "test_mse": 0.47025118385626613,
      "test_mse_normalized": 0.14309775829315186,
      "gate_accuracy": 0.974609375,
      "gate_entropy": 0.03241262584924698,
      "largest_expert_share": 0.5101568698883057,
      "utilization": [
        0.5101568698883057,
        0.48984313011169434
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 4,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.28945619009245277,
      "test_mse_normalized": 0.0880817174911499,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 5,
      "model": "linear",
      "best_epoch": 59,
      "test_mse": 2.6883159634319767,
      "test_mse_normalized": 0.8613490462303162,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 5,
      "model": "mlp",
      "best_epoch": 38,
      "test_mse": 1.7519434947022203,
      "test_mse_normalized": 0.5613309144973755,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 5,
      "model": "uniform_ensemble",
      "best_epoch": 59,
      "test_mse": 2.6923181967411023,
      "test_mse_normalized": 0.8626313805580139,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 5,
      "model": "aggregate_moe",
      "best_epoch": 106,
      "test_mse": 0.7300818882040113,
      "test_mse_normalized": 0.23392166197299957,
      "gate_accuracy": 0.957275390625,
      "gate_entropy": 0.06339498609304428,
      "largest_expert_share": 0.500528872013092,
      "utilization": [
        0.500528872013092,
        0.49947109818458557
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 5,
      "model": "sample_weight_moe",
      "best_epoch": 193,
      "test_mse": 0.9002179628136723,
      "test_mse_normalized": 0.28843405842781067,
      "gate_accuracy": 0.947998046875,
      "gate_entropy": 0.027433941140770912,
      "largest_expert_share": 0.5016282200813293,
      "utilization": [
        0.49837175011634827,
        0.5016282200813293
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 5,
      "model": "sample_weight_warmstart",
      "best_epoch": 100,
      "test_mse": 0.7519093978505037,
      "test_mse_normalized": 0.24091529846191406,
      "gate_accuracy": 0.955078125,
      "gate_entropy": 0.06479942798614502,
      "largest_expert_share": 0.5044921040534973,
      "utilization": [
        0.4955078959465027,
        0.5044921040534973
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 5,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2734546278663182,
      "test_mse_normalized": 0.08761614561080933,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 6,
      "model": "linear",
      "best_epoch": 44,
      "test_mse": 3.0509159586636345,
      "test_mse_normalized": 0.9607833027839661,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 6,
      "model": "mlp",
      "best_epoch": 40,
      "test_mse": 1.920174977108382,
      "test_mse_normalized": 0.6046944856643677,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 6,
      "model": "uniform_ensemble",
      "best_epoch": 38,
      "test_mse": 3.0398854131001256,
      "test_mse_normalized": 0.9573096036911011,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 6,
      "model": "aggregate_moe",
      "best_epoch": 59,
      "test_mse": 0.8614314958283551,
      "test_mse_normalized": 0.27127885818481445,
      "gate_accuracy": 0.960693359375,
      "gate_entropy": 0.07648718357086182,
      "largest_expert_share": 0.5040581226348877,
      "utilization": [
        0.4959418773651123,
        0.5040581226348877
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 6,
      "model": "sample_weight_moe",
      "best_epoch": 115,
      "test_mse": 0.7638291767611622,
      "test_mse_normalized": 0.24054229259490967,
      "gate_accuracy": 0.963134765625,
      "gate_entropy": 0.02856520190834999,
      "largest_expert_share": 0.5111198425292969,
      "utilization": [
        0.4888801574707031,
        0.5111198425292969
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 6,
      "model": "sample_weight_warmstart",
      "best_epoch": 104,
      "test_mse": 0.8533641826700401,
      "test_mse_normalized": 0.268738329410553,
      "gate_accuracy": 0.95556640625,
      "gate_entropy": 0.05144825577735901,
      "largest_expert_share": 0.5096657276153564,
      "utilization": [
        0.49033427238464355,
        0.5096657276153564
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 6,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.3181687042500373,
      "test_mse_normalized": 0.10019652545452118,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 7,
      "model": "linear",
      "best_epoch": 72,
      "test_mse": 3.3318148740219553,
      "test_mse_normalized": 0.9302859306335449,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 7,
      "model": "mlp",
      "best_epoch": 50,
      "test_mse": 1.1709309367786724,
      "test_mse_normalized": 0.32693910598754883,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 7,
      "model": "uniform_ensemble",
      "best_epoch": 71,
      "test_mse": 3.3387250198605805,
      "test_mse_normalized": 0.9322153329849243,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 7,
      "model": "aggregate_moe",
      "best_epoch": 47,
      "test_mse": 1.1497368342408172,
      "test_mse_normalized": 0.3210214376449585,
      "gate_accuracy": 0.922119140625,
      "gate_entropy": 0.1368822455406189,
      "largest_expert_share": 0.5136833786964417,
      "utilization": [
        0.48631665110588074,
        0.5136833786964417
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 7,
      "model": "sample_weight_moe",
      "best_epoch": 240,
      "test_mse": 0.5997436079284149,
      "test_mse_normalized": 0.16745619475841522,
      "gate_accuracy": 0.968017578125,
      "gate_entropy": 0.024177473038434982,
      "largest_expert_share": 0.5080214738845825,
      "utilization": [
        0.4919785261154175,
        0.5080214738845825
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 7,
      "model": "sample_weight_warmstart",
      "best_epoch": 124,
      "test_mse": 0.6406554817998164,
      "test_mse_normalized": 0.17887932062149048,
      "gate_accuracy": 0.9638671875,
      "gate_entropy": 0.03936120122671127,
      "largest_expert_share": 0.5161044001579285,
      "utilization": [
        0.48389557003974915,
        0.5161044001579285
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 7,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2894007555862668,
      "test_mse_normalized": 0.08080444484949112,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 8,
      "model": "linear",
      "best_epoch": 28,
      "test_mse": 2.674036976953313,
      "test_mse_normalized": 0.722259521484375,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 8,
      "model": "mlp",
      "best_epoch": 57,
      "test_mse": 1.4238170270355894,
      "test_mse_normalized": 0.38457411527633667,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 8,
      "model": "uniform_ensemble",
      "best_epoch": 27,
      "test_mse": 2.770641013330439,
      "test_mse_normalized": 0.7483523488044739,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 8,
      "model": "aggregate_moe",
      "best_epoch": 30,
      "test_mse": 2.362287069084433,
      "test_mse_normalized": 0.6380556225776672,
      "gate_accuracy": 0.5205078125,
      "gate_entropy": 0.2070201337337494,
      "largest_expert_share": 0.5037603378295898,
      "utilization": [
        0.5037603378295898,
        0.49623972177505493
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 8,
      "model": "sample_weight_moe",
      "best_epoch": 34,
      "test_mse": 2.1440078253209833,
      "test_mse_normalized": 0.5790982246398926,
      "gate_accuracy": 0.54248046875,
      "gate_entropy": 0.08336950838565826,
      "largest_expert_share": 0.5429280400276184,
      "utilization": [
        0.5429280400276184,
        0.4570719301700592
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 8,
      "model": "sample_weight_warmstart",
      "best_epoch": 100,
      "test_mse": 3.2969663463693366,
      "test_mse_normalized": 0.8905132412910461,
      "gate_accuracy": 0.5205078125,
      "gate_entropy": 0.2939206659793854,
      "largest_expert_share": 0.541403591632843,
      "utilization": [
        0.541403591632843,
        0.458596408367157
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 8,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.32469567407000793,
      "test_mse_normalized": 0.08770056068897247,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 9,
      "model": "linear",
      "best_epoch": 41,
      "test_mse": 3.3578063202872697,
      "test_mse_normalized": 1.0478993654251099,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 9,
      "model": "mlp",
      "best_epoch": 27,
      "test_mse": 1.8624463185840765,
      "test_mse_normalized": 0.581229567527771,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 9,
      "model": "uniform_ensemble",
      "best_epoch": 25,
      "test_mse": 3.305356359377301,
      "test_mse_normalized": 1.0315308570861816,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 9,
      "model": "aggregate_moe",
      "best_epoch": 85,
      "test_mse": 2.3498876652415306,
      "test_mse_normalized": 0.7333495616912842,
      "gate_accuracy": 0.82568359375,
      "gate_entropy": 0.3402236998081207,
      "largest_expert_share": 0.5272019505500793,
      "utilization": [
        0.47279810905456543,
        0.5272019505500793
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 9,
      "model": "sample_weight_moe",
      "best_epoch": 152,
      "test_mse": 0.6680931183507569,
      "test_mse_normalized": 0.2084975391626358,
      "gate_accuracy": 0.95751953125,
      "gate_entropy": 0.025151051580905914,
      "largest_expert_share": 0.5022428035736084,
      "utilization": [
        0.5022428035736084,
        0.4977571964263916
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 9,
      "model": "sample_weight_warmstart",
      "best_epoch": 599,
      "test_mse": 0.8907079337502615,
      "test_mse_normalized": 0.2779708504676819,
      "gate_accuracy": 0.93505859375,
      "gate_entropy": 0.03324618935585022,
      "largest_expert_share": 0.5194318294525146,
      "utilization": [
        0.5194318294525146,
        0.4805682301521301
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 9,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.31113248035417096,
      "test_mse_normalized": 0.09709777683019638,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 10,
      "model": "linear",
      "best_epoch": 39,
      "test_mse": 3.0920636874945098,
      "test_mse_normalized": 1.1027803421020508,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 10,
      "model": "mlp",
      "best_epoch": 40,
      "test_mse": 1.866734289811486,
      "test_mse_normalized": 0.6657682657241821,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 10,
      "model": "uniform_ensemble",
      "best_epoch": 49,
      "test_mse": 3.0955091214087838,
      "test_mse_normalized": 1.1040091514587402,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 10,
      "model": "aggregate_moe",
      "best_epoch": 60,
      "test_mse": 2.0847094591639297,
      "test_mse_normalized": 0.7435088157653809,
      "gate_accuracy": 0.53564453125,
      "gate_entropy": 0.25777146220207214,
      "largest_expert_share": 0.5042725801467896,
      "utilization": [
        0.49572744965553284,
        0.5042725801467896
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 10,
      "model": "sample_weight_moe",
      "best_epoch": 139,
      "test_mse": 2.4035267434822836,
      "test_mse_normalized": 0.8572145700454712,
      "gate_accuracy": 0.60791015625,
      "gate_entropy": 0.02806602418422699,
      "largest_expert_share": 0.5044293403625488,
      "utilization": [
        0.5044293403625488,
        0.49557068943977356
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 10,
      "model": "sample_weight_warmstart",
      "best_epoch": 161,
      "test_mse": 2.262651681299398,
      "test_mse_normalized": 0.8069716691970825,
      "gate_accuracy": 0.6162109375,
      "gate_entropy": 0.05409654229879379,
      "largest_expert_share": 0.512130856513977,
      "utilization": [
        0.48786914348602295,
        0.512130856513977
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 10,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.3011820181744475,
      "test_mse_normalized": 0.10741616040468216,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 11,
      "model": "linear",
      "best_epoch": 75,
      "test_mse": 2.6710073740280755,
      "test_mse_normalized": 0.7842597961425781,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 11,
      "model": "mlp",
      "best_epoch": 91,
      "test_mse": 2.0081876094472904,
      "test_mse_normalized": 0.5896430015563965,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 11,
      "model": "uniform_ensemble",
      "best_epoch": 42,
      "test_mse": 2.639746243540578,
      "test_mse_normalized": 0.7750809192657471,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 11,
      "model": "aggregate_moe",
      "best_epoch": 23,
      "test_mse": 1.9291119469263291,
      "test_mse_normalized": 0.5664248466491699,
      "gate_accuracy": 0.534912109375,
      "gate_entropy": 0.1867346465587616,
      "largest_expert_share": 0.5158056616783142,
      "utilization": [
        0.4841943383216858,
        0.5158056616783142
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 11,
      "model": "sample_weight_moe",
      "best_epoch": 35,
      "test_mse": 1.8211031536349447,
      "test_mse_normalized": 0.5347113609313965,
      "gate_accuracy": 0.5146484375,
      "gate_entropy": 0.08044688403606415,
      "largest_expert_share": 0.50537109375,
      "utilization": [
        0.4946288764476776,
        0.50537109375
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 11,
      "model": "sample_weight_warmstart",
      "best_epoch": 136,
      "test_mse": 1.9896697805153696,
      "test_mse_normalized": 0.5842058062553406,
      "gate_accuracy": 0.56201171875,
      "gate_entropy": 0.07704272866249084,
      "largest_expert_share": 0.5354406833648682,
      "utilization": [
        0.46455928683280945,
        0.5354406833648682
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 11,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.3439406939868971,
      "test_mse_normalized": 0.10098768770694733,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 12,
      "model": "linear",
      "best_epoch": 20,
      "test_mse": 3.14925844811637,
      "test_mse_normalized": 0.7509059906005859,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 12,
      "model": "mlp",
      "best_epoch": 35,
      "test_mse": 2.632880646016134,
      "test_mse_normalized": 0.6277813911437988,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 12,
      "model": "uniform_ensemble",
      "best_epoch": 19,
      "test_mse": 3.1595590664640305,
      "test_mse_normalized": 0.7533620595932007,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 12,
      "model": "aggregate_moe",
      "best_epoch": 20,
      "test_mse": 3.0224538016020026,
      "test_mse_normalized": 0.7206708192825317,
      "gate_accuracy": 0.743896484375,
      "gate_entropy": 0.12411151826381683,
      "largest_expert_share": 0.6506685018539429,
      "utilization": [
        0.34933149814605713,
        0.6506685018539429
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 12,
      "model": "sample_weight_moe",
      "best_epoch": 113,
      "test_mse": 0.9454298286036336,
      "test_mse_normalized": 0.22542732954025269,
      "gate_accuracy": 0.9501953125,
      "gate_entropy": 0.020655043423175812,
      "largest_expert_share": 0.5363210439682007,
      "utilization": [
        0.5363210439682007,
        0.4636789560317993
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 12,
      "model": "sample_weight_warmstart",
      "best_epoch": 193,
      "test_mse": 2.6505806310358992,
      "test_mse_normalized": 0.6320017576217651,
      "gate_accuracy": 0.81005859375,
      "gate_entropy": 0.028118515387177467,
      "largest_expert_share": 0.622387707233429,
      "utilization": [
        0.37761229276657104,
        0.622387707233429
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 12,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.32376160104003404,
      "test_mse_normalized": 0.07719738781452179,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 13,
      "model": "linear",
      "best_epoch": 117,
      "test_mse": 2.7685903897479953,
      "test_mse_normalized": 0.6535178422927856,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 13,
      "model": "mlp",
      "best_epoch": 52,
      "test_mse": 1.3776806530821633,
      "test_mse_normalized": 0.32519757747650146,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 13,
      "model": "uniform_ensemble",
      "best_epoch": 101,
      "test_mse": 2.768339645712901,
      "test_mse_normalized": 0.6534586548805237,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 13,
      "model": "aggregate_moe",
      "best_epoch": 67,
      "test_mse": 0.8354936443516905,
      "test_mse_normalized": 0.19721588492393494,
      "gate_accuracy": 0.9365234375,
      "gate_entropy": 0.08423761278390884,
      "largest_expert_share": 0.5304360389709473,
      "utilization": [
        0.46956396102905273,
        0.5304360389709473
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 13,
      "model": "sample_weight_moe",
      "best_epoch": 241,
      "test_mse": 0.657069314964498,
      "test_mse_normalized": 0.15509933233261108,
      "gate_accuracy": 0.9677734375,
      "gate_entropy": 0.023561915382742882,
      "largest_expert_share": 0.5148152112960815,
      "utilization": [
        0.48518478870391846,
        0.5148152112960815
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 13,
      "model": "sample_weight_warmstart",
      "best_epoch": 324,
      "test_mse": 0.6801876372374017,
      "test_mse_normalized": 0.1605563461780548,
      "gate_accuracy": 0.96142578125,
      "gate_entropy": 0.028575923293828964,
      "largest_expert_share": 0.5129532217979431,
      "utilization": [
        0.4870467782020569,
        0.5129532217979431
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 13,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.29644923737928425,
      "test_mse_normalized": 0.06997599452733994,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 14,
      "model": "linear",
      "best_epoch": 25,
      "test_mse": 2.7775989982840317,
      "test_mse_normalized": 0.8414010405540466,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 14,
      "model": "mlp",
      "best_epoch": 33,
      "test_mse": 2.3339361209630574,
      "test_mse_normalized": 0.707004964351654,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 14,
      "model": "uniform_ensemble",
      "best_epoch": 10,
      "test_mse": 2.7892179382597826,
      "test_mse_normalized": 0.8449206948280334,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 14,
      "model": "aggregate_moe",
      "best_epoch": 51,
      "test_mse": 0.7426094468796226,
      "test_mse_normalized": 0.22495412826538086,
      "gate_accuracy": 0.95654296875,
      "gate_entropy": 0.10956896841526031,
      "largest_expert_share": 0.501926064491272,
      "utilization": [
        0.49807390570640564,
        0.501926064491272
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 14,
      "model": "sample_weight_moe",
      "best_epoch": 173,
      "test_mse": 0.5233768237054136,
      "test_mse_normalized": 0.15854333341121674,
      "gate_accuracy": 0.97265625,
      "gate_entropy": 0.022349044680595398,
      "largest_expert_share": 0.5185525417327881,
      "utilization": [
        0.4814474582672119,
        0.5185525417327881
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 14,
      "model": "sample_weight_warmstart",
      "best_epoch": 141,
      "test_mse": 0.7512993514535271,
      "test_mse_normalized": 0.2275865077972412,
      "gate_accuracy": 0.951904296875,
      "gate_entropy": 0.040419138967990875,
      "largest_expert_share": 0.5272880792617798,
      "utilization": [
        0.4727118909358978,
        0.5272880792617798
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 14,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.291916717344446,
      "test_mse_normalized": 0.0884285420179367,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 15,
      "model": "linear",
      "best_epoch": 7,
      "test_mse": 2.797478685387934,
      "test_mse_normalized": 0.7951676249504089,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 15,
      "model": "mlp",
      "best_epoch": 29,
      "test_mse": 1.7413102996481389,
      "test_mse_normalized": 0.49495768547058105,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 15,
      "model": "uniform_ensemble",
      "best_epoch": 10,
      "test_mse": 2.9320259562124624,
      "test_mse_normalized": 0.8334119319915771,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 15,
      "model": "aggregate_moe",
      "best_epoch": 36,
      "test_mse": 2.2227619807018804,
      "test_mse_normalized": 0.6318076252937317,
      "gate_accuracy": 0.50146484375,
      "gate_entropy": 0.22742027044296265,
      "largest_expert_share": 0.593471884727478,
      "utilization": [
        0.593471884727478,
        0.4065280854701996
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 15,
      "model": "sample_weight_moe",
      "best_epoch": 35,
      "test_mse": 2.2245414529862186,
      "test_mse_normalized": 0.6323134303092957,
      "gate_accuracy": 0.5634765625,
      "gate_entropy": 0.08495679497718811,
      "largest_expert_share": 0.6124796867370605,
      "utilization": [
        0.6124796867370605,
        0.38752031326293945
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 15,
      "model": "sample_weight_warmstart",
      "best_epoch": 120,
      "test_mse": 2.3832263028474703,
      "test_mse_normalized": 0.6774187088012695,
      "gate_accuracy": 0.55712890625,
      "gate_entropy": 0.07018784433603287,
      "largest_expert_share": 0.5799677968025208,
      "utilization": [
        0.42003220319747925,
        0.5799677968025208
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 15,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.34953883548754183,
      "test_mse_normalized": 0.09935445338487625,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 16,
      "model": "linear",
      "best_epoch": 58,
      "test_mse": 3.341138094758939,
      "test_mse_normalized": 1.2083333730697632,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 16,
      "model": "mlp",
      "best_epoch": 110,
      "test_mse": 1.872912691230668,
      "test_mse_normalized": 0.6773449182510376,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 16,
      "model": "uniform_ensemble",
      "best_epoch": 98,
      "test_mse": 3.343633012668718,
      "test_mse_normalized": 1.209235668182373,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 16,
      "model": "aggregate_moe",
      "best_epoch": 117,
      "test_mse": 1.050206588394041,
      "test_mse_normalized": 0.3798106014728546,
      "gate_accuracy": 0.950927734375,
      "gate_entropy": 0.05771825835108757,
      "largest_expert_share": 0.5300170183181763,
      "utilization": [
        0.5300170183181763,
        0.46998298168182373
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 16,
      "model": "sample_weight_moe",
      "best_epoch": 55,
      "test_mse": 2.825465361172768,
      "test_mse_normalized": 1.021838665008545,
      "gate_accuracy": 0.687255859375,
      "gate_entropy": 0.04982669651508331,
      "largest_expert_share": 0.5751120448112488,
      "utilization": [
        0.5751120448112488,
        0.4248879849910736
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 16,
      "model": "sample_weight_warmstart",
      "best_epoch": 143,
      "test_mse": 0.9817183888604617,
      "test_mse_normalized": 0.35504162311553955,
      "gate_accuracy": 0.9482421875,
      "gate_entropy": 0.03962329030036926,
      "largest_expert_share": 0.5355895757675171,
      "utilization": [
        0.5355895757675171,
        0.4644105136394501
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 16,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.32409132802523904,
      "test_mse_normalized": 0.11720867455005646,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 17,
      "model": "linear",
      "best_epoch": 31,
      "test_mse": 2.756294604703167,
      "test_mse_normalized": 0.876028299331665,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 17,
      "model": "mlp",
      "best_epoch": 26,
      "test_mse": 1.4683114738267826,
      "test_mse_normalized": 0.4666708707809448,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 17,
      "model": "uniform_ensemble",
      "best_epoch": 25,
      "test_mse": 2.695560282849957,
      "test_mse_normalized": 0.8567252159118652,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 17,
      "model": "aggregate_moe",
      "best_epoch": 76,
      "test_mse": 2.19924738325139,
      "test_mse_normalized": 0.6989829540252686,
      "gate_accuracy": 0.52734375,
      "gate_entropy": 0.22911304235458374,
      "largest_expert_share": 0.6164594888687134,
      "utilization": [
        0.383540540933609,
        0.6164594888687134
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 17,
      "model": "sample_weight_moe",
      "best_epoch": 59,
      "test_mse": 2.1329417040077203,
      "test_mse_normalized": 0.6779091358184814,
      "gate_accuracy": 0.545166015625,
      "gate_entropy": 0.0588255450129509,
      "largest_expert_share": 0.5103546380996704,
      "utilization": [
        0.4896453619003296,
        0.5103546380996704
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 17,
      "model": "sample_weight_warmstart",
      "best_epoch": 136,
      "test_mse": 2.200810693982737,
      "test_mse_normalized": 0.6994798183441162,
      "gate_accuracy": 0.55517578125,
      "gate_entropy": 0.05480840057134628,
      "largest_expert_share": 0.5899151563644409,
      "utilization": [
        0.41008487343788147,
        0.5899151563644409
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 17,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.32687700293448874,
      "test_mse_normalized": 0.10389074683189392,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 18,
      "model": "linear",
      "best_epoch": 40,
      "test_mse": 2.6875034311691284,
      "test_mse_normalized": 0.970909833908081,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 18,
      "model": "mlp",
      "best_epoch": 51,
      "test_mse": 1.4936587484165382,
      "test_mse_normalized": 0.5396115779876709,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 18,
      "model": "uniform_ensemble",
      "best_epoch": 65,
      "test_mse": 2.7357897377401703,
      "test_mse_normalized": 0.9883541464805603,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 18,
      "model": "aggregate_moe",
      "best_epoch": 99,
      "test_mse": 0.8115922092190404,
      "test_mse_normalized": 0.29320254921913147,
      "gate_accuracy": 0.946533203125,
      "gate_entropy": 0.09983815252780914,
      "largest_expert_share": 0.5140467882156372,
      "utilization": [
        0.5140467882156372,
        0.4859532117843628
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 18,
      "model": "sample_weight_moe",
      "best_epoch": 116,
      "test_mse": 0.8540845780690253,
      "test_mse_normalized": 0.30855369567871094,
      "gate_accuracy": 0.947509765625,
      "gate_entropy": 0.03471095860004425,
      "largest_expert_share": 0.53632652759552,
      "utilization": [
        0.53632652759552,
        0.46367347240448
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 18,
      "model": "sample_weight_warmstart",
      "best_epoch": 100,
      "test_mse": 0.8323718511113719,
      "test_mse_normalized": 0.3007095754146576,
      "gate_accuracy": 0.944091796875,
      "gate_entropy": 0.09628874063491821,
      "largest_expert_share": 0.522918701171875,
      "utilization": [
        0.522918701171875,
        0.4770812690258026
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 18,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.29032740523112394,
      "test_mse_normalized": 0.10488609224557877,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 19,
      "model": "linear",
      "best_epoch": 35,
      "test_mse": 2.9774023474466587,
      "test_mse_normalized": 0.9062724113464355,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 19,
      "model": "mlp",
      "best_epoch": 33,
      "test_mse": 1.9558507411382613,
      "test_mse_normalized": 0.5953288674354553,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 19,
      "model": "uniform_ensemble",
      "best_epoch": 31,
      "test_mse": 2.9689139064198202,
      "test_mse_normalized": 0.9036886692047119,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 19,
      "model": "aggregate_moe",
      "best_epoch": 79,
      "test_mse": 1.9168778935216597,
      "test_mse_normalized": 0.5834661722183228,
      "gate_accuracy": 0.588134765625,
      "gate_entropy": 0.24408867955207825,
      "largest_expert_share": 0.5644124746322632,
      "utilization": [
        0.4355875253677368,
        0.5644124746322632
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 19,
      "model": "sample_weight_moe",
      "best_epoch": 25,
      "test_mse": 1.9787721559097504,
      "test_mse_normalized": 0.6023057699203491,
      "gate_accuracy": 0.557373046875,
      "gate_entropy": 0.10046960413455963,
      "largest_expert_share": 0.5668188333511353,
      "utilization": [
        0.5668188333511353,
        0.43318110704421997
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 19,
      "model": "sample_weight_warmstart",
      "best_epoch": 137,
      "test_mse": 2.1084413758155107,
      "test_mse_normalized": 0.6417749524116516,
      "gate_accuracy": 0.668701171875,
      "gate_entropy": 0.05533484369516373,
      "largest_expert_share": 0.5163915157318115,
      "utilization": [
        0.5163915157318115,
        0.4836084842681885
      ]
    },
    {
      "scenario": "heterogeneous_scarce",
      "seed": 19,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 0.2958062250748197,
      "test_mse_normalized": 0.09003856033086777,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 0,
      "model": "linear",
      "best_epoch": 76,
      "test_mse": 4.871204202302025,
      "test_mse_normalized": 0.942967414855957,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 0,
      "model": "mlp",
      "best_epoch": 31,
      "test_mse": 3.6319798969208428,
      "test_mse_normalized": 0.7030784487724304,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 0,
      "model": "uniform_ensemble",
      "best_epoch": 81,
      "test_mse": 4.88250932070962,
      "test_mse_normalized": 0.9451558589935303,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 0,
      "model": "aggregate_moe",
      "best_epoch": 96,
      "test_mse": 3.003551593157001,
      "test_mse_normalized": 0.5814273357391357,
      "gate_accuracy": 0.9423828125,
      "gate_entropy": 0.09419433027505875,
      "largest_expert_share": 0.5305002927780151,
      "utilization": [
        0.46949976682662964,
        0.5305002927780151
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 0,
      "model": "sample_weight_moe",
      "best_epoch": 90,
      "test_mse": 3.137715969552611,
      "test_mse_normalized": 0.6073988676071167,
      "gate_accuracy": 0.920166015625,
      "gate_entropy": 0.034814536571502686,
      "largest_expert_share": 0.5004375576972961,
      "utilization": [
        0.5004375576972961,
        0.49956244230270386
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 0,
      "model": "sample_weight_warmstart",
      "best_epoch": 211,
      "test_mse": 2.898521389975429,
      "test_mse_normalized": 0.5610955953598022,
      "gate_accuracy": 0.943603515625,
      "gate_entropy": 0.037617072463035583,
      "largest_expert_share": 0.5011255741119385,
      "utilization": [
        0.5011255741119385,
        0.49887436628341675
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 0,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.329762564062306,
      "test_mse_normalized": 0.4509952962398529,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 1,
      "model": "linear",
      "best_epoch": 116,
      "test_mse": 4.790172047403494,
      "test_mse_normalized": 0.8466132879257202,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 1,
      "model": "mlp",
      "best_epoch": 33,
      "test_mse": 3.49179482960939,
      "test_mse_normalized": 0.6171385645866394,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 1,
      "model": "uniform_ensemble",
      "best_epoch": 23,
      "test_mse": 4.732034636629188,
      "test_mse_normalized": 0.8363381028175354,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 1,
      "model": "aggregate_moe",
      "best_epoch": 75,
      "test_mse": 2.634484110728653,
      "test_mse_normalized": 0.4656177759170532,
      "gate_accuracy": 0.9658203125,
      "gate_entropy": 0.0662558451294899,
      "largest_expert_share": 0.5009009838104248,
      "utilization": [
        0.5009009838104248,
        0.4990989565849304
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 1,
      "model": "sample_weight_moe",
      "best_epoch": 75,
      "test_mse": 2.6548184954709635,
      "test_mse_normalized": 0.4692116677761078,
      "gate_accuracy": 0.965576171875,
      "gate_entropy": 0.03681004047393799,
      "largest_expert_share": 0.5162283182144165,
      "utilization": [
        0.5162283182144165,
        0.4837716817855835
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 1,
      "model": "sample_weight_warmstart",
      "best_epoch": 130,
      "test_mse": 2.649720861442348,
      "test_mse_normalized": 0.46831071376800537,
      "gate_accuracy": 0.968505859375,
      "gate_entropy": 0.033431027084589005,
      "largest_expert_share": 0.5042285919189453,
      "utilization": [
        0.5042285919189453,
        0.4957713484764099
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 1,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.389043259219941,
      "test_mse_normalized": 0.42223864793777466,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 2,
      "model": "linear",
      "best_epoch": 46,
      "test_mse": 4.80874015994612,
      "test_mse_normalized": 1.0933284759521484,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 2,
      "model": "mlp",
      "best_epoch": 26,
      "test_mse": 3.2098566441745717,
      "test_mse_normalized": 0.7298018932342529,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 2,
      "model": "uniform_ensemble",
      "best_epoch": 47,
      "test_mse": 4.799036172150848,
      "test_mse_normalized": 1.0911221504211426,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 2,
      "model": "aggregate_moe",
      "best_epoch": 31,
      "test_mse": 3.84444543412076,
      "test_mse_normalized": 0.8740837574005127,
      "gate_accuracy": 0.5390625,
      "gate_entropy": 0.3274727761745453,
      "largest_expert_share": 0.5694127082824707,
      "utilization": [
        0.4305872917175293,
        0.5694127082824707
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 2,
      "model": "sample_weight_moe",
      "best_epoch": 42,
      "test_mse": 4.078750757113238,
      "test_mse_normalized": 0.9273560643196106,
      "gate_accuracy": 0.509521484375,
      "gate_entropy": 0.06720440089702606,
      "largest_expert_share": 0.5985376834869385,
      "utilization": [
        0.4014623165130615,
        0.5985376834869385
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 2,
      "model": "sample_weight_warmstart",
      "best_epoch": 140,
      "test_mse": 3.9572571759631194,
      "test_mse_normalized": 0.8997329473495483,
      "gate_accuracy": 0.648681640625,
      "gate_entropy": 0.08992829918861389,
      "largest_expert_share": 0.5057246088981628,
      "utilization": [
        0.4942753314971924,
        0.5057246088981628
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 2,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.329103354234212,
      "test_mse_normalized": 0.5295513868331909,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 3,
      "model": "linear",
      "best_epoch": 38,
      "test_mse": 4.825380132745131,
      "test_mse_normalized": 0.8973936438560486,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 3,
      "model": "mlp",
      "best_epoch": 30,
      "test_mse": 3.3856289625671336,
      "test_mse_normalized": 0.6296378374099731,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 3,
      "model": "uniform_ensemble",
      "best_epoch": 37,
      "test_mse": 4.812863308082247,
      "test_mse_normalized": 0.8950658440589905,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 3,
      "model": "aggregate_moe",
      "best_epoch": 62,
      "test_mse": 3.0144663628079567,
      "test_mse_normalized": 0.560611367225647,
      "gate_accuracy": 0.966796875,
      "gate_entropy": 0.07448025047779083,
      "largest_expert_share": 0.5035718679428101,
      "utilization": [
        0.5035718679428101,
        0.49642816185951233
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 3,
      "model": "sample_weight_moe",
      "best_epoch": 43,
      "test_mse": 4.179082637734481,
      "test_mse_normalized": 0.7771993279457092,
      "gate_accuracy": 0.59326171875,
      "gate_entropy": 0.0548466220498085,
      "largest_expert_share": 0.5300668478012085,
      "utilization": [
        0.4699331820011139,
        0.5300668478012085
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 3,
      "model": "sample_weight_warmstart",
      "best_epoch": 109,
      "test_mse": 2.963750371241032,
      "test_mse_normalized": 0.5511795282363892,
      "gate_accuracy": 0.965087890625,
      "gate_entropy": 0.05219481140375137,
      "largest_expert_share": 0.5007427930831909,
      "utilization": [
        0.4992572069168091,
        0.5007427930831909
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 3,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.486843615012962,
      "test_mse_normalized": 0.46248742938041687,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 4,
      "model": "linear",
      "best_epoch": 34,
      "test_mse": 4.673772718630708,
      "test_mse_normalized": 0.914145827293396,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 4,
      "model": "mlp",
      "best_epoch": 36,
      "test_mse": 3.18972310325193,
      "test_mse_normalized": 0.6238797307014465,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 4,
      "model": "uniform_ensemble",
      "best_epoch": 27,
      "test_mse": 4.7134202530572855,
      "test_mse_normalized": 0.9219005107879639,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 4,
      "model": "aggregate_moe",
      "best_epoch": 42,
      "test_mse": 3.8548062116336688,
      "test_mse_normalized": 0.7539637088775635,
      "gate_accuracy": 0.54150390625,
      "gate_entropy": 0.27834901213645935,
      "largest_expert_share": 0.5211102962493896,
      "utilization": [
        0.47888967394828796,
        0.5211102962493896
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 4,
      "model": "sample_weight_moe",
      "best_epoch": 37,
      "test_mse": 4.092723720837592,
      "test_mse_normalized": 0.8004981279373169,
      "gate_accuracy": 0.625244140625,
      "gate_entropy": 0.08340968191623688,
      "largest_expert_share": 0.5076553225517273,
      "utilization": [
        0.5076553225517273,
        0.4923446774482727
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 4,
      "model": "sample_weight_warmstart",
      "best_epoch": 150,
      "test_mse": 3.960189018047519,
      "test_mse_normalized": 0.7745755910873413,
      "gate_accuracy": 0.561279296875,
      "gate_entropy": 0.06455878913402557,
      "largest_expert_share": 0.5942511558532715,
      "utilization": [
        0.5942511558532715,
        0.4057488441467285
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 4,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.328163121602169,
      "test_mse_normalized": 0.45536673069000244,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 5,
      "model": "linear",
      "best_epoch": 70,
      "test_mse": 4.699435649448043,
      "test_mse_normalized": 0.7967666387557983,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 5,
      "model": "mlp",
      "best_epoch": 30,
      "test_mse": 3.197025039517918,
      "test_mse_normalized": 0.5420401692390442,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 5,
      "model": "uniform_ensemble",
      "best_epoch": 16,
      "test_mse": 4.681787883498933,
      "test_mse_normalized": 0.7937745451927185,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 5,
      "model": "aggregate_moe",
      "best_epoch": 40,
      "test_mse": 3.2300723699929113,
      "test_mse_normalized": 0.5476431846618652,
      "gate_accuracy": 0.932861328125,
      "gate_entropy": 0.1955130249261856,
      "largest_expert_share": 0.5425934791564941,
      "utilization": [
        0.5425934791564941,
        0.45740649104118347
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 5,
      "model": "sample_weight_moe",
      "best_epoch": 50,
      "test_mse": 4.0266240207091,
      "test_mse_normalized": 0.682694673538208,
      "gate_accuracy": 0.518310546875,
      "gate_entropy": 0.052349358797073364,
      "largest_expert_share": 0.5224241018295288,
      "utilization": [
        0.4775758683681488,
        0.5224241018295288
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 5,
      "model": "sample_weight_warmstart",
      "best_epoch": 141,
      "test_mse": 2.7033446159061216,
      "test_mse_normalized": 0.45833903551101685,
      "gate_accuracy": 0.958984375,
      "gate_entropy": 0.03894221782684326,
      "largest_expert_share": 0.5212022066116333,
      "utilization": [
        0.4787977933883667,
        0.5212022066116333
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 5,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.347414428989075,
      "test_mse_normalized": 0.3979927897453308,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 6,
      "model": "linear",
      "best_epoch": 44,
      "test_mse": 4.6599970315767365,
      "test_mse_normalized": 0.811465322971344,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 6,
      "model": "mlp",
      "best_epoch": 23,
      "test_mse": 3.638961787576727,
      "test_mse_normalized": 0.633668065071106,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 6,
      "model": "uniform_ensemble",
      "best_epoch": 44,
      "test_mse": 4.6852947498966335,
      "test_mse_normalized": 0.8158705234527588,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 6,
      "model": "aggregate_moe",
      "best_epoch": 46,
      "test_mse": 3.141889878732216,
      "test_mse_normalized": 0.5471107959747314,
      "gate_accuracy": 0.94580078125,
      "gate_entropy": 0.17552641034126282,
      "largest_expert_share": 0.5004682540893555,
      "utilization": [
        0.5004682540893555,
        0.4995317757129669
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 6,
      "model": "sample_weight_moe",
      "best_epoch": 91,
      "test_mse": 2.6617947594255162,
      "test_mse_normalized": 0.46350976824760437,
      "gate_accuracy": 0.980224609375,
      "gate_entropy": 0.026840265840291977,
      "largest_expert_share": 0.5153958797454834,
      "utilization": [
        0.5153958797454834,
        0.4846040606498718
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 6,
      "model": "sample_weight_warmstart",
      "best_epoch": 153,
      "test_mse": 2.9812758778570356,
      "test_mse_normalized": 0.5191423892974854,
      "gate_accuracy": 0.95556640625,
      "gate_entropy": 0.03752575069665909,
      "largest_expert_share": 0.5029606223106384,
      "utilization": [
        0.5029606223106384,
        0.4970393776893616
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 6,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.453321940335963,
      "test_mse_normalized": 0.42720749974250793,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 7,
      "model": "linear",
      "best_epoch": 34,
      "test_mse": 5.081801364966066,
      "test_mse_normalized": 0.8873674869537354,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 7,
      "model": "mlp",
      "best_epoch": 26,
      "test_mse": 3.3278374059846816,
      "test_mse_normalized": 0.5810960531234741,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 7,
      "model": "uniform_ensemble",
      "best_epoch": 22,
      "test_mse": 5.1316897034611495,
      "test_mse_normalized": 0.8960788249969482,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 7,
      "model": "aggregate_moe",
      "best_epoch": 55,
      "test_mse": 2.5484341793259535,
      "test_mse_normalized": 0.44499921798706055,
      "gate_accuracy": 0.982666015625,
      "gate_entropy": 0.06607143580913544,
      "largest_expert_share": 0.5149497389793396,
      "utilization": [
        0.4850502610206604,
        0.5149497389793396
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 7,
      "model": "sample_weight_moe",
      "best_epoch": 108,
      "test_mse": 2.528462732416774,
      "test_mse_normalized": 0.441511869430542,
      "gate_accuracy": 0.982666015625,
      "gate_entropy": 0.017126033082604408,
      "largest_expert_share": 0.5122019052505493,
      "utilization": [
        0.4877980947494507,
        0.5122019052505493
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 7,
      "model": "sample_weight_warmstart",
      "best_epoch": 199,
      "test_mse": 2.567778745086423,
      "test_mse_normalized": 0.4483771026134491,
      "gate_accuracy": 0.975830078125,
      "gate_entropy": 0.017416175454854965,
      "largest_expert_share": 0.5008973479270935,
      "utilization": [
        0.5008973479270935,
        0.4991026520729065
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 7,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.3739296170119335,
      "test_mse_normalized": 0.4145278036594391,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 8,
      "model": "linear",
      "best_epoch": 28,
      "test_mse": 4.699235772814406,
      "test_mse_normalized": 0.8465386629104614,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 8,
      "model": "mlp",
      "best_epoch": 25,
      "test_mse": 3.0942230611788197,
      "test_mse_normalized": 0.557405412197113,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 8,
      "model": "uniform_ensemble",
      "best_epoch": 25,
      "test_mse": 4.826209048870152,
      "test_mse_normalized": 0.8694121241569519,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 8,
      "model": "aggregate_moe",
      "best_epoch": 40,
      "test_mse": 3.798417135395995,
      "test_mse_normalized": 0.684261679649353,
      "gate_accuracy": 0.508544921875,
      "gate_entropy": 0.4377371668815613,
      "largest_expert_share": 0.5053983926773071,
      "utilization": [
        0.5053983926773071,
        0.49460160732269287
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 8,
      "model": "sample_weight_moe",
      "best_epoch": 55,
      "test_mse": 3.864414267757653,
      "test_mse_normalized": 0.6961506605148315,
      "gate_accuracy": 0.52392578125,
      "gate_entropy": 0.06074770912528038,
      "largest_expert_share": 0.536580502986908,
      "utilization": [
        0.46341946721076965,
        0.536580502986908
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 8,
      "model": "sample_weight_warmstart",
      "best_epoch": 135,
      "test_mse": 4.089251327876034,
      "test_mse_normalized": 0.7366536855697632,
      "gate_accuracy": 0.543701171875,
      "gate_entropy": 0.0870388001203537,
      "largest_expert_share": 0.676451563835144,
      "utilization": [
        0.676451563835144,
        0.3235483765602112
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 8,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.2790272525185684,
      "test_mse_normalized": 0.41055285930633545,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 9,
      "model": "linear",
      "best_epoch": 52,
      "test_mse": 4.688399919147006,
      "test_mse_normalized": 0.8130601644515991,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 9,
      "model": "mlp",
      "best_epoch": 43,
      "test_mse": 3.454717303795878,
      "test_mse_normalized": 0.5991154909133911,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 9,
      "model": "uniform_ensemble",
      "best_epoch": 47,
      "test_mse": 4.7045793475062725,
      "test_mse_normalized": 0.8158659934997559,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 9,
      "model": "aggregate_moe",
      "best_epoch": 60,
      "test_mse": 4.139349020866614,
      "test_mse_normalized": 0.7178440093994141,
      "gate_accuracy": 0.850341796875,
      "gate_entropy": 0.41067230701446533,
      "largest_expert_share": 0.5645327568054199,
      "utilization": [
        0.5645327568054199,
        0.43546727299690247
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 9,
      "model": "sample_weight_moe",
      "best_epoch": 118,
      "test_mse": 2.450292692673287,
      "test_mse_normalized": 0.4249286353588104,
      "gate_accuracy": 0.97998046875,
      "gate_entropy": 0.025701401755213737,
      "largest_expert_share": 0.5094670057296753,
      "utilization": [
        0.5094670057296753,
        0.4905329942703247
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 9,
      "model": "sample_weight_warmstart",
      "best_epoch": 140,
      "test_mse": 3.2614409473089627,
      "test_mse_normalized": 0.5655975937843323,
      "gate_accuracy": 0.889892578125,
      "gate_entropy": 0.032694011926651,
      "largest_expert_share": 0.5684114694595337,
      "utilization": [
        0.5684114694595337,
        0.43158847093582153
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 9,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.3571451520305584,
      "test_mse_normalized": 0.4087750315666199,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 10,
      "model": "linear",
      "best_epoch": 33,
      "test_mse": 4.740677593061943,
      "test_mse_normalized": 0.9950475096702576,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 10,
      "model": "mlp",
      "best_epoch": 36,
      "test_mse": 3.4693079035670364,
      "test_mse_normalized": 0.7281925678253174,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 10,
      "model": "uniform_ensemble",
      "best_epoch": 31,
      "test_mse": 4.760338732205861,
      "test_mse_normalized": 0.9991742968559265,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 10,
      "model": "aggregate_moe",
      "best_epoch": 41,
      "test_mse": 3.954176084668285,
      "test_mse_normalized": 0.829964280128479,
      "gate_accuracy": 0.52783203125,
      "gate_entropy": 0.2952517867088318,
      "largest_expert_share": 0.5795753598213196,
      "utilization": [
        0.5795753598213196,
        0.42042461037635803
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 10,
      "model": "sample_weight_moe",
      "best_epoch": 163,
      "test_mse": 3.1680088122324,
      "test_mse_normalized": 0.6649512052536011,
      "gate_accuracy": 0.9248046875,
      "gate_entropy": 0.01971367374062538,
      "largest_expert_share": 0.530037522315979,
      "utilization": [
        0.4699625074863434,
        0.530037522315979
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 10,
      "model": "sample_weight_warmstart",
      "best_epoch": 204,
      "test_mse": 3.948101339036436,
      "test_mse_normalized": 0.8286892175674438,
      "gate_accuracy": 0.582763671875,
      "gate_entropy": 0.07420568913221359,
      "largest_expert_share": 0.5292022228240967,
      "utilization": [
        0.47079774737358093,
        0.5292022228240967
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 10,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.325038323355249,
      "test_mse_normalized": 0.48801538348197937,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 11,
      "model": "linear",
      "best_epoch": 16,
      "test_mse": 4.665615450292133,
      "test_mse_normalized": 0.8489928245544434,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 11,
      "model": "mlp",
      "best_epoch": 37,
      "test_mse": 3.391771386539541,
      "test_mse_normalized": 0.6171939373016357,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 11,
      "model": "uniform_ensemble",
      "best_epoch": 21,
      "test_mse": 4.6512924266981175,
      "test_mse_normalized": 0.8463864922523499,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 11,
      "model": "aggregate_moe",
      "best_epoch": 40,
      "test_mse": 3.866342452051072,
      "test_mse_normalized": 0.7035506963729858,
      "gate_accuracy": 0.523681640625,
      "gate_entropy": 0.2557927072048187,
      "largest_expert_share": 0.5094427466392517,
      "utilization": [
        0.5094427466392517,
        0.4905572831630707
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 11,
      "model": "sample_weight_moe",
      "best_epoch": 18,
      "test_mse": 3.8336596096844873,
      "test_mse_normalized": 0.6976034641265869,
      "gate_accuracy": 0.512939453125,
      "gate_entropy": 0.1357545256614685,
      "largest_expert_share": 0.5521104335784912,
      "utilization": [
        0.5521104335784912,
        0.44788962602615356
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 11,
      "model": "sample_weight_warmstart",
      "best_epoch": 118,
      "test_mse": 3.963188522553904,
      "test_mse_normalized": 0.7211735844612122,
      "gate_accuracy": 0.626708984375,
      "gate_entropy": 0.0875205397605896,
      "largest_expert_share": 0.5362824201583862,
      "utilization": [
        0.5362824201583862,
        0.4637175500392914
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 11,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.3214677503747625,
      "test_mse_normalized": 0.42243289947509766,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 12,
      "model": "linear",
      "best_epoch": 30,
      "test_mse": 4.746891769982589,
      "test_mse_normalized": 0.82387375831604,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 12,
      "model": "mlp",
      "best_epoch": 36,
      "test_mse": 3.1523947719155254,
      "test_mse_normalized": 0.5471317768096924,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 12,
      "model": "uniform_ensemble",
      "best_epoch": 28,
      "test_mse": 4.748277136406208,
      "test_mse_normalized": 0.824114203453064,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 12,
      "model": "aggregate_moe",
      "best_epoch": 54,
      "test_mse": 2.6060545396460353,
      "test_mse_normalized": 0.4523085951805115,
      "gate_accuracy": 0.971923828125,
      "gate_entropy": 0.0529094934463501,
      "largest_expert_share": 0.5151815414428711,
      "utilization": [
        0.4848184585571289,
        0.5151815414428711
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 12,
      "model": "sample_weight_moe",
      "best_epoch": 128,
      "test_mse": 2.6405729668490943,
      "test_mse_normalized": 0.4582996368408203,
      "gate_accuracy": 0.95263671875,
      "gate_entropy": 0.020728038623929024,
      "largest_expert_share": 0.5158811807632446,
      "utilization": [
        0.5158811807632446,
        0.48411881923675537
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 12,
      "model": "sample_weight_warmstart",
      "best_epoch": 113,
      "test_mse": 2.6593279572126636,
      "test_mse_normalized": 0.46155476570129395,
      "gate_accuracy": 0.952880859375,
      "gate_entropy": 0.04160555079579353,
      "largest_expert_share": 0.5129716992378235,
      "utilization": [
        0.5129716992378235,
        0.4870283007621765
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 12,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.2812522221416054,
      "test_mse_normalized": 0.39593568444252014,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 13,
      "model": "linear",
      "best_epoch": 56,
      "test_mse": 4.515806599891042,
      "test_mse_normalized": 0.7303892970085144,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 13,
      "model": "mlp",
      "best_epoch": 24,
      "test_mse": 3.3198682441989797,
      "test_mse_normalized": 0.5369575023651123,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 13,
      "model": "uniform_ensemble",
      "best_epoch": 36,
      "test_mse": 4.508739123737312,
      "test_mse_normalized": 0.729246199131012,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 13,
      "model": "aggregate_moe",
      "best_epoch": 55,
      "test_mse": 2.9226519831527646,
      "test_mse_normalized": 0.4727115035057068,
      "gate_accuracy": 0.945068359375,
      "gate_entropy": 0.06417043507099152,
      "largest_expert_share": 0.5284937024116516,
      "utilization": [
        0.5284937024116516,
        0.4715062975883484
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 13,
      "model": "sample_weight_moe",
      "best_epoch": 49,
      "test_mse": 2.798764042234403,
      "test_mse_normalized": 0.4526737928390503,
      "gate_accuracy": 0.945556640625,
      "gate_entropy": 0.04185924679040909,
      "largest_expert_share": 0.5195476412773132,
      "utilization": [
        0.5195476412773132,
        0.48045235872268677
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 13,
      "model": "sample_weight_warmstart",
      "best_epoch": 115,
      "test_mse": 2.760560313070526,
      "test_mse_normalized": 0.4464946985244751,
      "gate_accuracy": 0.94677734375,
      "gate_entropy": 0.04024278745055199,
      "largest_expert_share": 0.5276339650154114,
      "utilization": [
        0.5276339650154114,
        0.472366064786911
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 13,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.2495745641642944,
      "test_mse_normalized": 0.363847553730011,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 14,
      "model": "linear",
      "best_epoch": 38,
      "test_mse": 4.492690085047433,
      "test_mse_normalized": 0.8810135722160339,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 14,
      "model": "mlp",
      "best_epoch": 39,
      "test_mse": 3.4741794611617958,
      "test_mse_normalized": 0.6812843084335327,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 14,
      "model": "uniform_ensemble",
      "best_epoch": 40,
      "test_mse": 4.4881265606380225,
      "test_mse_normalized": 0.8801186680793762,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 14,
      "model": "aggregate_moe",
      "best_epoch": 68,
      "test_mse": 2.5678563402027796,
      "test_mse_normalized": 0.5035549402236938,
      "gate_accuracy": 0.954833984375,
      "gate_entropy": 0.06257399171590805,
      "largest_expert_share": 0.5388725996017456,
      "utilization": [
        0.5388725996017456,
        0.461127370595932
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 14,
      "model": "sample_weight_moe",
      "best_epoch": 168,
      "test_mse": 2.6349237964426426,
      "test_mse_normalized": 0.5167068243026733,
      "gate_accuracy": 0.95947265625,
      "gate_entropy": 0.017820660024881363,
      "largest_expert_share": 0.5065310001373291,
      "utilization": [
        0.5065310001373291,
        0.4934689998626709
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 14,
      "model": "sample_weight_warmstart",
      "best_epoch": 102,
      "test_mse": 2.5675578600529168,
      "test_mse_normalized": 0.5034964084625244,
      "gate_accuracy": 0.964599609375,
      "gate_entropy": 0.049751196056604385,
      "largest_expert_share": 0.5240110754966736,
      "utilization": [
        0.5240110754966736,
        0.4759889245033264
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 14,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.3094623480226244,
      "test_mse_normalized": 0.4528840482234955,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 15,
      "model": "linear",
      "best_epoch": 45,
      "test_mse": 4.826850211421094,
      "test_mse_normalized": 0.8774569034576416,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 15,
      "model": "mlp",
      "best_epoch": 29,
      "test_mse": 3.2847605030436693,
      "test_mse_normalized": 0.5971255898475647,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 15,
      "model": "uniform_ensemble",
      "best_epoch": 18,
      "test_mse": 4.795251854801871,
      "test_mse_normalized": 0.8717127442359924,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 15,
      "model": "aggregate_moe",
      "best_epoch": 7,
      "test_mse": 4.564693462574276,
      "test_mse_normalized": 0.8298003077507019,
      "gate_accuracy": 0.57275390625,
      "gate_entropy": 0.3910279870033264,
      "largest_expert_share": 0.5737752914428711,
      "utilization": [
        0.5737752914428711,
        0.4262246787548065
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 15,
      "model": "sample_weight_moe",
      "best_epoch": 48,
      "test_mse": 4.065524283960637,
      "test_mse_normalized": 0.7390580177307129,
      "gate_accuracy": 0.588623046875,
      "gate_entropy": 0.05701380968093872,
      "largest_expert_share": 0.5316215753555298,
      "utilization": [
        0.5316215753555298,
        0.4683784544467926
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 15,
      "model": "sample_weight_warmstart",
      "best_epoch": 142,
      "test_mse": 4.10200743286087,
      "test_mse_normalized": 0.7456901669502258,
      "gate_accuracy": 0.56005859375,
      "gate_entropy": 0.06293128430843353,
      "largest_expert_share": 0.5687432885169983,
      "utilization": [
        0.4312567114830017,
        0.5687432885169983
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 15,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.314466584894074,
      "test_mse_normalized": 0.4207391142845154,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 16,
      "model": "linear",
      "best_epoch": 57,
      "test_mse": 4.799877985934837,
      "test_mse_normalized": 0.7836267948150635,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 16,
      "model": "mlp",
      "best_epoch": 44,
      "test_mse": 3.322636179237355,
      "test_mse_normalized": 0.5424526929855347,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 16,
      "model": "uniform_ensemble",
      "best_epoch": 61,
      "test_mse": 4.792671091303488,
      "test_mse_normalized": 0.7824501991271973,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 16,
      "model": "aggregate_moe",
      "best_epoch": 54,
      "test_mse": 2.5524199431484065,
      "test_mse_normalized": 0.4167073965072632,
      "gate_accuracy": 0.984619140625,
      "gate_entropy": 0.07495619356632233,
      "largest_expert_share": 0.509305477142334,
      "utilization": [
        0.509305477142334,
        0.49069446325302124
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 16,
      "model": "sample_weight_moe",
      "best_epoch": 89,
      "test_mse": 2.667902032524785,
      "test_mse_normalized": 0.4355609714984894,
      "gate_accuracy": 0.961181640625,
      "gate_entropy": 0.03129192441701889,
      "largest_expert_share": 0.5400858521461487,
      "utilization": [
        0.5400858521461487,
        0.4599141478538513
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 16,
      "model": "sample_weight_warmstart",
      "best_epoch": 135,
      "test_mse": 2.676547932991468,
      "test_mse_normalized": 0.4369724988937378,
      "gate_accuracy": 0.969482421875,
      "gate_entropy": 0.03759805113077164,
      "largest_expert_share": 0.5050129890441895,
      "utilization": [
        0.5050129890441895,
        0.49498695135116577
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 16,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.266085893917539,
      "test_mse_normalized": 0.36996057629585266,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 17,
      "model": "linear",
      "best_epoch": 69,
      "test_mse": 4.668618903829039,
      "test_mse_normalized": 0.8211166858673096,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 17,
      "model": "mlp",
      "best_epoch": 25,
      "test_mse": 3.3182273387737395,
      "test_mse_normalized": 0.5836098194122314,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 17,
      "model": "uniform_ensemble",
      "best_epoch": 67,
      "test_mse": 4.6675690107708485,
      "test_mse_normalized": 0.8209320306777954,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 17,
      "model": "aggregate_moe",
      "best_epoch": 40,
      "test_mse": 3.960215254652272,
      "test_mse_normalized": 0.6965226531028748,
      "gate_accuracy": 0.559326171875,
      "gate_entropy": 0.23015090823173523,
      "largest_expert_share": 0.5458977222442627,
      "utilization": [
        0.4541022479534149,
        0.5458977222442627
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 17,
      "model": "sample_weight_moe",
      "best_epoch": 43,
      "test_mse": 3.878153471415203,
      "test_mse_normalized": 0.682089626789093,
      "gate_accuracy": 0.503662109375,
      "gate_entropy": 0.07146013528108597,
      "largest_expert_share": 0.5352787375450134,
      "utilization": [
        0.5352787375450134,
        0.4647212624549866
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 17,
      "model": "sample_weight_warmstart",
      "best_epoch": 240,
      "test_mse": 3.779076846095327,
      "test_mse_normalized": 0.6646640300750732,
      "gate_accuracy": 0.83349609375,
      "gate_entropy": 0.0774933397769928,
      "largest_expert_share": 0.567692220211029,
      "utilization": [
        0.567692220211029,
        0.43230780959129333
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 17,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.4119893380430035,
      "test_mse_normalized": 0.4242206811904907,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 18,
      "model": "linear",
      "best_epoch": 64,
      "test_mse": 4.513937298392349,
      "test_mse_normalized": 0.8570717573165894,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 18,
      "model": "mlp",
      "best_epoch": 35,
      "test_mse": 3.2951111488205886,
      "test_mse_normalized": 0.6256504058837891,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 18,
      "model": "uniform_ensemble",
      "best_epoch": 73,
      "test_mse": 4.507675230072345,
      "test_mse_normalized": 0.8558827638626099,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 18,
      "model": "aggregate_moe",
      "best_epoch": 77,
      "test_mse": 2.623054788622966,
      "test_mse_normalized": 0.49804550409317017,
      "gate_accuracy": 0.985595703125,
      "gate_entropy": 0.07573244720697403,
      "largest_expert_share": 0.5013435482978821,
      "utilization": [
        0.4986564517021179,
        0.5013435482978821
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 18,
      "model": "sample_weight_moe",
      "best_epoch": 66,
      "test_mse": 2.585704162905272,
      "test_mse_normalized": 0.490953654050827,
      "gate_accuracy": 0.973876953125,
      "gate_entropy": 0.036304742097854614,
      "largest_expert_share": 0.516228437423706,
      "utilization": [
        0.516228437423706,
        0.48377156257629395
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 18,
      "model": "sample_weight_warmstart",
      "best_epoch": 119,
      "test_mse": 2.623911004350999,
      "test_mse_normalized": 0.49820807576179504,
      "gate_accuracy": 0.974365234375,
      "gate_entropy": 0.04198427498340607,
      "largest_expert_share": 0.5149982571601868,
      "utilization": [
        0.5149982571601868,
        0.4850017726421356
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 18,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.3092000060434628,
      "test_mse_normalized": 0.43845316767692566,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 19,
      "model": "linear",
      "best_epoch": 44,
      "test_mse": 4.666968453272529,
      "test_mse_normalized": 0.9527472257614136,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 19,
      "model": "mlp",
      "best_epoch": 26,
      "test_mse": 3.316238198736656,
      "test_mse_normalized": 0.6769998073577881,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 19,
      "model": "uniform_ensemble",
      "best_epoch": 43,
      "test_mse": 4.684576833347034,
      "test_mse_normalized": 0.9563419222831726,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 19,
      "model": "aggregate_moe",
      "best_epoch": 34,
      "test_mse": 3.869063202075656,
      "test_mse_normalized": 0.7898573279380798,
      "gate_accuracy": 0.562744140625,
      "gate_entropy": 0.27393838763237,
      "largest_expert_share": 0.5351958274841309,
      "utilization": [
        0.46480420231819153,
        0.5351958274841309
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 19,
      "model": "sample_weight_moe",
      "best_epoch": 54,
      "test_mse": 3.8791752688567405,
      "test_mse_normalized": 0.7919216752052307,
      "gate_accuracy": 0.502685546875,
      "gate_entropy": 0.05134349316358566,
      "largest_expert_share": 0.5335049033164978,
      "utilization": [
        0.5335049033164978,
        0.4664950668811798
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 19,
      "model": "sample_weight_warmstart",
      "best_epoch": 127,
      "test_mse": 3.9291595472167726,
      "test_mse_normalized": 0.8021258115768433,
      "gate_accuracy": 0.51904296875,
      "gate_entropy": 0.0734749287366867,
      "largest_expert_share": 0.5623173713684082,
      "utilization": [
        0.5623173713684082,
        0.4376826286315918
      ]
    },
    {
      "scenario": "heterogeneous_noisy",
      "seed": 19,
      "model": "oracle_partition",
      "best_epoch": 0,
      "test_mse": 2.3774317465287296,
      "test_mse_normalized": 0.48534536361694336,
      "gate_accuracy": null,
      "gate_entropy": null,
      "largest_expert_share": null,
      "utilization": null
    }
  ]
}