{
  "mode": "dense",
  "args": {
    "mode": "dense",
    "supervision": "prefix_query",
    "seed": 53,
    "modulus": 31,
    "observe_mod": 5,
    "observe_prob": 0.3,
    "op_family": "full",
    "query_family": "all",
    "train_min_len": 1,
    "train_max_len": 8,
    "eval_lengths": "4,8,12,16,24",
    "eval_k": "0,1,2,4,8,12,16,24",
    "eval_query_types": "all",
    "state_dim": 256,
    "instr_dim": 128,
    "heads": 4,
    "compiler_layers": 2,
    "dropout": 0.0,
    "train_steps": 2000,
    "batch_size": 256,
    "eval_batch_size": 256,
    "eval_examples": 256,
    "lr": 0.001,
    "weight_decay": 0.0001,
    "query_loss_weight": 1.0,
    "belief_loss_weight": 1.0,
    "belief_distill_prob": 0.1,
    "probe_steps": 800,
    "probe_batch_size": 256,
    "probe_lr": 0.001,
    "probe_weight_decay": 0.0001,
    "probe_dropout": 0.0,
    "max_grad_norm": 1.0,
    "log_every": 250,
    "probe_log_every": 200,
    "output_dir": "experiments/dense_supervision_ladder/runs/main_prefix_query_mod31",
    "checkpoint_dir": "large_artifacts/dense_supervision_ladder/checkpoints/main_prefix_query_mod31"
  },
  "env": {
    "python": "3.12.3",
    "torch": "2.8.0+cu128",
    "device": "cuda",
    "cuda_name": "NVIDIA RTX 6000 Ada Generation"
  },
  "train": [
    {
      "step": 1,
      "loss": 3.4527366161346436,
      "grad_norm": 0.2706105709075928,
      "query_loss": 3.4527366161346436,
      "belief_loss": 0.0,
      "query_target_mass": 0.3547244369983673,
      "query_target_nll": 3.312554121017456,
      "query_top1_on_support": 0.50390625,
      "mean_query_support_size": 10.6796875
    },
    {
      "step": 250,
      "loss": 2.9553160667419434,
      "grad_norm": 0.21936741471290588,
      "query_loss": 2.9553160667419434,
      "belief_loss": 0.0,
      "query_target_mass": 0.4727950394153595,
      "query_target_nll": 3.0838723182678223,
      "query_top1_on_support": 0.55078125,
      "mean_query_support_size": 11.55859375
    },
    {
      "step": 500,
      "loss": 2.8422813415527344,
      "grad_norm": 0.26974430680274963,
      "query_loss": 2.8422813415527344,
      "belief_loss": 0.0,
      "query_target_mass": 0.5215315818786621,
      "query_target_nll": 2.9162304401397705,
      "query_top1_on_support": 0.625,
      "mean_query_support_size": 11.1875
    },
    {
      "step": 750,
      "loss": 2.76812481880188,
      "grad_norm": 0.2723493278026581,
      "query_loss": 2.76812481880188,
      "belief_loss": 0.0,
      "query_target_mass": 0.5363606214523315,
      "query_target_nll": 2.8929309844970703,
      "query_top1_on_support": 0.62109375,
      "mean_query_support_size": 11.765625
    },
    {
      "step": 1000,
      "loss": 2.557262659072876,
      "grad_norm": 0.44269007444381714,
      "query_loss": 2.557262659072876,
      "belief_loss": 0.0,
      "query_target_mass": 0.6047616004943848,
      "query_target_nll": 2.5184977054595947,
      "query_top1_on_support": 0.71875,
      "mean_query_support_size": 11.140625
    },
    {
      "step": 1250,
      "loss": 2.5128209590911865,
      "grad_norm": 0.43726879358291626,
      "query_loss": 2.5128209590911865,
      "belief_loss": 0.0,
      "query_target_mass": 0.6196250915527344,
      "query_target_nll": 2.4306323528289795,
      "query_top1_on_support": 0.7265625,
      "mean_query_support_size": 10.265625
    },
    {
      "step": 1500,
      "loss": 2.392376184463501,
      "grad_norm": 0.4905843436717987,
      "query_loss": 2.392376184463501,
      "belief_loss": 0.0,
      "query_target_mass": 0.6819772720336914,
      "query_target_nll": 2.3456368446350098,
      "query_top1_on_support": 0.78125,
      "mean_query_support_size": 11.32421875
    },
    {
      "step": 1750,
      "loss": 2.4568347930908203,
      "grad_norm": 0.3801973760128021,
      "query_loss": 2.4568347930908203,
      "belief_loss": 0.0,
      "query_target_mass": 0.6981567740440369,
      "query_target_nll": 2.3656301498413086,
      "query_top1_on_support": 0.77734375,
      "mean_query_support_size": 11.48828125
    },
    {
      "step": 2000,
      "loss": 2.4148457050323486,
      "grad_norm": 0.4910218119621277,
      "query_loss": 2.4148457050323486,
      "belief_loss": 0.0,
      "query_target_mass": 0.7071183919906616,
      "query_target_nll": 2.3361167907714844,
      "query_top1_on_support": 0.81640625,
      "mean_query_support_size": 11.22265625
    }
  ],
  "probe_train": [
    {
      "step": 1,
      "loss": 6.888701438903809,
      "grad_norm": 0.22798120975494385,
      "probe_belief_target_mass": 0.019772732630372047,
      "probe_belief_target_nll": 6.888701438903809,
      "probe_belief_top1_on_support": 0.014756944961845875,
      "mean_belief_support_size": 19.06727409362793
    },
    {
      "step": 200,
      "loss": 3.7351083755493164,
      "grad_norm": 0.5478934645652771,
      "probe_belief_target_mass": 0.5141795873641968,
      "probe_belief_target_nll": 3.7351083755493164,
      "probe_belief_top1_on_support": 0.6822916865348816,
      "mean_belief_support_size": 17.91232681274414
    },
    {
      "step": 400,
      "loss": 3.553785800933838,
      "grad_norm": 0.47997817397117615,
      "probe_belief_target_mass": 0.59576815366745,
      "probe_belief_target_nll": 3.553785800933838,
      "probe_belief_top1_on_support": 0.7183159589767456,
      "mean_belief_support_size": 17.484375
    },
    {
      "step": 600,
      "loss": 3.566924571990967,
      "grad_norm": 0.48027053475379944,
      "probe_belief_target_mass": 0.614950954914093,
      "probe_belief_target_nll": 3.566924571990967,
      "probe_belief_top1_on_support": 0.7404513955116272,
      "mean_belief_support_size": 17.92447853088379
    },
    {
      "step": 800,
      "loss": 3.5215656757354736,
      "grad_norm": 0.48000526428222656,
      "probe_belief_target_mass": 0.6367539167404175,
      "probe_belief_target_nll": 3.5215656757354736,
      "probe_belief_top1_on_support": 0.7204861044883728,
      "mean_belief_support_size": 18.07725715637207
    }
  ],
  "eval": [
    {
      "step": 2000,
      "rows": [
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.38705766201019287,
          "query_target_nll": 3.4422805309295654,
          "query_top1_on_support": 0.3828125,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.023775342851877213,
          "probe_belief_target_nll": 19.761674880981445,
          "probe_belief_top1_on_support": 0.02734375,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014472655951976776,
          "decoder_belief_target_nll": 6.883075714111328,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.4353370666503906,
          "query_target_nll": 5.367556095123291,
          "query_top1_on_support": 0.46484375,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.04659551382064819,
          "probe_belief_target_nll": 22.26012420654297,
          "probe_belief_top1_on_support": 0.05078125,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014465348795056343,
          "decoder_belief_target_nll": 6.888731002807617,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.5370990037918091,
          "query_target_nll": 4.593379020690918,
          "query_top1_on_support": 0.55078125,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.11001268029212952,
          "probe_belief_target_nll": 16.383955001831055,
          "probe_belief_top1_on_support": 0.125,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014426682144403458,
          "decoder_belief_target_nll": 6.889157772064209,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.7659492492675781,
          "query_target_nll": 2.3938450813293457,
          "query_top1_on_support": 0.83203125,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.37117692828178406,
          "probe_belief_target_nll": 3.9751386642456055,
          "probe_belief_top1_on_support": 0.4921875,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014404285699129105,
          "decoder_belief_target_nll": 6.888184070587158,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.7659492492675781,
          "query_target_nll": 2.3938450813293457,
          "query_top1_on_support": 0.83203125,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.37117692828178406,
          "probe_belief_target_nll": 3.9751386642456055,
          "probe_belief_top1_on_support": 0.4921875,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014404285699129105,
          "decoder_belief_target_nll": 6.888184070587158,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.7659492492675781,
          "query_target_nll": 2.3938450813293457,
          "query_top1_on_support": 0.83203125,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.37117692828178406,
          "probe_belief_target_nll": 3.9751386642456055,
          "probe_belief_top1_on_support": 0.4921875,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014404285699129105,
          "decoder_belief_target_nll": 6.888184070587158,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.7659492492675781,
          "query_target_nll": 2.3938450813293457,
          "query_top1_on_support": 0.83203125,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.37117692828178406,
          "probe_belief_target_nll": 3.9751386642456055,
          "probe_belief_top1_on_support": 0.4921875,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014404285699129105,
          "decoder_belief_target_nll": 6.888184070587158,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.7659492492675781,
          "query_target_nll": 2.3938450813293457,
          "query_top1_on_support": 0.83203125,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.37117692828178406,
          "probe_belief_target_nll": 3.9751386642456055,
          "probe_belief_top1_on_support": 0.4921875,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014404285699129105,
          "decoder_belief_target_nll": 6.888184070587158,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.35879045724868774,
          "query_target_nll": 3.4360742568969727,
          "query_top1_on_support": 0.39453125,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.02438316121697426,
          "probe_belief_target_nll": 19.46014404296875,
          "probe_belief_top1_on_support": 0.0234375,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013883915729820728,
          "decoder_belief_target_nll": 6.88845682144165,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.42012789845466614,
          "query_target_nll": 5.533373832702637,
          "query_top1_on_support": 0.44140625,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.07515901327133179,
          "probe_belief_target_nll": 20.747650146484375,
          "probe_belief_top1_on_support": 0.0625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013872548937797546,
          "decoder_belief_target_nll": 6.888821125030518,
          "decoder_belief_top1_on_support": 0.02734375,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.5174762010574341,
          "query_target_nll": 4.855392932891846,
          "query_top1_on_support": 0.55859375,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.12379883229732513,
          "probe_belief_target_nll": 16.84773063659668,
          "probe_belief_top1_on_support": 0.12890625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013870046474039555,
          "decoder_belief_target_nll": 6.882267475128174,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.7792646288871765,
          "query_target_nll": 2.2371184825897217,
          "query_top1_on_support": 0.8515625,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.4056084454059601,
          "probe_belief_target_nll": 3.7160611152648926,
          "probe_belief_top1_on_support": 0.515625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013856426812708378,
          "decoder_belief_target_nll": 6.888169288635254,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.7792646288871765,
          "query_target_nll": 2.2371184825897217,
          "query_top1_on_support": 0.8515625,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.4056084454059601,
          "probe_belief_target_nll": 3.7160611152648926,
          "probe_belief_top1_on_support": 0.515625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013856426812708378,
          "decoder_belief_target_nll": 6.888169288635254,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.7792646288871765,
          "query_target_nll": 2.2371184825897217,
          "query_top1_on_support": 0.8515625,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.4056084454059601,
          "probe_belief_target_nll": 3.7160611152648926,
          "probe_belief_top1_on_support": 0.515625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013856426812708378,
          "decoder_belief_target_nll": 6.888169288635254,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.7792646288871765,
          "query_target_nll": 2.2371184825897217,
          "query_top1_on_support": 0.8515625,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.4056084454059601,
          "probe_belief_target_nll": 3.7160611152648926,
          "probe_belief_top1_on_support": 0.515625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013856426812708378,
          "decoder_belief_target_nll": 6.888169288635254,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.7792646288871765,
          "query_target_nll": 2.2371184825897217,
          "query_top1_on_support": 0.8515625,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.4056084454059601,
          "probe_belief_target_nll": 3.7160611152648926,
          "probe_belief_top1_on_support": 0.515625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013856426812708378,
          "decoder_belief_target_nll": 6.888169288635254,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.369209349155426,
          "query_target_nll": 3.43990421295166,
          "query_top1_on_support": 0.3671875,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.03135659545660019,
          "probe_belief_target_nll": 18.225967407226562,
          "probe_belief_top1_on_support": 0.01953125,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.01265290379524231,
          "decoder_belief_target_nll": 6.881394386291504,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.3821432292461395,
          "query_target_nll": 4.269763946533203,
          "query_top1_on_support": 0.375,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.08863086998462677,
          "probe_belief_target_nll": 20.06021499633789,
          "probe_belief_top1_on_support": 0.1015625,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.01258667279034853,
          "decoder_belief_target_nll": 6.882438659667969,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.40400439500808716,
          "query_target_nll": 4.421138763427734,
          "query_top1_on_support": 0.40625,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.135393887758255,
          "probe_belief_target_nll": 16.21369743347168,
          "probe_belief_top1_on_support": 0.1640625,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.012624487280845642,
          "decoder_belief_target_nll": 6.882946968078613,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.6412630081176758,
          "query_target_nll": 2.7085342407226562,
          "query_top1_on_support": 0.7265625,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.3959798812866211,
          "probe_belief_target_nll": 3.7639708518981934,
          "probe_belief_top1_on_support": 0.53125,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.012584710493683815,
          "decoder_belief_target_nll": 6.895322322845459,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.6412630081176758,
          "query_target_nll": 2.7085342407226562,
          "query_top1_on_support": 0.7265625,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.3959798812866211,
          "probe_belief_target_nll": 3.7639708518981934,
          "probe_belief_top1_on_support": 0.53125,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.012584710493683815,
          "decoder_belief_target_nll": 6.895322322845459,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.6412630081176758,
          "query_target_nll": 2.7085342407226562,
          "query_top1_on_support": 0.7265625,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.3959798812866211,
          "probe_belief_target_nll": 3.7639708518981934,
          "probe_belief_top1_on_support": 0.53125,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.012584710493683815,
          "decoder_belief_target_nll": 6.895322322845459,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.6412630081176758,
          "query_target_nll": 2.7085342407226562,
          "query_top1_on_support": 0.7265625,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.3959798812866211,
          "probe_belief_target_nll": 3.7639708518981934,
          "probe_belief_top1_on_support": 0.53125,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.012584710493683815,
          "decoder_belief_target_nll": 6.895322322845459,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.6412630081176758,
          "query_target_nll": 2.7085342407226562,
          "query_top1_on_support": 0.7265625,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.3959798812866211,
          "probe_belief_target_nll": 3.7639708518981934,
          "probe_belief_top1_on_support": 0.53125,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.012584710493683815,
          "decoder_belief_target_nll": 6.895322322845459,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.42962419986724854,
          "query_target_nll": 15.341827392578125,
          "query_top1_on_support": 0.4296875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.034793250262737274,
          "probe_belief_target_nll": 19.32046890258789,
          "probe_belief_top1_on_support": 0.0390625,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.015007955953478813,
          "decoder_belief_target_nll": 6.8856377601623535,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.4202759861946106,
          "query_target_nll": 16.399024963378906,
          "query_top1_on_support": 0.4296875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.06750762462615967,
          "probe_belief_target_nll": 21.911357879638672,
          "probe_belief_top1_on_support": 0.0703125,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.014946565963327885,
          "decoder_belief_target_nll": 6.889062404632568,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.447967529296875,
          "query_target_nll": 11.275036811828613,
          "query_top1_on_support": 0.44921875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.1378546804189682,
          "probe_belief_target_nll": 16.333772659301758,
          "probe_belief_top1_on_support": 0.15625,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.014939721673727036,
          "decoder_belief_target_nll": 6.892369747161865,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.7249109745025635,
          "query_target_nll": 2.286309242248535,
          "query_top1_on_support": 0.80859375,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.4085642099380493,
          "probe_belief_target_nll": 3.8549559116363525,
          "probe_belief_top1_on_support": 0.5546875,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.01496715284883976,
          "decoder_belief_target_nll": 6.8933820724487305,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.7249109745025635,
          "query_target_nll": 2.286309242248535,
          "query_top1_on_support": 0.80859375,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.4085642099380493,
          "probe_belief_target_nll": 3.8549559116363525,
          "probe_belief_top1_on_support": 0.5546875,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.01496715284883976,
          "decoder_belief_target_nll": 6.8933820724487305,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.7249109745025635,
          "query_target_nll": 2.286309242248535,
          "query_top1_on_support": 0.80859375,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.4085642099380493,
          "probe_belief_target_nll": 3.8549559116363525,
          "probe_belief_top1_on_support": 0.5546875,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.01496715284883976,
          "decoder_belief_target_nll": 6.8933820724487305,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.7249109745025635,
          "query_target_nll": 2.286309242248535,
          "query_top1_on_support": 0.80859375,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.4085642099380493,
          "probe_belief_target_nll": 3.8549559116363525,
          "probe_belief_top1_on_support": 0.5546875,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.01496715284883976,
          "decoder_belief_target_nll": 6.8933820724487305,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.7249109745025635,
          "query_target_nll": 2.286309242248535,
          "query_top1_on_support": 0.80859375,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.4085642099380493,
          "probe_belief_target_nll": 3.8549559116363525,
          "probe_belief_top1_on_support": 0.5546875,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.01496715284883976,
          "decoder_belief_target_nll": 6.8933820724487305,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.1708594113588333,
          "query_target_nll": 3.438274383544922,
          "query_top1_on_support": 0.1875,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.007624031510204077,
          "probe_belief_target_nll": 19.956857681274414,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.007429078686982393,
          "decoder_belief_target_nll": 6.892161846160889,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.1974165141582489,
          "query_target_nll": 7.106866836547852,
          "query_top1_on_support": 0.21484375,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.01159866526722908,
          "probe_belief_target_nll": 24.91458511352539,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.007435379084199667,
          "decoder_belief_target_nll": 6.879210472106934,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.21260371804237366,
          "query_target_nll": 7.233747482299805,
          "query_top1_on_support": 0.21484375,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.015389278531074524,
          "probe_belief_target_nll": 21.311294555664062,
          "probe_belief_top1_on_support": 0.01953125,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.007456360384821892,
          "decoder_belief_target_nll": 6.877274513244629,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.23280656337738037,
          "query_target_nll": 8.238622665405273,
          "query_top1_on_support": 0.2578125,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.029274556785821915,
          "probe_belief_target_nll": 16.760391235351562,
          "probe_belief_top1_on_support": 0.03515625,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.007509816437959671,
          "decoder_belief_target_nll": 6.870516777038574,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.48615944385528564,
          "query_target_nll": 2.157172918319702,
          "query_top1_on_support": 0.5625,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.15027503669261932,
          "probe_belief_target_nll": 4.0262274742126465,
          "probe_belief_top1_on_support": 0.265625,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.0074435994029045105,
          "decoder_belief_target_nll": 6.882933616638184,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.48615944385528564,
          "query_target_nll": 2.157172918319702,
          "query_top1_on_support": 0.5625,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.15027503669261932,
          "probe_belief_target_nll": 4.0262274742126465,
          "probe_belief_top1_on_support": 0.265625,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.0074435994029045105,
          "decoder_belief_target_nll": 6.882933616638184,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.48615944385528564,
          "query_target_nll": 2.157172918319702,
          "query_top1_on_support": 0.5625,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.15027503669261932,
          "probe_belief_target_nll": 4.0262274742126465,
          "probe_belief_top1_on_support": 0.265625,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.0074435994029045105,
          "decoder_belief_target_nll": 6.882933616638184,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.48615944385528564,
          "query_target_nll": 2.157172918319702,
          "query_top1_on_support": 0.5625,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.15027503669261932,
          "probe_belief_target_nll": 4.0262274742126465,
          "probe_belief_top1_on_support": 0.265625,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.0074435994029045105,
          "decoder_belief_target_nll": 6.882933616638184,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.18815207481384277,
          "query_target_nll": 3.440260887145996,
          "query_top1_on_support": 0.19921875,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.007323496509343386,
          "probe_belief_target_nll": 19.54556655883789,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007351590320467949,
          "decoder_belief_target_nll": 6.897672653198242,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.20082512497901917,
          "query_target_nll": 6.873013496398926,
          "query_top1_on_support": 0.203125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.01240649912506342,
          "probe_belief_target_nll": 24.224842071533203,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007363783195614815,
          "decoder_belief_target_nll": 6.8907365798950195,
          "decoder_belief_top1_on_support": 0.02734375,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.21924051642417908,
          "query_target_nll": 7.279502868652344,
          "query_top1_on_support": 0.234375,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.02029505930840969,
          "probe_belief_target_nll": 20.627397537231445,
          "probe_belief_top1_on_support": 0.02734375,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007424953393638134,
          "decoder_belief_target_nll": 6.881694316864014,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.2716737985610962,
          "query_target_nll": 7.187814712524414,
          "query_top1_on_support": 0.28515625,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.02498161420226097,
          "probe_belief_target_nll": 17.26917266845703,
          "probe_belief_top1_on_support": 0.02734375,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.0073839155957102776,
          "decoder_belief_target_nll": 6.883820533752441,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.5234196186065674,
          "query_target_nll": 2.266279697418213,
          "query_top1_on_support": 0.625,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.1530669927597046,
          "probe_belief_target_nll": 4.120212078094482,
          "probe_belief_top1_on_support": 0.203125,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007353910245001316,
          "decoder_belief_target_nll": 6.893914699554443,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.5234196186065674,
          "query_target_nll": 2.266279697418213,
          "query_top1_on_support": 0.625,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.1530669927597046,
          "probe_belief_target_nll": 4.120212078094482,
          "probe_belief_top1_on_support": 0.203125,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007353910245001316,
          "decoder_belief_target_nll": 6.893914699554443,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.5234196186065674,
          "query_target_nll": 2.266279697418213,
          "query_top1_on_support": 0.625,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.1530669927597046,
          "probe_belief_target_nll": 4.120212078094482,
          "probe_belief_top1_on_support": 0.203125,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007353910245001316,
          "decoder_belief_target_nll": 6.893914699554443,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.5234196186065674,
          "query_target_nll": 2.266279697418213,
          "query_top1_on_support": 0.625,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.1530669927597046,
          "probe_belief_target_nll": 4.120212078094482,
          "probe_belief_top1_on_support": 0.203125,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007353910245001316,
          "decoder_belief_target_nll": 6.893914699554443,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.17293241620063782,
          "query_target_nll": 3.436974048614502,
          "query_top1_on_support": 0.19921875,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.009891730733215809,
          "probe_belief_target_nll": 20.210498809814453,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.0060095530934631824,
          "decoder_belief_target_nll": 6.883863925933838,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.17760148644447327,
          "query_target_nll": 4.483442783355713,
          "query_top1_on_support": 0.16015625,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.007999001070857048,
          "probe_belief_target_nll": 25.262100219726562,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.005976366810500622,
          "decoder_belief_target_nll": 6.8923659324646,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.16798873245716095,
          "query_target_nll": 4.889930725097656,
          "query_top1_on_support": 0.15625,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.012068947777152061,
          "probe_belief_target_nll": 21.681987762451172,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.005985239055007696,
          "decoder_belief_target_nll": 6.8893256187438965,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.18335244059562683,
          "query_target_nll": 5.112384796142578,
          "query_top1_on_support": 0.2109375,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.028168626129627228,
          "probe_belief_target_nll": 16.98348617553711,
          "probe_belief_top1_on_support": 0.03125,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.005944425240159035,
          "decoder_belief_target_nll": 6.9054155349731445,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.37715545296669006,
          "query_target_nll": 2.5968496799468994,
          "query_top1_on_support": 0.48828125,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.14692886173725128,
          "probe_belief_target_nll": 4.138134479522705,
          "probe_belief_top1_on_support": 0.24609375,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.0059567163698375225,
          "decoder_belief_target_nll": 6.8985724449157715,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.37715545296669006,
          "query_target_nll": 2.5968496799468994,
          "query_top1_on_support": 0.48828125,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.14692886173725128,
          "probe_belief_target_nll": 4.138134479522705,
          "probe_belief_top1_on_support": 0.24609375,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.0059567163698375225,
          "decoder_belief_target_nll": 6.8985724449157715,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.37715545296669006,
          "query_target_nll": 2.5968496799468994,
          "query_top1_on_support": 0.48828125,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.14692886173725128,
          "probe_belief_target_nll": 4.138134479522705,
          "probe_belief_top1_on_support": 0.24609375,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.0059567163698375225,
          "decoder_belief_target_nll": 6.8985724449157715,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.37715545296669006,
          "query_target_nll": 2.5968496799468994,
          "query_top1_on_support": 0.48828125,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.14692886173725128,
          "probe_belief_target_nll": 4.138134479522705,
          "probe_belief_top1_on_support": 0.24609375,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.0059567163698375225,
          "decoder_belief_target_nll": 6.8985724449157715,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.179687038064003,
          "query_target_nll": 16.295040130615234,
          "query_top1_on_support": 0.1796875,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.006235072389245033,
          "probe_belief_target_nll": 20.384193420410156,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.0072906846180558205,
          "decoder_belief_target_nll": 6.872095584869385,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.19173365831375122,
          "query_target_nll": 17.849843978881836,
          "query_top1_on_support": 0.203125,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.016051705926656723,
          "probe_belief_target_nll": 24.484886169433594,
          "probe_belief_top1_on_support": 0.01953125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007307816296815872,
          "decoder_belief_target_nll": 6.877316951751709,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.21616721153259277,
          "query_target_nll": 14.532102584838867,
          "query_top1_on_support": 0.19921875,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.0212603397667408,
          "probe_belief_target_nll": 21.778732299804688,
          "probe_belief_top1_on_support": 0.03125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007277940399944782,
          "decoder_belief_target_nll": 6.88322114944458,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.217250794172287,
          "query_target_nll": 9.398311614990234,
          "query_top1_on_support": 0.2265625,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.023295501247048378,
          "probe_belief_target_nll": 17.408466339111328,
          "probe_belief_top1_on_support": 0.03125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007231113966554403,
          "decoder_belief_target_nll": 6.894679546356201,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.4658415913581848,
          "query_target_nll": 2.4701361656188965,
          "query_top1_on_support": 0.5859375,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.15983015298843384,
          "probe_belief_target_nll": 4.058141708374023,
          "probe_belief_top1_on_support": 0.23828125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007205056957900524,
          "decoder_belief_target_nll": 6.8897552490234375,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.4658415913581848,
          "query_target_nll": 2.4701361656188965,
          "query_top1_on_support": 0.5859375,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.15983015298843384,
          "probe_belief_target_nll": 4.058141708374023,
          "probe_belief_top1_on_support": 0.23828125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007205056957900524,
          "decoder_belief_target_nll": 6.8897552490234375,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.4658415913581848,
          "query_target_nll": 2.4701361656188965,
          "query_top1_on_support": 0.5859375,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.15983015298843384,
          "probe_belief_target_nll": 4.058141708374023,
          "probe_belief_top1_on_support": 0.23828125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007205056957900524,
          "decoder_belief_target_nll": 6.8897552490234375,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.4658415913581848,
          "query_target_nll": 2.4701361656188965,
          "query_top1_on_support": 0.5859375,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.15983015298843384,
          "probe_belief_target_nll": 4.058141708374023,
          "probe_belief_top1_on_support": 0.23828125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007205056957900524,
          "decoder_belief_target_nll": 6.8897552490234375,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.11276483535766602,
          "query_target_nll": 3.441875457763672,
          "query_top1_on_support": 0.08984375,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.00518011162057519,
          "probe_belief_target_nll": 19.880165100097656,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003915381617844105,
          "decoder_belief_target_nll": 6.898362636566162,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.11401056498289108,
          "query_target_nll": 7.029582977294922,
          "query_top1_on_support": 0.12109375,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.004338055849075317,
          "probe_belief_target_nll": 24.123945236206055,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.00393335148692131,
          "decoder_belief_target_nll": 6.892082691192627,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.13237495720386505,
          "query_target_nll": 7.009456157684326,
          "query_top1_on_support": 0.1484375,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.007267926353961229,
          "probe_belief_target_nll": 20.492698669433594,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003984685055911541,
          "decoder_belief_target_nll": 6.876181602478027,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.1315198391675949,
          "query_target_nll": 7.927379608154297,
          "query_top1_on_support": 0.14453125,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.0053534600883722305,
          "probe_belief_target_nll": 17.826053619384766,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003962852992117405,
          "decoder_belief_target_nll": 6.887714862823486,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.15262067317962646,
          "query_target_nll": 7.159872531890869,
          "query_top1_on_support": 0.17578125,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.011084875091910362,
          "probe_belief_target_nll": 13.964679718017578,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003909060265868902,
          "decoder_belief_target_nll": 6.889837741851807,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.35237014293670654,
          "query_target_nll": 2.3328115940093994,
          "query_top1_on_support": 0.40625,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.07952834665775299,
          "probe_belief_target_nll": 4.2635087966918945,
          "probe_belief_top1_on_support": 0.10546875,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003897815477102995,
          "decoder_belief_target_nll": 6.909980297088623,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.35237014293670654,
          "query_target_nll": 2.3328115940093994,
          "query_top1_on_support": 0.40625,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.07952834665775299,
          "probe_belief_target_nll": 4.2635087966918945,
          "probe_belief_top1_on_support": 0.10546875,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003897815477102995,
          "decoder_belief_target_nll": 6.909980297088623,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.35237014293670654,
          "query_target_nll": 2.3328115940093994,
          "query_top1_on_support": 0.40625,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.07952834665775299,
          "probe_belief_target_nll": 4.2635087966918945,
          "probe_belief_top1_on_support": 0.10546875,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003897815477102995,
          "decoder_belief_target_nll": 6.909980297088623,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.11950834095478058,
          "query_target_nll": 3.4512205123901367,
          "query_top1_on_support": 0.1171875,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.005568355321884155,
          "probe_belief_target_nll": 19.85717010498047,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004829892888665199,
          "decoder_belief_target_nll": 6.896785736083984,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.12343788146972656,
          "query_target_nll": 6.828400135040283,
          "query_top1_on_support": 0.1015625,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.004737775772809982,
          "probe_belief_target_nll": 23.657325744628906,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004839593544602394,
          "decoder_belief_target_nll": 6.879848480224609,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.11222097277641296,
          "query_target_nll": 7.794222831726074,
          "query_top1_on_support": 0.11328125,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.00890122726559639,
          "probe_belief_target_nll": 21.816408157348633,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004840843379497528,
          "decoder_belief_target_nll": 6.888619899749756,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.13898593187332153,
          "query_target_nll": 7.835103511810303,
          "query_top1_on_support": 0.1484375,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.013148553669452667,
          "probe_belief_target_nll": 18.1126708984375,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004810236394405365,
          "decoder_belief_target_nll": 6.886903762817383,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.17666494846343994,
          "query_target_nll": 6.691835403442383,
          "query_top1_on_support": 0.1796875,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.017364999279379845,
          "probe_belief_target_nll": 13.728299140930176,
          "probe_belief_top1_on_support": 0.0234375,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004889604169875383,
          "decoder_belief_target_nll": 6.87407112121582,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.38467466831207275,
          "query_target_nll": 2.2460739612579346,
          "query_top1_on_support": 0.42578125,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.09174764901399612,
          "probe_belief_target_nll": 4.259307384490967,
          "probe_belief_top1_on_support": 0.16796875,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004883850924670696,
          "decoder_belief_target_nll": 6.875706195831299,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.38467466831207275,
          "query_target_nll": 2.2460739612579346,
          "query_top1_on_support": 0.42578125,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.09174764901399612,
          "probe_belief_target_nll": 4.259307384490967,
          "probe_belief_top1_on_support": 0.16796875,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004883850924670696,
          "decoder_belief_target_nll": 6.875706195831299,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.38467466831207275,
          "query_target_nll": 2.2460739612579346,
          "query_top1_on_support": 0.42578125,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.09174764901399612,
          "probe_belief_target_nll": 4.259307384490967,
          "probe_belief_top1_on_support": 0.16796875,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004883850924670696,
          "decoder_belief_target_nll": 6.875706195831299,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.1288180947303772,
          "query_target_nll": 3.4412217140197754,
          "query_top1_on_support": 0.09765625,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.003570724744349718,
          "probe_belief_target_nll": 20.108592987060547,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.0045904358848929405,
          "decoder_belief_target_nll": 6.894543647766113,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.1225067526102066,
          "query_target_nll": 4.531197547912598,
          "query_top1_on_support": 0.125,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.0069199963472783566,
          "probe_belief_target_nll": 24.85212516784668,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004555365536361933,
          "decoder_belief_target_nll": 6.8941192626953125,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.1344933807849884,
          "query_target_nll": 4.790386199951172,
          "query_top1_on_support": 0.140625,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.007465899456292391,
          "probe_belief_target_nll": 22.036052703857422,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004594624508172274,
          "decoder_belief_target_nll": 6.889666557312012,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.13253086805343628,
          "query_target_nll": 5.043881416320801,
          "query_top1_on_support": 0.140625,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.005762216169387102,
          "probe_belief_target_nll": 18.720703125,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004602443426847458,
          "decoder_belief_target_nll": 6.886865615844727,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.13638928532600403,
          "query_target_nll": 5.294670104980469,
          "query_top1_on_support": 0.12109375,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.01882031559944153,
          "probe_belief_target_nll": 14.781570434570312,
          "probe_belief_top1_on_support": 0.0234375,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004587552510201931,
          "decoder_belief_target_nll": 6.899934768676758,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.3008682429790497,
          "query_target_nll": 2.5704030990600586,
          "query_top1_on_support": 0.3828125,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.09235451370477676,
          "probe_belief_target_nll": 4.1539201736450195,
          "probe_belief_top1_on_support": 0.125,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004591761622577906,
          "decoder_belief_target_nll": 6.897476673126221,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.3008682429790497,
          "query_target_nll": 2.5704030990600586,
          "query_top1_on_support": 0.3828125,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.09235451370477676,
          "probe_belief_target_nll": 4.1539201736450195,
          "probe_belief_top1_on_support": 0.125,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004591761622577906,
          "decoder_belief_target_nll": 6.897476673126221,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.3008682429790497,
          "query_target_nll": 2.5704030990600586,
          "query_top1_on_support": 0.3828125,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.09235451370477676,
          "probe_belief_target_nll": 4.1539201736450195,
          "probe_belief_top1_on_support": 0.125,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004591761622577906,
          "decoder_belief_target_nll": 6.897476673126221,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.15624690055847168,
          "query_target_nll": 15.751791000366211,
          "query_top1_on_support": 0.15625,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.005244688596576452,
          "probe_belief_target_nll": 19.713298797607422,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.0044704619795084,
          "decoder_belief_target_nll": 6.890408992767334,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.1478605568408966,
          "query_target_nll": 16.791156768798828,
          "query_top1_on_support": 0.14453125,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.007483605295419693,
          "probe_belief_target_nll": 23.440528869628906,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004511772654950619,
          "decoder_belief_target_nll": 6.883469581604004,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.12865856289863586,
          "query_target_nll": 13.483122825622559,
          "query_top1_on_support": 0.15234375,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.004707403015345335,
          "probe_belief_target_nll": 21.190412521362305,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004520360380411148,
          "decoder_belief_target_nll": 6.880026340484619,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.1264120191335678,
          "query_target_nll": 9.601359367370605,
          "query_top1_on_support": 0.1328125,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.004089131951332092,
          "probe_belief_target_nll": 18.256046295166016,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004446214996278286,
          "decoder_belief_target_nll": 6.90001106262207,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.1522977501153946,
          "query_target_nll": 6.582859039306641,
          "query_top1_on_support": 0.16015625,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.016224252060055733,
          "probe_belief_target_nll": 13.654471397399902,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004474747460335493,
          "decoder_belief_target_nll": 6.882687091827393,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.3272551894187927,
          "query_target_nll": 2.561065912246704,
          "query_top1_on_support": 0.375,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.08941244333982468,
          "probe_belief_target_nll": 4.200148105621338,
          "probe_belief_top1_on_support": 0.09765625,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004453534260392189,
          "decoder_belief_target_nll": 6.886258125305176,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.3272551894187927,
          "query_target_nll": 2.561065912246704,
          "query_top1_on_support": 0.375,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.08941244333982468,
          "probe_belief_target_nll": 4.200148105621338,
          "probe_belief_top1_on_support": 0.09765625,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004453534260392189,
          "decoder_belief_target_nll": 6.886258125305176,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.3272551894187927,
          "query_target_nll": 2.561065912246704,
          "query_top1_on_support": 0.375,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.08941244333982468,
          "probe_belief_target_nll": 4.200148105621338,
          "probe_belief_top1_on_support": 0.09765625,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004453534260392189,
          "decoder_belief_target_nll": 6.886258125305176,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.07348477840423584,
          "query_target_nll": 3.452986240386963,
          "query_top1_on_support": 0.05078125,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.002934915479272604,
          "probe_belief_target_nll": 20.415767669677734,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.0026604547165334225,
          "decoder_belief_target_nll": 6.9035444259643555,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.07705361396074295,
          "query_target_nll": 6.906700134277344,
          "query_top1_on_support": 0.0703125,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.004059894476085901,
          "probe_belief_target_nll": 25.165843963623047,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.0026970477774739265,
          "decoder_belief_target_nll": 6.897449493408203,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.06957447528839111,
          "query_target_nll": 7.691681385040283,
          "query_top1_on_support": 0.05078125,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.002034185454249382,
          "probe_belief_target_nll": 21.27972412109375,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.002697181422263384,
          "decoder_belief_target_nll": 6.90098762512207,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.07169881463050842,
          "query_target_nll": 8.121694564819336,
          "query_top1_on_support": 0.05859375,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.0018628359539434314,
          "probe_belief_target_nll": 18.19491195678711,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.002702105790376663,
          "decoder_belief_target_nll": 6.895764350891113,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.0900784581899643,
          "query_target_nll": 7.76046085357666,
          "query_top1_on_support": 0.08984375,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.0041028596460819244,
          "probe_belief_target_nll": 15.74509334564209,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.0026772543787956238,
          "decoder_belief_target_nll": 6.899971961975098,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.12839734554290771,
          "query_target_nll": 7.272041320800781,
          "query_top1_on_support": 0.16015625,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.007800178602337837,
          "probe_belief_target_nll": 14.173428535461426,
          "probe_belief_top1_on_support": 0.03125,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.00268940394744277,
          "decoder_belief_target_nll": 6.896151065826416,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.28160130977630615,
          "query_target_nll": 2.1941614151000977,
          "query_top1_on_support": 0.3359375,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.06209713965654373,
          "probe_belief_target_nll": 4.120807647705078,
          "probe_belief_top1_on_support": 0.09375,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.0026957732625305653,
          "decoder_belief_target_nll": 6.892484188079834,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.28160130977630615,
          "query_target_nll": 2.1941614151000977,
          "query_top1_on_support": 0.3359375,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.06209713965654373,
          "probe_belief_target_nll": 4.120807647705078,
          "probe_belief_top1_on_support": 0.09375,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.0026957732625305653,
          "decoder_belief_target_nll": 6.892484188079834,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.0782669186592102,
          "query_target_nll": 3.448391914367676,
          "query_top1_on_support": 0.06640625,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.003317037597298622,
          "probe_belief_target_nll": 20.473102569580078,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.002909916453063488,
          "decoder_belief_target_nll": 6.902646064758301,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.07741416990756989,
          "query_target_nll": 6.961420059204102,
          "query_top1_on_support": 0.0703125,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.003578848671168089,
          "probe_belief_target_nll": 24.161338806152344,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0029762168414890766,
          "decoder_belief_target_nll": 6.884449005126953,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.07927694916725159,
          "query_target_nll": 7.803620338439941,
          "query_top1_on_support": 0.1015625,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.0037822481244802475,
          "probe_belief_target_nll": 21.203731536865234,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0029760173056274652,
          "decoder_belief_target_nll": 6.8760528564453125,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.07899937033653259,
          "query_target_nll": 8.105550765991211,
          "query_top1_on_support": 0.0859375,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.002757556736469269,
          "probe_belief_target_nll": 17.616884231567383,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0030044815503060818,
          "decoder_belief_target_nll": 6.870024681091309,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.08611300587654114,
          "query_target_nll": 8.490521430969238,
          "query_top1_on_support": 0.07421875,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.0037707285955548286,
          "probe_belief_target_nll": 15.728646278381348,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0029691611416637897,
          "decoder_belief_target_nll": 6.883388519287109,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.12730759382247925,
          "query_target_nll": 7.788299083709717,
          "query_top1_on_support": 0.16796875,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.00781446322798729,
          "probe_belief_target_nll": 14.747979164123535,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.002958457451313734,
          "decoder_belief_target_nll": 6.886859893798828,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.3055473566055298,
          "query_target_nll": 2.198978900909424,
          "query_top1_on_support": 0.390625,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.06119300052523613,
          "probe_belief_target_nll": 4.223379135131836,
          "probe_belief_top1_on_support": 0.09375,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0029708058573305607,
          "decoder_belief_target_nll": 6.875208854675293,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.3055473566055298,
          "query_target_nll": 2.198978900909424,
          "query_top1_on_support": 0.390625,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.06119300052523613,
          "probe_belief_target_nll": 4.223379135131836,
          "probe_belief_top1_on_support": 0.09375,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0029708058573305607,
          "decoder_belief_target_nll": 6.875208854675293,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.07480615377426147,
          "query_target_nll": 3.4384548664093018,
          "query_top1_on_support": 0.06640625,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.0022336929105222225,
          "probe_belief_target_nll": 20.752574920654297,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.0025790343061089516,
          "decoder_belief_target_nll": 6.881529808044434,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.06903809309005737,
          "query_target_nll": 4.680848121643066,
          "query_top1_on_support": 0.05859375,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.003615797031670809,
          "probe_belief_target_nll": 24.98881721496582,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.0025893342681229115,
          "decoder_belief_target_nll": 6.877750396728516,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.06781695783138275,
          "query_target_nll": 4.912189960479736,
          "query_top1_on_support": 0.08203125,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.002162487944588065,
          "probe_belief_target_nll": 22.749919891357422,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.002619115635752678,
          "decoder_belief_target_nll": 6.870867729187012,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.08387811481952667,
          "query_target_nll": 4.860463619232178,
          "query_top1_on_support": 0.1015625,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.0036972928792238235,
          "probe_belief_target_nll": 19.178634643554688,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.0025894981808960438,
          "decoder_belief_target_nll": 6.879504680633545,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.08739548921585083,
          "query_target_nll": 5.412723541259766,
          "query_top1_on_support": 0.078125,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.004780299961566925,
          "probe_belief_target_nll": 15.799546241760254,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.0025395958218723536,
          "decoder_belief_target_nll": 6.896904945373535,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.07966957986354828,
          "query_target_nll": 5.716312885284424,
          "query_top1_on_support": 0.08203125,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.007991540245711803,
          "probe_belief_target_nll": 14.535344123840332,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.0025451357942074537,
          "decoder_belief_target_nll": 6.895641803741455,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.20397691428661346,
          "query_target_nll": 2.587148904800415,
          "query_top1_on_support": 0.2734375,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.06457306444644928,
          "probe_belief_target_nll": 4.198293685913086,
          "probe_belief_top1_on_support": 0.09765625,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.0025600786320865154,
          "decoder_belief_target_nll": 6.891134262084961,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.20397691428661346,
          "query_target_nll": 2.587148904800415,
          "query_top1_on_support": 0.2734375,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.06457306444644928,
          "probe_belief_target_nll": 4.198293685913086,
          "probe_belief_top1_on_support": 0.09765625,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.0025600786320865154,
          "decoder_belief_target_nll": 6.891134262084961,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.09763532876968384,
          "query_target_nll": 16.15235137939453,
          "query_top1_on_support": 0.09765625,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.003254866460338235,
          "probe_belief_target_nll": 20.113357543945312,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.0028198393993079662,
          "decoder_belief_target_nll": 6.8823957443237305,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.07329590618610382,
          "query_target_nll": 18.601478576660156,
          "query_top1_on_support": 0.07421875,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.0027464230079203844,
          "probe_belief_target_nll": 26.000213623046875,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.0028210121672600508,
          "decoder_belief_target_nll": 6.884746551513672,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.08802542090415955,
          "query_target_nll": 13.386114120483398,
          "query_top1_on_support": 0.09375,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.0026879389770329,
          "probe_belief_target_nll": 22.848737716674805,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.0027880328707396984,
          "decoder_belief_target_nll": 6.889951229095459,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.07676629722118378,
          "query_target_nll": 9.577436447143555,
          "query_top1_on_support": 0.0703125,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.0018413644284009933,
          "probe_belief_target_nll": 19.735492706298828,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.0028073040302842855,
          "decoder_belief_target_nll": 6.883782863616943,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.07445725053548813,
          "query_target_nll": 7.877204895019531,
          "query_top1_on_support": 0.07421875,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.0029023061506450176,
          "probe_belief_target_nll": 16.467117309570312,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.0028476607985794544,
          "decoder_belief_target_nll": 6.8751068115234375,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.08445500582456589,
          "query_target_nll": 6.866785049438477,
          "query_top1_on_support": 0.08203125,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.00678869616240263,
          "probe_belief_target_nll": 14.032744407653809,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.002833156380802393,
          "decoder_belief_target_nll": 6.876945495605469,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.24245470762252808,
          "query_target_nll": 2.5814380645751953,
          "query_top1_on_support": 0.30859375,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.058385737240314484,
          "probe_belief_target_nll": 4.322669982910156,
          "probe_belief_top1_on_support": 0.10546875,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.0028187143616378307,
          "decoder_belief_target_nll": 6.888120651245117,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.24245470762252808,
          "query_target_nll": 2.5814380645751953,
          "query_top1_on_support": 0.30859375,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.058385737240314484,
          "probe_belief_target_nll": 4.322669982910156,
          "probe_belief_top1_on_support": 0.10546875,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.0028187143616378307,
          "decoder_belief_target_nll": 6.888120651245117,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.05017354339361191,
          "query_target_nll": 3.4476430416107178,
          "query_top1_on_support": 0.0390625,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.001616863301023841,
          "probe_belief_target_nll": 20.119205474853516,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.001698299078270793,
          "decoder_belief_target_nll": 6.891463279724121,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.04968327656388283,
          "query_target_nll": 8.438928604125977,
          "query_top1_on_support": 0.04296875,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.0016862424090504646,
          "probe_belief_target_nll": 25.49517059326172,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0016870300751179457,
          "decoder_belief_target_nll": 6.904887676239014,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.04869475215673447,
          "query_target_nll": 8.197154998779297,
          "query_top1_on_support": 0.06640625,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.000921140774153173,
          "probe_belief_target_nll": 22.673786163330078,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0017060119425877929,
          "decoder_belief_target_nll": 6.891151428222656,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.05248050019145012,
          "query_target_nll": 8.11836051940918,
          "query_top1_on_support": 0.05859375,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.0021202233619987965,
          "probe_belief_target_nll": 18.11824607849121,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0016904829535633326,
          "decoder_belief_target_nll": 6.894390106201172,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.05046410486102104,
          "query_target_nll": 8.591760635375977,
          "query_top1_on_support": 0.046875,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.002413226291537285,
          "probe_belief_target_nll": 16.076919555664062,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0017197767738252878,
          "decoder_belief_target_nll": 6.876595497131348,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.04422399401664734,
          "query_target_nll": 8.516816139221191,
          "query_top1_on_support": 0.05078125,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.0034813201054930687,
          "probe_belief_target_nll": 15.248098373413086,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0017100072000175714,
          "decoder_belief_target_nll": 6.884110450744629,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.053389228880405426,
          "query_target_nll": 8.403359413146973,
          "query_top1_on_support": 0.04296875,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.002840933622792363,
          "probe_belief_target_nll": 15.126404762268066,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.001691261655651033,
          "decoder_belief_target_nll": 6.8985185623168945,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.21258780360221863,
          "query_target_nll": 2.2957499027252197,
          "query_top1_on_support": 0.23828125,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.04361128807067871,
          "probe_belief_target_nll": 4.271106719970703,
          "probe_belief_top1_on_support": 0.07421875,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.001701124245300889,
          "decoder_belief_target_nll": 6.884617805480957,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.05117081105709076,
          "query_target_nll": 3.4391367435455322,
          "query_top1_on_support": 0.0703125,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.002518337219953537,
          "probe_belief_target_nll": 19.72718620300293,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018666792893782258,
          "decoder_belief_target_nll": 6.888033866882324,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.0489947684109211,
          "query_target_nll": 6.194244861602783,
          "query_top1_on_support": 0.03125,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.002501262351870537,
          "probe_belief_target_nll": 24.461471557617188,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018721958622336388,
          "decoder_belief_target_nll": 6.880016803741455,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.05083242058753967,
          "query_target_nll": 6.851609230041504,
          "query_top1_on_support": 0.046875,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.003745660185813904,
          "probe_belief_target_nll": 20.86083984375,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018780829850584269,
          "decoder_belief_target_nll": 6.868471622467041,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.04570219665765762,
          "query_target_nll": 7.919045448303223,
          "query_top1_on_support": 0.04296875,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0021366195287555456,
          "probe_belief_target_nll": 18.517431259155273,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018819107208400965,
          "decoder_belief_target_nll": 6.868364334106445,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.058645397424697876,
          "query_target_nll": 8.370320320129395,
          "query_top1_on_support": 0.078125,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0015147766098380089,
          "probe_belief_target_nll": 16.876468658447266,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018661112990230322,
          "decoder_belief_target_nll": 6.88861083984375,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.053307753056287766,
          "query_target_nll": 8.289237976074219,
          "query_top1_on_support": 0.0546875,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0014964013826102018,
          "probe_belief_target_nll": 15.700399398803711,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018387166783213615,
          "decoder_belief_target_nll": 6.891017913818359,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.05342286080121994,
          "query_target_nll": 8.437881469726562,
          "query_top1_on_support": 0.03515625,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0028853160329163074,
          "probe_belief_target_nll": 15.506274223327637,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.001841970719397068,
          "decoder_belief_target_nll": 6.8899736404418945,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.2070869505405426,
          "query_target_nll": 2.2706923484802246,
          "query_top1_on_support": 0.25,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.03979146108031273,
          "probe_belief_target_nll": 4.384524345397949,
          "probe_belief_top1_on_support": 0.05859375,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018164580687880516,
          "decoder_belief_target_nll": 6.908079147338867,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.0485808439552784,
          "query_target_nll": 3.446023464202881,
          "query_top1_on_support": 0.03515625,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0014616334810853004,
          "probe_belief_target_nll": 20.63762855529785,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0016624853014945984,
          "decoder_belief_target_nll": 6.875309944152832,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.0516655296087265,
          "query_target_nll": 4.310399055480957,
          "query_top1_on_support": 0.046875,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0007089381106197834,
          "probe_belief_target_nll": 26.036659240722656,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0016572216991335154,
          "decoder_belief_target_nll": 6.879852771759033,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.04622167348861694,
          "query_target_nll": 4.716606616973877,
          "query_top1_on_support": 0.03515625,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0004291324003133923,
          "probe_belief_target_nll": 23.550975799560547,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0016583265969529748,
          "decoder_belief_target_nll": 6.87769889831543,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.05017517879605293,
          "query_target_nll": 5.350135326385498,
          "query_top1_on_support": 0.05078125,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0021135229617357254,
          "probe_belief_target_nll": 19.395191192626953,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0016445205546915531,
          "decoder_belief_target_nll": 6.877184867858887,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.04461327940225601,
          "query_target_nll": 5.5991740226745605,
          "query_top1_on_support": 0.046875,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.001076688407920301,
          "probe_belief_target_nll": 16.42465591430664,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0016522963996976614,
          "decoder_belief_target_nll": 6.875729084014893,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.04636133089661598,
          "query_target_nll": 5.285262107849121,
          "query_top1_on_support": 0.02734375,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0013651885092258453,
          "probe_belief_target_nll": 15.790002822875977,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0016649415483698249,
          "decoder_belief_target_nll": 6.878782749176025,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.05064039304852486,
          "query_target_nll": 5.355766296386719,
          "query_top1_on_support": 0.0625,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0022416184656322002,
          "probe_belief_target_nll": 15.343729972839355,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0016588832950219512,
          "decoder_belief_target_nll": 6.881117820739746,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.14974680542945862,
          "query_target_nll": 2.5899906158447266,
          "query_top1_on_support": 0.23046875,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.044005267322063446,
          "probe_belief_target_nll": 4.2567138671875,
          "probe_belief_top1_on_support": 0.09375,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.001671205973252654,
          "decoder_belief_target_nll": 6.869377136230469,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.058579154312610626,
          "query_target_nll": 15.675984382629395,
          "query_top1_on_support": 0.05859375,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.001728939707390964,
          "probe_belief_target_nll": 19.70183563232422,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0016273008659482002,
          "decoder_belief_target_nll": 6.876195430755615,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.052529819309711456,
          "query_target_nll": 17.74203872680664,
          "query_top1_on_support": 0.05078125,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0019493747968226671,
          "probe_belief_target_nll": 24.729114532470703,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0016124190296977758,
          "decoder_belief_target_nll": 6.8799662590026855,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.035761065781116486,
          "query_target_nll": 15.339466094970703,
          "query_top1_on_support": 0.03515625,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0003421097935643047,
          "probe_belief_target_nll": 23.272377014160156,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0015986895887181163,
          "decoder_belief_target_nll": 6.886950492858887,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.04592133313417435,
          "query_target_nll": 9.601738929748535,
          "query_top1_on_support": 0.0390625,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0028943405486643314,
          "probe_belief_target_nll": 18.56441307067871,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0016191659960895777,
          "decoder_belief_target_nll": 6.882864475250244,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.05179509148001671,
          "query_target_nll": 7.470364093780518,
          "query_top1_on_support": 0.0546875,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0022029494866728783,
          "probe_belief_target_nll": 15.941792488098145,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0016171273309737444,
          "decoder_belief_target_nll": 6.880516052246094,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.05718230828642845,
          "query_target_nll": 6.767477989196777,
          "query_top1_on_support": 0.0625,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0014988547191023827,
          "probe_belief_target_nll": 15.343648910522461,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0016061989590525627,
          "decoder_belief_target_nll": 6.885695934295654,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.046720754355192184,
          "query_target_nll": 6.603315353393555,
          "query_top1_on_support": 0.05859375,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0020820230711251497,
          "probe_belief_target_nll": 15.1683988571167,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0016094316961243749,
          "decoder_belief_target_nll": 6.884510040283203,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "prefix_query",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.17083002626895905,
          "query_target_nll": 2.5915961265563965,
          "query_top1_on_support": 0.20703125,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.044244300574064255,
          "probe_belief_target_nll": 4.328843593597412,
          "probe_belief_top1_on_support": 0.078125,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0015900236321613193,
          "decoder_belief_target_nll": 6.904009819030762,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.55078125
        }
      ]
    }
  ],
  "checkpoints": [
    "large_artifacts/dense_supervision_ladder/checkpoints/main_prefix_query_mod31/checkpoint_final.pt"
  ]
}