{
  "mode": "dense",
  "args": {
    "mode": "dense",
    "supervision": "sparse_belief",
    "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_sparse_belief_mod31",
    "checkpoint_dir": "large_artifacts/dense_supervision_ladder/checkpoints/main_sparse_belief_mod31"
  },
  "env": {
    "python": "3.12.3",
    "torch": "2.8.0+cu128",
    "device": "cuda",
    "cuda_name": "NVIDIA RTX 6000 Ada Generation"
  },
  "train": [
    {
      "step": 1,
      "loss": 10.346678733825684,
      "grad_norm": 0.5029891133308411,
      "query_loss": 3.4527366161346436,
      "belief_loss": 6.893941879272461,
      "query_target_mass": 0.3516972064971924,
      "query_target_nll": 3.3472394943237305,
      "query_top1_on_support": 0.4375,
      "mean_query_support_size": 10.6796875
    },
    {
      "step": 250,
      "loss": 8.023082733154297,
      "grad_norm": 0.790693998336792,
      "query_loss": 2.980053186416626,
      "belief_loss": 5.043029308319092,
      "query_target_mass": 0.46608293056488037,
      "query_target_nll": 3.1213951110839844,
      "query_top1_on_support": 0.515625,
      "mean_query_support_size": 11.55859375
    },
    {
      "step": 500,
      "loss": 7.920917987823486,
      "grad_norm": 0.6890249252319336,
      "query_loss": 2.8972067832946777,
      "belief_loss": 5.023711204528809,
      "query_target_mass": 0.49971306324005127,
      "query_target_nll": 2.97933030128479,
      "query_top1_on_support": 0.56640625,
      "mean_query_support_size": 11.1875
    },
    {
      "step": 750,
      "loss": 7.64913272857666,
      "grad_norm": 0.8053268790245056,
      "query_loss": 2.8448407649993896,
      "belief_loss": 4.80429220199585,
      "query_target_mass": 0.508887529373169,
      "query_target_nll": 3.051349401473999,
      "query_top1_on_support": 0.546875,
      "mean_query_support_size": 11.765625
    },
    {
      "step": 1000,
      "loss": 7.255433559417725,
      "grad_norm": 0.7969368100166321,
      "query_loss": 2.7585806846618652,
      "belief_loss": 4.496852874755859,
      "query_target_mass": 0.5332392454147339,
      "query_target_nll": 2.866823196411133,
      "query_top1_on_support": 0.5859375,
      "mean_query_support_size": 11.140625
    },
    {
      "step": 1250,
      "loss": 7.320431709289551,
      "grad_norm": 0.7849659323692322,
      "query_loss": 2.7434914112091064,
      "belief_loss": 4.576940536499023,
      "query_target_mass": 0.5186703205108643,
      "query_target_nll": 2.7609829902648926,
      "query_top1_on_support": 0.58984375,
      "mean_query_support_size": 10.265625
    },
    {
      "step": 1500,
      "loss": 6.676576614379883,
      "grad_norm": 1.2391321659088135,
      "query_loss": 2.5376083850860596,
      "belief_loss": 4.138968467712402,
      "query_target_mass": 0.6108914613723755,
      "query_target_nll": 2.564842939376831,
      "query_top1_on_support": 0.68359375,
      "mean_query_support_size": 11.32421875
    },
    {
      "step": 1750,
      "loss": 6.535120010375977,
      "grad_norm": 1.3672958612442017,
      "query_loss": 2.5427494049072266,
      "belief_loss": 3.992370367050171,
      "query_target_mass": 0.6518100500106812,
      "query_target_nll": 2.5013632774353027,
      "query_top1_on_support": 0.765625,
      "mean_query_support_size": 11.48828125
    },
    {
      "step": 2000,
      "loss": 6.349118709564209,
      "grad_norm": 1.2815567255020142,
      "query_loss": 2.5030953884124756,
      "belief_loss": 3.8460233211517334,
      "query_target_mass": 0.6406388282775879,
      "query_target_nll": 2.4879150390625,
      "query_top1_on_support": 0.73828125,
      "mean_query_support_size": 11.22265625
    }
  ],
  "probe_train": [
    {
      "step": 1,
      "loss": 6.884690761566162,
      "grad_norm": 0.2188841849565506,
      "probe_belief_target_mass": 0.019869903102517128,
      "probe_belief_target_nll": 6.884690761566162,
      "probe_belief_top1_on_support": 0.0125868059694767,
      "mean_belief_support_size": 19.06727409362793
    },
    {
      "step": 200,
      "loss": 3.9569430351257324,
      "grad_norm": 0.5092809200286865,
      "probe_belief_target_mass": 0.4424940347671509,
      "probe_belief_target_nll": 3.9569430351257324,
      "probe_belief_top1_on_support": 0.5559895634651184,
      "mean_belief_support_size": 17.91232681274414
    },
    {
      "step": 400,
      "loss": 3.8271915912628174,
      "grad_norm": 0.47220146656036377,
      "probe_belief_target_mass": 0.49107837677001953,
      "probe_belief_target_nll": 3.8271915912628174,
      "probe_belief_top1_on_support": 0.581163227558136,
      "mean_belief_support_size": 17.484375
    },
    {
      "step": 600,
      "loss": 3.7699644565582275,
      "grad_norm": 0.4462476372718811,
      "probe_belief_target_mass": 0.5247055888175964,
      "probe_belief_target_nll": 3.7699644565582275,
      "probe_belief_top1_on_support": 0.6362847089767456,
      "mean_belief_support_size": 17.92447853088379
    },
    {
      "step": 800,
      "loss": 3.783337116241455,
      "grad_norm": 0.4843798577785492,
      "probe_belief_target_mass": 0.530522346496582,
      "probe_belief_target_nll": 3.783337116241455,
      "probe_belief_top1_on_support": 0.5933159589767456,
      "mean_belief_support_size": 18.07725715637207
    }
  ],
  "eval": [
    {
      "step": 2000,
      "rows": [
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.3875007927417755,
          "query_target_nll": 3.440969705581665,
          "query_top1_on_support": 0.39453125,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.024484684690833092,
          "probe_belief_target_nll": 19.04496955871582,
          "probe_belief_top1_on_support": 0.05078125,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.023425227031111717,
          "decoder_belief_target_nll": 18.22724151611328,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.43672144412994385,
          "query_target_nll": 5.14561653137207,
          "query_top1_on_support": 0.4296875,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.045280396938323975,
          "probe_belief_target_nll": 19.90811538696289,
          "probe_belief_top1_on_support": 0.0546875,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.044319309294223785,
          "decoder_belief_target_nll": 15.847434043884277,
          "decoder_belief_top1_on_support": 0.05078125,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.5321581959724426,
          "query_target_nll": 4.252549171447754,
          "query_top1_on_support": 0.55859375,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.09684959053993225,
          "probe_belief_target_nll": 14.468912124633789,
          "probe_belief_top1_on_support": 0.09765625,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.08612294495105743,
          "decoder_belief_target_nll": 11.834957122802734,
          "decoder_belief_top1_on_support": 0.09765625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.7215992212295532,
          "query_target_nll": 2.468386173248291,
          "query_top1_on_support": 0.80859375,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.26993387937545776,
          "probe_belief_target_nll": 4.24449348449707,
          "probe_belief_top1_on_support": 0.36328125,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.24888475239276886,
          "decoder_belief_target_nll": 4.352989673614502,
          "decoder_belief_top1_on_support": 0.37890625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.7215992212295532,
          "query_target_nll": 2.468386173248291,
          "query_top1_on_support": 0.80859375,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.26993387937545776,
          "probe_belief_target_nll": 4.24449348449707,
          "probe_belief_top1_on_support": 0.36328125,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.24888475239276886,
          "decoder_belief_target_nll": 4.352989673614502,
          "decoder_belief_top1_on_support": 0.37890625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.7215992212295532,
          "query_target_nll": 2.468386173248291,
          "query_top1_on_support": 0.80859375,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.26993387937545776,
          "probe_belief_target_nll": 4.24449348449707,
          "probe_belief_top1_on_support": 0.36328125,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.24888475239276886,
          "decoder_belief_target_nll": 4.352989673614502,
          "decoder_belief_top1_on_support": 0.37890625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.7215992212295532,
          "query_target_nll": 2.468386173248291,
          "query_top1_on_support": 0.80859375,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.26993387937545776,
          "probe_belief_target_nll": 4.24449348449707,
          "probe_belief_top1_on_support": 0.36328125,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.24888475239276886,
          "decoder_belief_target_nll": 4.352989673614502,
          "decoder_belief_top1_on_support": 0.37890625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.7215992212295532,
          "query_target_nll": 2.468386173248291,
          "query_top1_on_support": 0.80859375,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.26993387937545776,
          "probe_belief_target_nll": 4.24449348449707,
          "probe_belief_top1_on_support": 0.36328125,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.24888475239276886,
          "decoder_belief_target_nll": 4.352989673614502,
          "decoder_belief_top1_on_support": 0.37890625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.3589063584804535,
          "query_target_nll": 3.4375829696655273,
          "query_top1_on_support": 0.38671875,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.025087185204029083,
          "probe_belief_target_nll": 18.72535514831543,
          "probe_belief_top1_on_support": 0.03515625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.02517564594745636,
          "decoder_belief_target_nll": 17.797527313232422,
          "decoder_belief_top1_on_support": 0.03125,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.4214079976081848,
          "query_target_nll": 5.246138572692871,
          "query_top1_on_support": 0.43359375,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.07447656244039536,
          "probe_belief_target_nll": 18.396291732788086,
          "probe_belief_top1_on_support": 0.0859375,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.0727461725473404,
          "decoder_belief_target_nll": 14.802971839904785,
          "decoder_belief_top1_on_support": 0.07421875,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.5083845257759094,
          "query_target_nll": 4.612387657165527,
          "query_top1_on_support": 0.515625,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.10502368211746216,
          "probe_belief_target_nll": 14.896615982055664,
          "probe_belief_top1_on_support": 0.125,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.09556949138641357,
          "decoder_belief_target_nll": 12.009841918945312,
          "decoder_belief_top1_on_support": 0.08984375,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.7437162399291992,
          "query_target_nll": 2.325209617614746,
          "query_top1_on_support": 0.8046875,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.31419429183006287,
          "probe_belief_target_nll": 3.9891862869262695,
          "probe_belief_top1_on_support": 0.44140625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.2872324585914612,
          "decoder_belief_target_nll": 4.139433860778809,
          "decoder_belief_top1_on_support": 0.38671875,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.7437162399291992,
          "query_target_nll": 2.325209617614746,
          "query_top1_on_support": 0.8046875,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.31419429183006287,
          "probe_belief_target_nll": 3.9891862869262695,
          "probe_belief_top1_on_support": 0.44140625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.2872324585914612,
          "decoder_belief_target_nll": 4.139433860778809,
          "decoder_belief_top1_on_support": 0.38671875,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.7437162399291992,
          "query_target_nll": 2.325209617614746,
          "query_top1_on_support": 0.8046875,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.31419429183006287,
          "probe_belief_target_nll": 3.9891862869262695,
          "probe_belief_top1_on_support": 0.44140625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.2872324585914612,
          "decoder_belief_target_nll": 4.139433860778809,
          "decoder_belief_top1_on_support": 0.38671875,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.7437162399291992,
          "query_target_nll": 2.325209617614746,
          "query_top1_on_support": 0.8046875,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.31419429183006287,
          "probe_belief_target_nll": 3.9891862869262695,
          "probe_belief_top1_on_support": 0.44140625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.2872324585914612,
          "decoder_belief_target_nll": 4.139433860778809,
          "decoder_belief_top1_on_support": 0.38671875,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.7437162399291992,
          "query_target_nll": 2.325209617614746,
          "query_top1_on_support": 0.8046875,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.31419429183006287,
          "probe_belief_target_nll": 3.9891862869262695,
          "probe_belief_top1_on_support": 0.44140625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.2872324585914612,
          "decoder_belief_target_nll": 4.139433860778809,
          "decoder_belief_top1_on_support": 0.38671875,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.36911359429359436,
          "query_target_nll": 3.443976640701294,
          "query_top1_on_support": 0.38671875,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.030931979417800903,
          "probe_belief_target_nll": 17.633487701416016,
          "probe_belief_top1_on_support": 0.03515625,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.031489476561546326,
          "decoder_belief_target_nll": 16.966472625732422,
          "decoder_belief_top1_on_support": 0.03515625,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.38265588879585266,
          "query_target_nll": 4.132856845855713,
          "query_top1_on_support": 0.3671875,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.07971543073654175,
          "probe_belief_target_nll": 17.707393646240234,
          "probe_belief_top1_on_support": 0.06640625,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.07861289381980896,
          "decoder_belief_target_nll": 14.504945755004883,
          "decoder_belief_top1_on_support": 0.0859375,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.3965659737586975,
          "query_target_nll": 4.093966960906982,
          "query_top1_on_support": 0.37109375,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.11517485976219177,
          "probe_belief_target_nll": 14.37051773071289,
          "probe_belief_top1_on_support": 0.12109375,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.10834366083145142,
          "decoder_belief_target_nll": 11.700006484985352,
          "decoder_belief_top1_on_support": 0.10546875,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.5703474879264832,
          "query_target_nll": 2.8445568084716797,
          "query_top1_on_support": 0.64453125,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.30277150869369507,
          "probe_belief_target_nll": 3.981679916381836,
          "probe_belief_top1_on_support": 0.40625,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.2770496606826782,
          "decoder_belief_target_nll": 4.122934341430664,
          "decoder_belief_top1_on_support": 0.3515625,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.5703474879264832,
          "query_target_nll": 2.8445568084716797,
          "query_top1_on_support": 0.64453125,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.30277150869369507,
          "probe_belief_target_nll": 3.981679916381836,
          "probe_belief_top1_on_support": 0.40625,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.2770496606826782,
          "decoder_belief_target_nll": 4.122934341430664,
          "decoder_belief_top1_on_support": 0.3515625,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.5703474879264832,
          "query_target_nll": 2.8445568084716797,
          "query_top1_on_support": 0.64453125,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.30277150869369507,
          "probe_belief_target_nll": 3.981679916381836,
          "probe_belief_top1_on_support": 0.40625,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.2770496606826782,
          "decoder_belief_target_nll": 4.122934341430664,
          "decoder_belief_top1_on_support": 0.3515625,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.5703474879264832,
          "query_target_nll": 2.8445568084716797,
          "query_top1_on_support": 0.64453125,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.30277150869369507,
          "probe_belief_target_nll": 3.981679916381836,
          "probe_belief_top1_on_support": 0.40625,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.2770496606826782,
          "decoder_belief_target_nll": 4.122934341430664,
          "decoder_belief_top1_on_support": 0.3515625,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.5703474879264832,
          "query_target_nll": 2.8445568084716797,
          "query_top1_on_support": 0.64453125,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.30277150869369507,
          "probe_belief_target_nll": 3.981679916381836,
          "probe_belief_top1_on_support": 0.40625,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.2770496606826782,
          "decoder_belief_target_nll": 4.122934341430664,
          "decoder_belief_top1_on_support": 0.3515625,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.4296571612358093,
          "query_target_nll": 14.867488861083984,
          "query_top1_on_support": 0.4296875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.03407829999923706,
          "probe_belief_target_nll": 18.473583221435547,
          "probe_belief_top1_on_support": 0.03125,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.03519056737422943,
          "decoder_belief_target_nll": 17.672714233398438,
          "decoder_belief_top1_on_support": 0.03515625,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.40726780891418457,
          "query_target_nll": 12.548791885375977,
          "query_top1_on_support": 0.40625,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.06477857381105423,
          "probe_belief_target_nll": 19.456819534301758,
          "probe_belief_top1_on_support": 0.06640625,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.06328153610229492,
          "decoder_belief_target_nll": 15.532881736755371,
          "decoder_belief_top1_on_support": 0.0625,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.4271826148033142,
          "query_target_nll": 9.052154541015625,
          "query_top1_on_support": 0.4375,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.1117912158370018,
          "probe_belief_target_nll": 14.37959098815918,
          "probe_belief_top1_on_support": 0.12890625,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.10177533328533173,
          "decoder_belief_target_nll": 11.735183715820312,
          "decoder_belief_top1_on_support": 0.10546875,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.6323733329772949,
          "query_target_nll": 2.480210781097412,
          "query_top1_on_support": 0.71875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.29049959778785706,
          "probe_belief_target_nll": 4.1593523025512695,
          "probe_belief_top1_on_support": 0.375,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.25877174735069275,
          "decoder_belief_target_nll": 4.321881294250488,
          "decoder_belief_top1_on_support": 0.3203125,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.6323733329772949,
          "query_target_nll": 2.480210781097412,
          "query_top1_on_support": 0.71875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.29049959778785706,
          "probe_belief_target_nll": 4.1593523025512695,
          "probe_belief_top1_on_support": 0.375,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.25877174735069275,
          "decoder_belief_target_nll": 4.321881294250488,
          "decoder_belief_top1_on_support": 0.3203125,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.6323733329772949,
          "query_target_nll": 2.480210781097412,
          "query_top1_on_support": 0.71875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.29049959778785706,
          "probe_belief_target_nll": 4.1593523025512695,
          "probe_belief_top1_on_support": 0.375,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.25877174735069275,
          "decoder_belief_target_nll": 4.321881294250488,
          "decoder_belief_top1_on_support": 0.3203125,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.6323733329772949,
          "query_target_nll": 2.480210781097412,
          "query_top1_on_support": 0.71875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.29049959778785706,
          "probe_belief_target_nll": 4.1593523025512695,
          "probe_belief_top1_on_support": 0.375,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.25877174735069275,
          "decoder_belief_target_nll": 4.321881294250488,
          "decoder_belief_top1_on_support": 0.3203125,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.6323733329772949,
          "query_target_nll": 2.480210781097412,
          "query_top1_on_support": 0.71875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.29049959778785706,
          "probe_belief_target_nll": 4.1593523025512695,
          "probe_belief_top1_on_support": 0.375,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.25877174735069275,
          "decoder_belief_target_nll": 4.321881294250488,
          "decoder_belief_top1_on_support": 0.3203125,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.1706024557352066,
          "query_target_nll": 3.436906337738037,
          "query_top1_on_support": 0.16796875,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.007643490098416805,
          "probe_belief_target_nll": 19.466833114624023,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.007952200248837471,
          "decoder_belief_target_nll": 18.549083709716797,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.19752246141433716,
          "query_target_nll": 6.546826362609863,
          "query_top1_on_support": 0.2109375,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.012325944378972054,
          "probe_belief_target_nll": 22.259620666503906,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.011836519464850426,
          "decoder_belief_target_nll": 17.616546630859375,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.19947117567062378,
          "query_target_nll": 6.520274639129639,
          "query_top1_on_support": 0.171875,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.012504801154136658,
          "probe_belief_target_nll": 19.247608184814453,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.013551769778132439,
          "decoder_belief_target_nll": 15.210927963256836,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.22358599305152893,
          "query_target_nll": 7.2086358070373535,
          "query_top1_on_support": 0.21484375,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.026085754856467247,
          "probe_belief_target_nll": 14.874342918395996,
          "probe_belief_top1_on_support": 0.0390625,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.02318401262164116,
          "decoder_belief_target_nll": 12.578977584838867,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.4509386420249939,
          "query_target_nll": 2.273059368133545,
          "query_top1_on_support": 0.56640625,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.12398464977741241,
          "probe_belief_target_nll": 4.314483642578125,
          "probe_belief_top1_on_support": 0.19921875,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.11542787402868271,
          "decoder_belief_target_nll": 4.386770248413086,
          "decoder_belief_top1_on_support": 0.17578125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.4509386420249939,
          "query_target_nll": 2.273059368133545,
          "query_top1_on_support": 0.56640625,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.12398464977741241,
          "probe_belief_target_nll": 4.314483642578125,
          "probe_belief_top1_on_support": 0.19921875,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.11542787402868271,
          "decoder_belief_target_nll": 4.386770248413086,
          "decoder_belief_top1_on_support": 0.17578125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.4509386420249939,
          "query_target_nll": 2.273059368133545,
          "query_top1_on_support": 0.56640625,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.12398464977741241,
          "probe_belief_target_nll": 4.314483642578125,
          "probe_belief_top1_on_support": 0.19921875,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.11542787402868271,
          "decoder_belief_target_nll": 4.386770248413086,
          "decoder_belief_top1_on_support": 0.17578125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.4509386420249939,
          "query_target_nll": 2.273059368133545,
          "query_top1_on_support": 0.56640625,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.12398464977741241,
          "probe_belief_target_nll": 4.314483642578125,
          "probe_belief_top1_on_support": 0.19921875,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.11542787402868271,
          "decoder_belief_target_nll": 4.386770248413086,
          "decoder_belief_top1_on_support": 0.17578125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.18847832083702087,
          "query_target_nll": 3.432114601135254,
          "query_top1_on_support": 0.19921875,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.007009274326264858,
          "probe_belief_target_nll": 19.504064559936523,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.00785432942211628,
          "decoder_belief_target_nll": 18.540292739868164,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.199398010969162,
          "query_target_nll": 6.41583251953125,
          "query_top1_on_support": 0.1953125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.01143988873809576,
          "probe_belief_target_nll": 21.42977523803711,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.012483743950724602,
          "decoder_belief_target_nll": 17.3012638092041,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.2124846875667572,
          "query_target_nll": 6.582586288452148,
          "query_top1_on_support": 0.23046875,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.01576620526611805,
          "probe_belief_target_nll": 18.20132827758789,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.015600129961967468,
          "decoder_belief_target_nll": 14.734748840332031,
          "decoder_belief_top1_on_support": 0.0234375,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.27014607191085815,
          "query_target_nll": 6.40441370010376,
          "query_top1_on_support": 0.3125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.024796482175588608,
          "probe_belief_target_nll": 14.894513130187988,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.023792799562215805,
          "decoder_belief_target_nll": 12.703166007995605,
          "decoder_belief_top1_on_support": 0.0234375,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.49653497338294983,
          "query_target_nll": 2.333217144012451,
          "query_top1_on_support": 0.58203125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.12155873328447342,
          "probe_belief_target_nll": 4.397405624389648,
          "probe_belief_top1_on_support": 0.15234375,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.11305311322212219,
          "decoder_belief_target_nll": 4.507279396057129,
          "decoder_belief_top1_on_support": 0.15234375,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.49653497338294983,
          "query_target_nll": 2.333217144012451,
          "query_top1_on_support": 0.58203125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.12155873328447342,
          "probe_belief_target_nll": 4.397405624389648,
          "probe_belief_top1_on_support": 0.15234375,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.11305311322212219,
          "decoder_belief_target_nll": 4.507279396057129,
          "decoder_belief_top1_on_support": 0.15234375,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.49653497338294983,
          "query_target_nll": 2.333217144012451,
          "query_top1_on_support": 0.58203125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.12155873328447342,
          "probe_belief_target_nll": 4.397405624389648,
          "probe_belief_top1_on_support": 0.15234375,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.11305311322212219,
          "decoder_belief_target_nll": 4.507279396057129,
          "decoder_belief_top1_on_support": 0.15234375,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.49653497338294983,
          "query_target_nll": 2.333217144012451,
          "query_top1_on_support": 0.58203125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.12155873328447342,
          "probe_belief_target_nll": 4.397405624389648,
          "probe_belief_top1_on_support": 0.15234375,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.11305311322212219,
          "decoder_belief_target_nll": 4.507279396057129,
          "decoder_belief_top1_on_support": 0.15234375,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.1719774454832077,
          "query_target_nll": 3.4427218437194824,
          "query_top1_on_support": 0.15234375,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.010080765001475811,
          "probe_belief_target_nll": 19.571687698364258,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.010065397247672081,
          "decoder_belief_target_nll": 18.896669387817383,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.17779923975467682,
          "query_target_nll": 4.2880964279174805,
          "query_top1_on_support": 0.203125,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.008135096170008183,
          "probe_belief_target_nll": 22.270095825195312,
          "probe_belief_top1_on_support": 0.01953125,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.006512343883514404,
          "decoder_belief_target_nll": 17.882152557373047,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.17151613533496857,
          "query_target_nll": 4.362220287322998,
          "query_top1_on_support": 0.1640625,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.011537476442754269,
          "probe_belief_target_nll": 19.302173614501953,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.009746475145220757,
          "decoder_belief_target_nll": 15.513193130493164,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.17777255177497864,
          "query_target_nll": 4.645449161529541,
          "query_top1_on_support": 0.19140625,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.025319939479231834,
          "probe_belief_target_nll": 15.213836669921875,
          "probe_belief_top1_on_support": 0.0546875,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.024455517530441284,
          "decoder_belief_target_nll": 13.004908561706543,
          "decoder_belief_top1_on_support": 0.04296875,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.3339286744594574,
          "query_target_nll": 2.738668918609619,
          "query_top1_on_support": 0.41796875,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.11206071823835373,
          "probe_belief_target_nll": 4.395392417907715,
          "probe_belief_top1_on_support": 0.203125,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.10165075957775116,
          "decoder_belief_target_nll": 4.466423988342285,
          "decoder_belief_top1_on_support": 0.17578125,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.3339286744594574,
          "query_target_nll": 2.738668918609619,
          "query_top1_on_support": 0.41796875,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.11206071823835373,
          "probe_belief_target_nll": 4.395392417907715,
          "probe_belief_top1_on_support": 0.203125,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.10165075957775116,
          "decoder_belief_target_nll": 4.466423988342285,
          "decoder_belief_top1_on_support": 0.17578125,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.3339286744594574,
          "query_target_nll": 2.738668918609619,
          "query_top1_on_support": 0.41796875,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.11206071823835373,
          "probe_belief_target_nll": 4.395392417907715,
          "probe_belief_top1_on_support": 0.203125,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.10165075957775116,
          "decoder_belief_target_nll": 4.466423988342285,
          "decoder_belief_top1_on_support": 0.17578125,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.3339286744594574,
          "query_target_nll": 2.738668918609619,
          "query_top1_on_support": 0.41796875,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.11206071823835373,
          "probe_belief_target_nll": 4.395392417907715,
          "probe_belief_top1_on_support": 0.203125,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.10165075957775116,
          "decoder_belief_target_nll": 4.466423988342285,
          "decoder_belief_top1_on_support": 0.17578125,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.1796865463256836,
          "query_target_nll": 15.86644172668457,
          "query_top1_on_support": 0.1796875,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.006553840823471546,
          "probe_belief_target_nll": 20.38970184326172,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.006188365630805492,
          "decoder_belief_target_nll": 19.168895721435547,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.18893635272979736,
          "query_target_nll": 13.664530754089355,
          "query_top1_on_support": 0.18359375,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.01555169653147459,
          "probe_belief_target_nll": 21.660839080810547,
          "probe_belief_top1_on_support": 0.01953125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.015851549804210663,
          "decoder_belief_target_nll": 17.244176864624023,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.21496272087097168,
          "query_target_nll": 11.352614402770996,
          "query_top1_on_support": 0.21484375,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.017878178507089615,
          "probe_belief_target_nll": 19.30864906311035,
          "probe_belief_top1_on_support": 0.01953125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.018546536564826965,
          "decoder_belief_target_nll": 15.368427276611328,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.21760249137878418,
          "query_target_nll": 8.012003898620605,
          "query_top1_on_support": 0.24609375,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.018833838403224945,
          "probe_belief_target_nll": 15.453044891357422,
          "probe_belief_top1_on_support": 0.0234375,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.018054740503430367,
          "decoder_belief_target_nll": 13.120460510253906,
          "decoder_belief_top1_on_support": 0.0234375,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.41286590695381165,
          "query_target_nll": 2.675182819366455,
          "query_top1_on_support": 0.49609375,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.11454309523105621,
          "probe_belief_target_nll": 4.479894638061523,
          "probe_belief_top1_on_support": 0.20703125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.09964849054813385,
          "decoder_belief_target_nll": 4.557765960693359,
          "decoder_belief_top1_on_support": 0.140625,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.41286590695381165,
          "query_target_nll": 2.675182819366455,
          "query_top1_on_support": 0.49609375,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.11454309523105621,
          "probe_belief_target_nll": 4.479894638061523,
          "probe_belief_top1_on_support": 0.20703125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.09964849054813385,
          "decoder_belief_target_nll": 4.557765960693359,
          "decoder_belief_top1_on_support": 0.140625,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.41286590695381165,
          "query_target_nll": 2.675182819366455,
          "query_top1_on_support": 0.49609375,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.11454309523105621,
          "probe_belief_target_nll": 4.479894638061523,
          "probe_belief_top1_on_support": 0.20703125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.09964849054813385,
          "decoder_belief_target_nll": 4.557765960693359,
          "decoder_belief_top1_on_support": 0.140625,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.41286590695381165,
          "query_target_nll": 2.675182819366455,
          "query_top1_on_support": 0.49609375,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.11454309523105621,
          "probe_belief_target_nll": 4.479894638061523,
          "probe_belief_top1_on_support": 0.20703125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.09964849054813385,
          "decoder_belief_target_nll": 4.557765960693359,
          "decoder_belief_top1_on_support": 0.140625,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.11344887316226959,
          "query_target_nll": 3.4335951805114746,
          "query_top1_on_support": 0.12890625,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.00496266596019268,
          "probe_belief_target_nll": 19.658788681030273,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.0051659876480698586,
          "decoder_belief_target_nll": 18.80882453918457,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.11397624015808105,
          "query_target_nll": 6.47315788269043,
          "query_top1_on_support": 0.109375,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.0039999354630708694,
          "probe_belief_target_nll": 21.13372039794922,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.004033536650240421,
          "decoder_belief_target_nll": 17.04985237121582,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.13291800022125244,
          "query_target_nll": 6.548402786254883,
          "query_top1_on_support": 0.15234375,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.006850685924291611,
          "probe_belief_target_nll": 18.266216278076172,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.005891788750886917,
          "decoder_belief_target_nll": 14.93015193939209,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.1271679550409317,
          "query_target_nll": 7.27414608001709,
          "query_top1_on_support": 0.14453125,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.005045966245234013,
          "probe_belief_target_nll": 15.937360763549805,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003688839264214039,
          "decoder_belief_target_nll": 13.65073013305664,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.151889830827713,
          "query_target_nll": 6.422077655792236,
          "query_top1_on_support": 0.1796875,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.009367505088448524,
          "probe_belief_target_nll": 12.810604095458984,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.008340679109096527,
          "decoder_belief_target_nll": 11.598613739013672,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.32908329367637634,
          "query_target_nll": 2.451228380203247,
          "query_top1_on_support": 0.359375,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.0625566765666008,
          "probe_belief_target_nll": 4.605355739593506,
          "probe_belief_top1_on_support": 0.07421875,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.06016342714428902,
          "decoder_belief_target_nll": 4.622800350189209,
          "decoder_belief_top1_on_support": 0.08984375,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.32908329367637634,
          "query_target_nll": 2.451228380203247,
          "query_top1_on_support": 0.359375,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.0625566765666008,
          "probe_belief_target_nll": 4.605355739593506,
          "probe_belief_top1_on_support": 0.07421875,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.06016342714428902,
          "decoder_belief_target_nll": 4.622800350189209,
          "decoder_belief_top1_on_support": 0.08984375,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.32908329367637634,
          "query_target_nll": 2.451228380203247,
          "query_top1_on_support": 0.359375,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.0625566765666008,
          "probe_belief_target_nll": 4.605355739593506,
          "probe_belief_top1_on_support": 0.07421875,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.06016342714428902,
          "decoder_belief_target_nll": 4.622800350189209,
          "decoder_belief_top1_on_support": 0.08984375,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.12032783031463623,
          "query_target_nll": 3.4458491802215576,
          "query_top1_on_support": 0.1171875,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.005839702673256397,
          "probe_belief_target_nll": 19.794395446777344,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.005333170294761658,
          "decoder_belief_target_nll": 18.89089584350586,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.1236012652516365,
          "query_target_nll": 6.406680583953857,
          "query_top1_on_support": 0.1328125,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.005266737192869186,
          "probe_belief_target_nll": 20.935829162597656,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.005310480482876301,
          "decoder_belief_target_nll": 17.04189109802246,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.11125990003347397,
          "query_target_nll": 7.168655872344971,
          "query_top1_on_support": 0.12109375,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.008704837411642075,
          "probe_belief_target_nll": 19.109020233154297,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.008088851347565651,
          "decoder_belief_target_nll": 15.543173789978027,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.13699400424957275,
          "query_target_nll": 7.320117473602295,
          "query_top1_on_support": 0.14453125,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.01154402457177639,
          "probe_belief_target_nll": 15.780960083007812,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.009535056538879871,
          "decoder_belief_target_nll": 13.618119239807129,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.16943268477916718,
          "query_target_nll": 6.383355140686035,
          "query_top1_on_support": 0.19140625,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.014614560641348362,
          "probe_belief_target_nll": 12.661849975585938,
          "probe_belief_top1_on_support": 0.03125,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.013264046981930733,
          "decoder_belief_target_nll": 11.51179027557373,
          "decoder_belief_top1_on_support": 0.03125,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.36286887526512146,
          "query_target_nll": 2.315375804901123,
          "query_top1_on_support": 0.421875,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.06953997910022736,
          "probe_belief_target_nll": 4.527782440185547,
          "probe_belief_top1_on_support": 0.10546875,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.06450187414884567,
          "decoder_belief_target_nll": 4.607532024383545,
          "decoder_belief_top1_on_support": 0.09765625,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.36286887526512146,
          "query_target_nll": 2.315375804901123,
          "query_top1_on_support": 0.421875,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.06953997910022736,
          "probe_belief_target_nll": 4.527782440185547,
          "probe_belief_top1_on_support": 0.10546875,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.06450187414884567,
          "decoder_belief_target_nll": 4.607532024383545,
          "decoder_belief_top1_on_support": 0.09765625,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.36286887526512146,
          "query_target_nll": 2.315375804901123,
          "query_top1_on_support": 0.421875,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.06953997910022736,
          "probe_belief_target_nll": 4.527782440185547,
          "probe_belief_top1_on_support": 0.10546875,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.06450187414884567,
          "decoder_belief_target_nll": 4.607532024383545,
          "decoder_belief_top1_on_support": 0.09765625,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.12847954034805298,
          "query_target_nll": 3.4478890895843506,
          "query_top1_on_support": 0.12890625,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.0032682563178241253,
          "probe_belief_target_nll": 20.222347259521484,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.003290888387709856,
          "decoder_belief_target_nll": 19.359220504760742,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.12242038547992706,
          "query_target_nll": 4.37800407409668,
          "query_top1_on_support": 0.13671875,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.005911804735660553,
          "probe_belief_target_nll": 22.572282791137695,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.006507849786430597,
          "decoder_belief_target_nll": 18.063629150390625,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.1329362690448761,
          "query_target_nll": 4.457697868347168,
          "query_top1_on_support": 0.15234375,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.006311976816505194,
          "probe_belief_target_nll": 19.743663787841797,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.006533990614116192,
          "decoder_belief_target_nll": 16.057052612304688,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.130281463265419,
          "query_target_nll": 4.573910713195801,
          "query_top1_on_support": 0.1171875,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.004481140524148941,
          "probe_belief_target_nll": 16.622343063354492,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004375847522169352,
          "decoder_belief_target_nll": 14.202886581420898,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.13860629498958588,
          "query_target_nll": 4.798951148986816,
          "query_top1_on_support": 0.125,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.012816887348890305,
          "probe_belief_target_nll": 13.183917999267578,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.011666757054626942,
          "decoder_belief_target_nll": 12.035192489624023,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.26077622175216675,
          "query_target_nll": 2.7539002895355225,
          "query_top1_on_support": 0.296875,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.07432469725608826,
          "probe_belief_target_nll": 4.467698097229004,
          "probe_belief_top1_on_support": 0.08984375,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.06852079182863235,
          "decoder_belief_target_nll": 4.5663299560546875,
          "decoder_belief_top1_on_support": 0.1015625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.26077622175216675,
          "query_target_nll": 2.7539002895355225,
          "query_top1_on_support": 0.296875,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.07432469725608826,
          "probe_belief_target_nll": 4.467698097229004,
          "probe_belief_top1_on_support": 0.08984375,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.06852079182863235,
          "decoder_belief_target_nll": 4.5663299560546875,
          "decoder_belief_top1_on_support": 0.1015625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.26077622175216675,
          "query_target_nll": 2.7539002895355225,
          "query_top1_on_support": 0.296875,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.07432469725608826,
          "probe_belief_target_nll": 4.467698097229004,
          "probe_belief_top1_on_support": 0.08984375,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.06852079182863235,
          "decoder_belief_target_nll": 4.5663299560546875,
          "decoder_belief_top1_on_support": 0.1015625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.15624620020389557,
          "query_target_nll": 15.42966365814209,
          "query_top1_on_support": 0.15625,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.005121015943586826,
          "probe_belief_target_nll": 19.43918228149414,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.005358696915209293,
          "decoder_belief_target_nll": 18.6466064453125,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.14589419960975647,
          "query_target_nll": 12.966175079345703,
          "query_top1_on_support": 0.16796875,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.006978422403335571,
          "probe_belief_target_nll": 20.616241455078125,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.007258635014295578,
          "decoder_belief_target_nll": 16.661775588989258,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.12808607518672943,
          "query_target_nll": 10.493486404418945,
          "query_top1_on_support": 0.11328125,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.004614963661879301,
          "probe_belief_target_nll": 18.70041275024414,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.005015830975025892,
          "decoder_belief_target_nll": 15.311408042907715,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.13022330403327942,
          "query_target_nll": 7.988440990447998,
          "query_top1_on_support": 0.12890625,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.004940406419336796,
          "probe_belief_target_nll": 16.149913787841797,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.005648567341268063,
          "decoder_belief_target_nll": 13.694581031799316,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.14233975112438202,
          "query_target_nll": 6.3473920822143555,
          "query_top1_on_support": 0.12109375,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.013718162663280964,
          "probe_belief_target_nll": 12.738033294677734,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.011703357100486755,
          "decoder_belief_target_nll": 11.671591758728027,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.3070523738861084,
          "query_target_nll": 2.7574517726898193,
          "query_top1_on_support": 0.3515625,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.07369403541088104,
          "probe_belief_target_nll": 4.563894748687744,
          "probe_belief_top1_on_support": 0.1015625,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.06677305698394775,
          "decoder_belief_target_nll": 4.636654853820801,
          "decoder_belief_top1_on_support": 0.08203125,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.3070523738861084,
          "query_target_nll": 2.7574517726898193,
          "query_top1_on_support": 0.3515625,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.07369403541088104,
          "probe_belief_target_nll": 4.563894748687744,
          "probe_belief_top1_on_support": 0.1015625,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.06677305698394775,
          "decoder_belief_target_nll": 4.636654853820801,
          "decoder_belief_top1_on_support": 0.08203125,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.3070523738861084,
          "query_target_nll": 2.7574517726898193,
          "query_top1_on_support": 0.3515625,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.07369403541088104,
          "probe_belief_target_nll": 4.563894748687744,
          "probe_belief_top1_on_support": 0.1015625,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.06677305698394775,
          "decoder_belief_target_nll": 4.636654853820801,
          "decoder_belief_top1_on_support": 0.08203125,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.07402318716049194,
          "query_target_nll": 3.4452571868896484,
          "query_top1_on_support": 0.04296875,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.003032865934073925,
          "probe_belief_target_nll": 20.236000061035156,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.003243163228034973,
          "decoder_belief_target_nll": 19.429166793823242,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.07755617797374725,
          "query_target_nll": 6.46082878112793,
          "query_top1_on_support": 0.078125,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.0036960281431674957,
          "probe_belief_target_nll": 22.266399383544922,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.003946537151932716,
          "decoder_belief_target_nll": 18.249771118164062,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.07524259388446808,
          "query_target_nll": 6.958983898162842,
          "query_top1_on_support": 0.08984375,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.0025011252146214247,
          "probe_belief_target_nll": 18.800182342529297,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.002920888364315033,
          "decoder_belief_target_nll": 15.415680885314941,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.0797172337770462,
          "query_target_nll": 7.1692304611206055,
          "query_top1_on_support": 0.09375,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.0019436674192547798,
          "probe_belief_target_nll": 16.0010986328125,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.0020900974050164223,
          "decoder_belief_target_nll": 13.799245834350586,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.09081721305847168,
          "query_target_nll": 6.8494110107421875,
          "query_top1_on_support": 0.08203125,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.0033264183439314365,
          "probe_belief_target_nll": 14.107030868530273,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.003413079772144556,
          "decoder_belief_target_nll": 12.775857925415039,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.11301425844430923,
          "query_target_nll": 6.442441940307617,
          "query_top1_on_support": 0.1171875,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.006531961727887392,
          "probe_belief_target_nll": 12.909035682678223,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.005276944488286972,
          "decoder_belief_target_nll": 11.876649856567383,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.2497502565383911,
          "query_target_nll": 2.390657663345337,
          "query_top1_on_support": 0.30078125,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.047619521617889404,
          "probe_belief_target_nll": 4.517749786376953,
          "probe_belief_top1_on_support": 0.08984375,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.043144166469573975,
          "decoder_belief_target_nll": 4.644145488739014,
          "decoder_belief_top1_on_support": 0.046875,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.2497502565383911,
          "query_target_nll": 2.390657663345337,
          "query_top1_on_support": 0.30078125,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.047619521617889404,
          "probe_belief_target_nll": 4.517749786376953,
          "probe_belief_top1_on_support": 0.08984375,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.043144166469573975,
          "decoder_belief_target_nll": 4.644145488739014,
          "decoder_belief_top1_on_support": 0.046875,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.07855866849422455,
          "query_target_nll": 3.448561906814575,
          "query_top1_on_support": 0.07421875,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.003454031888395548,
          "probe_belief_target_nll": 20.193527221679688,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0033936845138669014,
          "decoder_belief_target_nll": 19.265953063964844,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.07931599020957947,
          "query_target_nll": 6.5073394775390625,
          "query_top1_on_support": 0.0859375,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.0030961609445512295,
          "probe_belief_target_nll": 21.41629409790039,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0031768956687301397,
          "decoder_belief_target_nll": 17.41656494140625,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.07894865423440933,
          "query_target_nll": 7.00504207611084,
          "query_top1_on_support": 0.07421875,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.0036955266259610653,
          "probe_belief_target_nll": 18.873943328857422,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.004003722220659256,
          "decoder_belief_target_nll": 15.396686553955078,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.07844914495944977,
          "query_target_nll": 7.282888412475586,
          "query_top1_on_support": 0.08203125,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.0023273173719644547,
          "probe_belief_target_nll": 15.869953155517578,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0023768111132085323,
          "decoder_belief_target_nll": 13.527278900146484,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.08736065030097961,
          "query_target_nll": 7.727080345153809,
          "query_top1_on_support": 0.08984375,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.0030708236154168844,
          "probe_belief_target_nll": 14.608489990234375,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.002903833519667387,
          "decoder_belief_target_nll": 13.182997703552246,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.11832715570926666,
          "query_target_nll": 6.988277435302734,
          "query_top1_on_support": 0.1640625,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.005537400953471661,
          "probe_belief_target_nll": 13.51080322265625,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.005420057103037834,
          "decoder_belief_target_nll": 12.092500686645508,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.2714126706123352,
          "query_target_nll": 2.37730073928833,
          "query_top1_on_support": 0.36328125,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.04833938926458359,
          "probe_belief_target_nll": 4.650869369506836,
          "probe_belief_top1_on_support": 0.06640625,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.047781672328710556,
          "decoder_belief_target_nll": 4.677072525024414,
          "decoder_belief_top1_on_support": 0.1015625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.2714126706123352,
          "query_target_nll": 2.37730073928833,
          "query_top1_on_support": 0.36328125,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.04833938926458359,
          "probe_belief_target_nll": 4.650869369506836,
          "probe_belief_top1_on_support": 0.06640625,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.047781672328710556,
          "decoder_belief_target_nll": 4.677072525024414,
          "decoder_belief_top1_on_support": 0.1015625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.0752170979976654,
          "query_target_nll": 3.4349684715270996,
          "query_top1_on_support": 0.09375,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.0022239170502871275,
          "probe_belief_target_nll": 20.54547882080078,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.002165406011044979,
          "decoder_belief_target_nll": 19.535465240478516,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.06921474635601044,
          "query_target_nll": 4.5010600090026855,
          "query_top1_on_support": 0.0625,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.0035087044816464186,
          "probe_belief_target_nll": 21.62310791015625,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.0036528243217617273,
          "decoder_belief_target_nll": 17.76543426513672,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.06596168875694275,
          "query_target_nll": 4.491710186004639,
          "query_top1_on_support": 0.07421875,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.0026964442804455757,
          "probe_belief_target_nll": 19.55191421508789,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.0029695869889110327,
          "decoder_belief_target_nll": 16.07668685913086,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.07674743980169296,
          "query_target_nll": 4.5481414794921875,
          "query_top1_on_support": 0.07421875,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.0024252182338386774,
          "probe_belief_target_nll": 16.752029418945312,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.002456870162859559,
          "decoder_belief_target_nll": 14.333343505859375,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.08135445415973663,
          "query_target_nll": 4.840205192565918,
          "query_top1_on_support": 0.08984375,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.003845163621008396,
          "probe_belief_target_nll": 14.107419967651367,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.0035599556285887957,
          "decoder_belief_target_nll": 12.924406051635742,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.07777456939220428,
          "query_target_nll": 5.132898330688477,
          "query_top1_on_support": 0.109375,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.006000799126923084,
          "probe_belief_target_nll": 13.248591423034668,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.005011127330362797,
          "decoder_belief_target_nll": 12.148578643798828,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.17345690727233887,
          "query_target_nll": 2.7504591941833496,
          "query_top1_on_support": 0.23046875,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.05253813415765762,
          "probe_belief_target_nll": 4.537432670593262,
          "probe_belief_top1_on_support": 0.109375,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.042853519320487976,
          "decoder_belief_target_nll": 4.629834175109863,
          "decoder_belief_top1_on_support": 0.0546875,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.17345690727233887,
          "query_target_nll": 2.7504591941833496,
          "query_top1_on_support": 0.23046875,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.05253813415765762,
          "probe_belief_target_nll": 4.537432670593262,
          "probe_belief_top1_on_support": 0.109375,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.042853519320487976,
          "decoder_belief_target_nll": 4.629834175109863,
          "decoder_belief_top1_on_support": 0.0546875,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.09764684736728668,
          "query_target_nll": 15.740348815917969,
          "query_top1_on_support": 0.09765625,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.00311747333034873,
          "probe_belief_target_nll": 19.832834243774414,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.002986658364534378,
          "decoder_belief_target_nll": 19.237194061279297,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.07808974385261536,
          "query_target_nll": 14.48492431640625,
          "query_top1_on_support": 0.078125,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.0024493085220456123,
          "probe_belief_target_nll": 22.955602645874023,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.0029524252749979496,
          "decoder_belief_target_nll": 18.391000747680664,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.09379839897155762,
          "query_target_nll": 10.637611389160156,
          "query_top1_on_support": 0.10546875,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.002584426896646619,
          "probe_belief_target_nll": 20.141164779663086,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.0035496961791068316,
          "decoder_belief_target_nll": 16.11939239501953,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.06995917111635208,
          "query_target_nll": 8.543733596801758,
          "query_top1_on_support": 0.046875,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.002806904725730419,
          "probe_belief_target_nll": 17.235408782958984,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.002616694662719965,
          "decoder_belief_target_nll": 14.459754943847656,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.07623346149921417,
          "query_target_nll": 7.412636756896973,
          "query_top1_on_support": 0.0546875,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.0025096540339291096,
          "probe_belief_target_nll": 14.58630657196045,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.002323069144040346,
          "decoder_belief_target_nll": 13.003660202026367,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.07942146062850952,
          "query_target_nll": 6.793580055236816,
          "query_top1_on_support": 0.078125,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.0058068884536623955,
          "probe_belief_target_nll": 12.867358207702637,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.005408570636063814,
          "decoder_belief_target_nll": 11.87070369720459,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.21822145581245422,
          "query_target_nll": 2.8131871223449707,
          "query_top1_on_support": 0.2734375,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.044539645314216614,
          "probe_belief_target_nll": 4.750622749328613,
          "probe_belief_top1_on_support": 0.0546875,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.04355961084365845,
          "decoder_belief_target_nll": 4.773484230041504,
          "decoder_belief_top1_on_support": 0.05078125,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.21822145581245422,
          "query_target_nll": 2.8131871223449707,
          "query_top1_on_support": 0.2734375,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.044539645314216614,
          "probe_belief_target_nll": 4.750622749328613,
          "probe_belief_top1_on_support": 0.0546875,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.04355961084365845,
          "decoder_belief_target_nll": 4.773484230041504,
          "decoder_belief_top1_on_support": 0.05078125,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.05002513527870178,
          "query_target_nll": 3.450819492340088,
          "query_top1_on_support": 0.04296875,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.0014453905168920755,
          "probe_belief_target_nll": 20.260562896728516,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.001650748890824616,
          "decoder_belief_target_nll": 18.878887176513672,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.0497707799077034,
          "query_target_nll": 7.5727128982543945,
          "query_top1_on_support": 0.04296875,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.001368933473713696,
          "probe_belief_target_nll": 22.575393676757812,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0016681901179254055,
          "decoder_belief_target_nll": 17.98963165283203,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.04925272986292839,
          "query_target_nll": 7.240996360778809,
          "query_top1_on_support": 0.0546875,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.0013651440385729074,
          "probe_belief_target_nll": 19.766162872314453,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0018616081215441227,
          "decoder_belief_target_nll": 15.837435722351074,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.05743411183357239,
          "query_target_nll": 7.180184841156006,
          "query_top1_on_support": 0.05859375,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.002713646274060011,
          "probe_belief_target_nll": 15.905848503112793,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.003298343624919653,
          "decoder_belief_target_nll": 13.643674850463867,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.0538628026843071,
          "query_target_nll": 7.667764663696289,
          "query_top1_on_support": 0.0703125,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.0018114014528691769,
          "probe_belief_target_nll": 14.5148286819458,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.001488272799178958,
          "decoder_belief_target_nll": 13.174311637878418,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.044274479150772095,
          "query_target_nll": 7.829044818878174,
          "query_top1_on_support": 0.05078125,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.001982542220503092,
          "probe_belief_target_nll": 14.23430347442627,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0025835842825472355,
          "decoder_belief_target_nll": 13.164031028747559,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.04685056954622269,
          "query_target_nll": 7.822125434875488,
          "query_top1_on_support": 0.04296875,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.0027436143718659878,
          "probe_belief_target_nll": 14.349641799926758,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0023177736438810825,
          "decoder_belief_target_nll": 13.119730949401855,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.1819154918193817,
          "query_target_nll": 2.4233901500701904,
          "query_top1_on_support": 0.21875,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.03182527422904968,
          "probe_belief_target_nll": 4.596618175506592,
          "probe_belief_top1_on_support": 0.04296875,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.02976546436548233,
          "decoder_belief_target_nll": 4.72791862487793,
          "decoder_belief_top1_on_support": 0.0390625,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.050779879093170166,
          "query_target_nll": 3.451326847076416,
          "query_top1_on_support": 0.0546875,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0024720351211726665,
          "probe_belief_target_nll": 19.60150146484375,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0023382389917969704,
          "decoder_belief_target_nll": 18.677921295166016,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.04878048226237297,
          "query_target_nll": 5.844198226928711,
          "query_top1_on_support": 0.046875,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0024446258321404457,
          "probe_belief_target_nll": 21.429794311523438,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.002308167517185211,
          "decoder_belief_target_nll": 17.16750144958496,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.04928814619779587,
          "query_target_nll": 6.281500816345215,
          "query_top1_on_support": 0.046875,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.003886797931045294,
          "probe_belief_target_nll": 18.366352081298828,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0043664369732141495,
          "decoder_belief_target_nll": 14.80514907836914,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.04513581842184067,
          "query_target_nll": 6.96187686920166,
          "query_top1_on_support": 0.04296875,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.002204642165452242,
          "probe_belief_target_nll": 15.89588451385498,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0017680180026218295,
          "decoder_belief_target_nll": 13.535005569458008,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.05859971046447754,
          "query_target_nll": 7.33762788772583,
          "query_top1_on_support": 0.078125,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0021698209457099438,
          "probe_belief_target_nll": 14.869701385498047,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0013755192048847675,
          "decoder_belief_target_nll": 13.166829109191895,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.051497284322977066,
          "query_target_nll": 7.519469261169434,
          "query_top1_on_support": 0.04296875,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0014119461411610246,
          "probe_belief_target_nll": 14.198015213012695,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.001350542064756155,
          "decoder_belief_target_nll": 13.001680374145508,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.05403326824307442,
          "query_target_nll": 7.820031642913818,
          "query_top1_on_support": 0.046875,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0022325210738927126,
          "probe_belief_target_nll": 14.210277557373047,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0017770074773579836,
          "decoder_belief_target_nll": 13.00777530670166,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.1810535043478012,
          "query_target_nll": 2.4974493980407715,
          "query_top1_on_support": 0.22265625,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.02998519502580166,
          "probe_belief_target_nll": 4.928955078125,
          "probe_belief_top1_on_support": 0.03515625,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.02715037390589714,
          "decoder_belief_target_nll": 4.941681861877441,
          "decoder_belief_top1_on_support": 0.0390625,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.04844088852405548,
          "query_target_nll": 3.45188570022583,
          "query_top1_on_support": 0.05078125,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0014527615858241916,
          "probe_belief_target_nll": 20.579395294189453,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0012658210471272469,
          "decoder_belief_target_nll": 19.700345993041992,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.05137209966778755,
          "query_target_nll": 4.149942874908447,
          "query_top1_on_support": 0.06640625,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0006099788006395102,
          "probe_belief_target_nll": 23.00002670288086,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0006875827093608677,
          "decoder_belief_target_nll": 18.41862678527832,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.045517005026340485,
          "query_target_nll": 4.309512615203857,
          "query_top1_on_support": 0.05078125,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0006448991480283439,
          "probe_belief_target_nll": 20.554697036743164,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0006266424898058176,
          "decoder_belief_target_nll": 16.562572479248047,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.05200439319014549,
          "query_target_nll": 4.819108486175537,
          "query_top1_on_support": 0.078125,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0019996543414890766,
          "probe_belief_target_nll": 16.892929077148438,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0018461078871041536,
          "decoder_belief_target_nll": 14.333735466003418,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.04960928484797478,
          "query_target_nll": 4.935050964355469,
          "query_top1_on_support": 0.04296875,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0018606201047077775,
          "probe_belief_target_nll": 14.779552459716797,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.001799222081899643,
          "decoder_belief_target_nll": 13.441563606262207,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.05017296224832535,
          "query_target_nll": 4.789178848266602,
          "query_top1_on_support": 0.0546875,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0014068552991375327,
          "probe_belief_target_nll": 14.349069595336914,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0013330877991393209,
          "decoder_belief_target_nll": 13.091439247131348,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.05108410865068436,
          "query_target_nll": 4.914534568786621,
          "query_top1_on_support": 0.06640625,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.00190033542457968,
          "probe_belief_target_nll": 14.042010307312012,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0020595756359398365,
          "decoder_belief_target_nll": 12.983100891113281,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.1265985369682312,
          "query_target_nll": 2.7512974739074707,
          "query_top1_on_support": 0.171875,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.03159945458173752,
          "probe_belief_target_nll": 4.854114532470703,
          "probe_belief_top1_on_support": 0.05078125,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.030767232179641724,
          "decoder_belief_target_nll": 4.842070579528809,
          "decoder_belief_top1_on_support": 0.0546875,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.058593183755874634,
          "query_target_nll": 15.36107063293457,
          "query_top1_on_support": 0.05859375,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0017285235226154327,
          "probe_belief_target_nll": 19.49241828918457,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0018246042309328914,
          "decoder_belief_target_nll": 18.785240173339844,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.055491410195827484,
          "query_target_nll": 13.689867973327637,
          "query_top1_on_support": 0.05078125,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.002825981704518199,
          "probe_belief_target_nll": 22.079261779785156,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.002895868383347988,
          "decoder_belief_target_nll": 17.608562469482422,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.039778247475624084,
          "query_target_nll": 12.007421493530273,
          "query_top1_on_support": 0.01953125,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.00046357838436961174,
          "probe_belief_target_nll": 20.492263793945312,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0005122085567563772,
          "decoder_belief_target_nll": 16.473514556884766,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.04901470988988876,
          "query_target_nll": 8.137460708618164,
          "query_top1_on_support": 0.03515625,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0027336550410836935,
          "probe_belief_target_nll": 16.375436782836914,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0026952989865094423,
          "decoder_belief_target_nll": 13.534523010253906,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.053211405873298645,
          "query_target_nll": 7.1107563972473145,
          "query_top1_on_support": 0.03515625,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0021127897780388594,
          "probe_belief_target_nll": 14.36923599243164,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.001955299871042371,
          "decoder_belief_target_nll": 12.803702354431152,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.06147283688187599,
          "query_target_nll": 6.81730318069458,
          "query_top1_on_support": 0.0703125,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0011884011328220367,
          "probe_belief_target_nll": 13.979998588562012,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0011697225272655487,
          "decoder_belief_target_nll": 12.960027694702148,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.05081668496131897,
          "query_target_nll": 6.405533313751221,
          "query_top1_on_support": 0.05859375,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0020119291730225086,
          "probe_belief_target_nll": 13.969758987426758,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.002159461611881852,
          "decoder_belief_target_nll": 12.739486694335938,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sparse_belief",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.16033422946929932,
          "query_target_nll": 2.8388171195983887,
          "query_top1_on_support": 0.21484375,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.03438716381788254,
          "probe_belief_target_nll": 4.7839531898498535,
          "probe_belief_top1_on_support": 0.0546875,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0326245054602623,
          "decoder_belief_target_nll": 4.821120262145996,
          "decoder_belief_top1_on_support": 0.0703125,
          "mean_decoder_belief_support_size": 1.55078125
        }
      ]
    }
  ],
  "checkpoints": [
    "large_artifacts/dense_supervision_ladder/checkpoints/main_sparse_belief_mod31/checkpoint_final.pt"
  ]
}