{
  "mode": "dense",
  "args": {
    "mode": "dense",
    "supervision": "sampled_final",
    "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_sampled_final_mod31",
    "checkpoint_dir": "large_artifacts/dense_supervision_ladder/checkpoints/main_sampled_final_mod31"
  },
  "env": {
    "python": "3.12.3",
    "torch": "2.8.0+cu128",
    "device": "cuda",
    "cuda_name": "NVIDIA RTX 6000 Ada Generation"
  },
  "train": [
    {
      "step": 1,
      "loss": 3.458657741546631,
      "grad_norm": 0.7787356376647949,
      "query_loss": 3.458657741546631,
      "belief_loss": 0.0,
      "query_target_mass": 0.355380117893219,
      "query_target_nll": 3.31845760345459,
      "query_top1_on_support": 0.65625,
      "mean_query_support_size": 10.6796875
    },
    {
      "step": 250,
      "loss": 3.256669282913208,
      "grad_norm": 0.608508825302124,
      "query_loss": 3.256669282913208,
      "belief_loss": 0.0,
      "query_target_mass": 0.4335976541042328,
      "query_target_nll": 3.2420663833618164,
      "query_top1_on_support": 0.4921875,
      "mean_query_support_size": 11.55859375
    },
    {
      "step": 500,
      "loss": 3.197326183319092,
      "grad_norm": 0.6760830879211426,
      "query_loss": 3.197326183319092,
      "belief_loss": 0.0,
      "query_target_mass": 0.43896034359931946,
      "query_target_nll": 3.1738734245300293,
      "query_top1_on_support": 0.4921875,
      "mean_query_support_size": 11.1875
    },
    {
      "step": 750,
      "loss": 3.2419610023498535,
      "grad_norm": 0.6131318211555481,
      "query_loss": 3.2419610023498535,
      "belief_loss": 0.0,
      "query_target_mass": 0.44917112588882446,
      "query_target_nll": 3.211925983428955,
      "query_top1_on_support": 0.46875,
      "mean_query_support_size": 11.765625
    },
    {
      "step": 1000,
      "loss": 3.054413318634033,
      "grad_norm": 0.6427000761032104,
      "query_loss": 3.054413318634033,
      "belief_loss": 0.0,
      "query_target_mass": 0.4639519453048706,
      "query_target_nll": 3.0754473209381104,
      "query_top1_on_support": 0.54296875,
      "mean_query_support_size": 11.140625
    },
    {
      "step": 1250,
      "loss": 3.0142078399658203,
      "grad_norm": 0.6964521408081055,
      "query_loss": 3.0142078399658203,
      "belief_loss": 0.0,
      "query_target_mass": 0.4530593156814575,
      "query_target_nll": 3.034543037414551,
      "query_top1_on_support": 0.5234375,
      "mean_query_support_size": 10.265625
    },
    {
      "step": 1500,
      "loss": 2.997404098510742,
      "grad_norm": 0.7726538777351379,
      "query_loss": 2.997404098510742,
      "belief_loss": 0.0,
      "query_target_mass": 0.49872449040412903,
      "query_target_nll": 2.9595303535461426,
      "query_top1_on_support": 0.56640625,
      "mean_query_support_size": 11.32421875
    },
    {
      "step": 1750,
      "loss": 2.9219069480895996,
      "grad_norm": 0.744170606136322,
      "query_loss": 2.9219069480895996,
      "belief_loss": 0.0,
      "query_target_mass": 0.5330148339271545,
      "query_target_nll": 2.9054243564605713,
      "query_top1_on_support": 0.5859375,
      "mean_query_support_size": 11.48828125
    },
    {
      "step": 2000,
      "loss": 2.9793601036071777,
      "grad_norm": 0.7833150029182434,
      "query_loss": 2.9793601036071777,
      "belief_loss": 0.0,
      "query_target_mass": 0.5056822299957275,
      "query_target_nll": 2.9254961013793945,
      "query_top1_on_support": 0.56640625,
      "mean_query_support_size": 11.22265625
    }
  ],
  "probe_train": [
    {
      "step": 1,
      "loss": 6.887564659118652,
      "grad_norm": 0.19753976166248322,
      "probe_belief_target_mass": 0.01982787437736988,
      "probe_belief_target_nll": 6.887564659118652,
      "probe_belief_top1_on_support": 0.01171875,
      "mean_belief_support_size": 19.06727409362793
    },
    {
      "step": 200,
      "loss": 4.908459663391113,
      "grad_norm": 0.35194405913352966,
      "probe_belief_target_mass": 0.31116625666618347,
      "probe_belief_target_nll": 4.908459663391113,
      "probe_belief_top1_on_support": 0.3927951455116272,
      "mean_belief_support_size": 17.91232681274414
    },
    {
      "step": 400,
      "loss": 4.76699686050415,
      "grad_norm": 0.33680516481399536,
      "probe_belief_target_mass": 0.33993053436279297,
      "probe_belief_target_nll": 4.76699686050415,
      "probe_belief_top1_on_support": 0.3858506977558136,
      "mean_belief_support_size": 17.484375
    },
    {
      "step": 600,
      "loss": 4.704586029052734,
      "grad_norm": 0.31722620129585266,
      "probe_belief_target_mass": 0.36774519085884094,
      "probe_belief_target_nll": 4.704586029052734,
      "probe_belief_top1_on_support": 0.4114583432674408,
      "mean_belief_support_size": 17.92447853088379
    },
    {
      "step": 800,
      "loss": 4.646040916442871,
      "grad_norm": 0.3301858901977539,
      "probe_belief_target_mass": 0.3828005790710449,
      "probe_belief_target_nll": 4.646040916442871,
      "probe_belief_top1_on_support": 0.4283854067325592,
      "mean_belief_support_size": 18.07725715637207
    }
  ],
  "eval": [
    {
      "step": 2000,
      "rows": [
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.3885788321495056,
          "query_target_nll": 3.8880763053894043,
          "query_top1_on_support": 0.41015625,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.02420428767800331,
          "probe_belief_target_nll": 18.1962890625,
          "probe_belief_top1_on_support": 0.0234375,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.01447401661425829,
          "decoder_belief_target_nll": 6.882529258728027,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.42850545048713684,
          "query_target_nll": 4.217644691467285,
          "query_top1_on_support": 0.4375,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.043166570365428925,
          "probe_belief_target_nll": 11.479613304138184,
          "probe_belief_top1_on_support": 0.05859375,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014443970285356045,
          "decoder_belief_target_nll": 6.890974998474121,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.4727054834365845,
          "query_target_nll": 3.7510786056518555,
          "query_top1_on_support": 0.50390625,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.06103072315454483,
          "probe_belief_target_nll": 8.463632583618164,
          "probe_belief_top1_on_support": 0.078125,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014424953609704971,
          "decoder_belief_target_nll": 6.889986038208008,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.5676214694976807,
          "query_target_nll": 2.9163055419921875,
          "query_top1_on_support": 0.625,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.10574312508106232,
          "probe_belief_target_nll": 5.3946123123168945,
          "probe_belief_top1_on_support": 0.15625,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014377398416399956,
          "decoder_belief_target_nll": 6.891692161560059,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.5676214694976807,
          "query_target_nll": 2.9163055419921875,
          "query_top1_on_support": 0.625,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.10574312508106232,
          "probe_belief_target_nll": 5.3946123123168945,
          "probe_belief_top1_on_support": 0.15625,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014377398416399956,
          "decoder_belief_target_nll": 6.891692161560059,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.5676214694976807,
          "query_target_nll": 2.9163055419921875,
          "query_top1_on_support": 0.625,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.10574312508106232,
          "probe_belief_target_nll": 5.3946123123168945,
          "probe_belief_top1_on_support": 0.15625,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014377398416399956,
          "decoder_belief_target_nll": 6.891692161560059,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.5676214694976807,
          "query_target_nll": 2.9163055419921875,
          "query_top1_on_support": 0.625,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.10574312508106232,
          "probe_belief_target_nll": 5.3946123123168945,
          "probe_belief_top1_on_support": 0.15625,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014377398416399956,
          "decoder_belief_target_nll": 6.891692161560059,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.5676214694976807,
          "query_target_nll": 2.9163055419921875,
          "query_top1_on_support": 0.625,
          "mean_query_support_size": 12.01171875,
          "probe_belief_target_mass": 0.10574312508106232,
          "probe_belief_target_nll": 5.3946123123168945,
          "probe_belief_top1_on_support": 0.15625,
          "mean_belief_support_size": 13.84765625,
          "decoder_belief_target_mass": 0.014377398416399956,
          "decoder_belief_target_nll": 6.891692161560059,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 13.84765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.3628043234348297,
          "query_target_nll": 3.899934768676758,
          "query_top1_on_support": 0.41796875,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.02521379664540291,
          "probe_belief_target_nll": 17.832368850708008,
          "probe_belief_top1_on_support": 0.03125,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.01388023141771555,
          "decoder_belief_target_nll": 6.888823509216309,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.41219937801361084,
          "query_target_nll": 4.214325428009033,
          "query_top1_on_support": 0.41796875,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.07149087637662888,
          "probe_belief_target_nll": 11.212239265441895,
          "probe_belief_top1_on_support": 0.06640625,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013923258520662785,
          "decoder_belief_target_nll": 6.886358261108398,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.46003228425979614,
          "query_target_nll": 3.7781646251678467,
          "query_top1_on_support": 0.4765625,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.0803302675485611,
          "probe_belief_target_nll": 8.849077224731445,
          "probe_belief_top1_on_support": 0.09375,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.01389826275408268,
          "decoder_belief_target_nll": 6.884922027587891,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.5803507566452026,
          "query_target_nll": 2.757957935333252,
          "query_top1_on_support": 0.640625,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.14212942123413086,
          "probe_belief_target_nll": 5.09412956237793,
          "probe_belief_top1_on_support": 0.1953125,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013922741636633873,
          "decoder_belief_target_nll": 6.884544372558594,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.5803507566452026,
          "query_target_nll": 2.757957935333252,
          "query_top1_on_support": 0.640625,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.14212942123413086,
          "probe_belief_target_nll": 5.09412956237793,
          "probe_belief_top1_on_support": 0.1953125,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013922741636633873,
          "decoder_belief_target_nll": 6.884544372558594,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.5803507566452026,
          "query_target_nll": 2.757957935333252,
          "query_top1_on_support": 0.640625,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.14212942123413086,
          "probe_belief_target_nll": 5.09412956237793,
          "probe_belief_top1_on_support": 0.1953125,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013922741636633873,
          "decoder_belief_target_nll": 6.884544372558594,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.5803507566452026,
          "query_target_nll": 2.757957935333252,
          "query_top1_on_support": 0.640625,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.14212942123413086,
          "probe_belief_target_nll": 5.09412956237793,
          "probe_belief_top1_on_support": 0.1953125,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013922741636633873,
          "decoder_belief_target_nll": 6.884544372558594,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.5803507566452026,
          "query_target_nll": 2.757957935333252,
          "query_top1_on_support": 0.640625,
          "mean_query_support_size": 11.109375,
          "probe_belief_target_mass": 0.14212942123413086,
          "probe_belief_target_nll": 5.09412956237793,
          "probe_belief_top1_on_support": 0.1953125,
          "mean_belief_support_size": 13.328125,
          "decoder_belief_target_mass": 0.013922741636633873,
          "decoder_belief_target_nll": 6.884544372558594,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 13.328125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.3693024218082428,
          "query_target_nll": 3.7612686157226562,
          "query_top1_on_support": 0.3359375,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.03149298205971718,
          "probe_belief_target_nll": 16.827865600585938,
          "probe_belief_top1_on_support": 0.03515625,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.01264966931194067,
          "decoder_belief_target_nll": 6.881804466247559,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.3720166087150574,
          "query_target_nll": 3.762467861175537,
          "query_top1_on_support": 0.34765625,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.07729542255401611,
          "probe_belief_target_nll": 11.266532897949219,
          "probe_belief_top1_on_support": 0.078125,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.012594973668456078,
          "decoder_belief_target_nll": 6.884030342102051,
          "decoder_belief_top1_on_support": 0.01953125,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.38513365387916565,
          "query_target_nll": 3.656439781188965,
          "query_top1_on_support": 0.41015625,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.09306921809911728,
          "probe_belief_target_nll": 8.843932151794434,
          "probe_belief_top1_on_support": 0.1015625,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.012600000947713852,
          "decoder_belief_target_nll": 6.883859634399414,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.4510878026485443,
          "query_target_nll": 3.190380573272705,
          "query_top1_on_support": 0.5,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.13995683193206787,
          "probe_belief_target_nll": 5.1092119216918945,
          "probe_belief_top1_on_support": 0.1953125,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.012549594044685364,
          "decoder_belief_target_nll": 6.88762092590332,
          "decoder_belief_top1_on_support": 0.0234375,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.4510878026485443,
          "query_target_nll": 3.190380573272705,
          "query_top1_on_support": 0.5,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.13995683193206787,
          "probe_belief_target_nll": 5.1092119216918945,
          "probe_belief_top1_on_support": 0.1953125,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.012549594044685364,
          "decoder_belief_target_nll": 6.88762092590332,
          "decoder_belief_top1_on_support": 0.0234375,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.4510878026485443,
          "query_target_nll": 3.190380573272705,
          "query_top1_on_support": 0.5,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.13995683193206787,
          "probe_belief_target_nll": 5.1092119216918945,
          "probe_belief_top1_on_support": 0.1953125,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.012549594044685364,
          "decoder_belief_target_nll": 6.88762092590332,
          "decoder_belief_top1_on_support": 0.0234375,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.4510878026485443,
          "query_target_nll": 3.190380573272705,
          "query_top1_on_support": 0.5,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.13995683193206787,
          "probe_belief_target_nll": 5.1092119216918945,
          "probe_belief_top1_on_support": 0.1953125,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.012549594044685364,
          "decoder_belief_target_nll": 6.88762092590332,
          "decoder_belief_top1_on_support": 0.0234375,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.4510878026485443,
          "query_target_nll": 3.190380573272705,
          "query_top1_on_support": 0.5,
          "mean_query_support_size": 11.453125,
          "probe_belief_target_mass": 0.13995683193206787,
          "probe_belief_target_nll": 5.1092119216918945,
          "probe_belief_top1_on_support": 0.1953125,
          "mean_belief_support_size": 12.0859375,
          "decoder_belief_target_mass": 0.012549594044685364,
          "decoder_belief_target_nll": 6.88762092590332,
          "decoder_belief_top1_on_support": 0.0234375,
          "mean_decoder_belief_support_size": 12.0859375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.4296427071094513,
          "query_target_nll": 11.551993370056152,
          "query_top1_on_support": 0.4296875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.03516547754406929,
          "probe_belief_target_nll": 17.526256561279297,
          "probe_belief_top1_on_support": 0.0390625,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.015009617432951927,
          "decoder_belief_target_nll": 6.885549545288086,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.3949224352836609,
          "query_target_nll": 4.592586994171143,
          "query_top1_on_support": 0.421875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.06300163269042969,
          "probe_belief_target_nll": 11.438828468322754,
          "probe_belief_top1_on_support": 0.0625,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.014922873117029667,
          "decoder_belief_target_nll": 6.895334243774414,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.3960111737251282,
          "query_target_nll": 3.890069007873535,
          "query_top1_on_support": 0.40234375,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.08244329690933228,
          "probe_belief_target_nll": 8.649209022521973,
          "probe_belief_top1_on_support": 0.0859375,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.01500011794269085,
          "decoder_belief_target_nll": 6.885737419128418,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.44220778346061707,
          "query_target_nll": 3.176767349243164,
          "query_top1_on_support": 0.46875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.12629294395446777,
          "probe_belief_target_nll": 5.354042053222656,
          "probe_belief_top1_on_support": 0.171875,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.014984514564275742,
          "decoder_belief_target_nll": 6.887640953063965,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.44220778346061707,
          "query_target_nll": 3.176767349243164,
          "query_top1_on_support": 0.46875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.12629294395446777,
          "probe_belief_target_nll": 5.354042053222656,
          "probe_belief_top1_on_support": 0.171875,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.014984514564275742,
          "decoder_belief_target_nll": 6.887640953063965,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.44220778346061707,
          "query_target_nll": 3.176767349243164,
          "query_top1_on_support": 0.46875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.12629294395446777,
          "probe_belief_target_nll": 5.354042053222656,
          "probe_belief_top1_on_support": 0.171875,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.014984514564275742,
          "decoder_belief_target_nll": 6.887640953063965,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.44220778346061707,
          "query_target_nll": 3.176767349243164,
          "query_top1_on_support": 0.46875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.12629294395446777,
          "probe_belief_target_nll": 5.354042053222656,
          "probe_belief_top1_on_support": 0.171875,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.014984514564275742,
          "decoder_belief_target_nll": 6.887640953063965,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 4,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.44220778346061707,
          "query_target_nll": 3.176767349243164,
          "query_top1_on_support": 0.46875,
          "mean_query_support_size": 11.60546875,
          "probe_belief_target_mass": 0.12629294395446777,
          "probe_belief_target_nll": 5.354042053222656,
          "probe_belief_top1_on_support": 0.171875,
          "mean_belief_support_size": 14.40625,
          "decoder_belief_target_mass": 0.014984514564275742,
          "decoder_belief_target_nll": 6.887640953063965,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 14.40625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.17243902385234833,
          "query_target_nll": 3.971475601196289,
          "query_top1_on_support": 0.18359375,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.008015068247914314,
          "probe_belief_target_nll": 18.550508499145508,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.007428823970258236,
          "decoder_belief_target_nll": 6.892124652862549,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.1965627372264862,
          "query_target_nll": 4.810474395751953,
          "query_top1_on_support": 0.21484375,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.013360894285142422,
          "probe_belief_target_nll": 13.635334014892578,
          "probe_belief_top1_on_support": 0.0234375,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.007437261752784252,
          "decoder_belief_target_nll": 6.891261577606201,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.19899779558181763,
          "query_target_nll": 4.556830406188965,
          "query_top1_on_support": 0.1953125,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.010605107992887497,
          "probe_belief_target_nll": 11.09372329711914,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.007386397570371628,
          "decoder_belief_target_nll": 6.898001670837402,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.1989201009273529,
          "query_target_nll": 4.5083818435668945,
          "query_top1_on_support": 0.21875,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.015459328889846802,
          "probe_belief_target_nll": 9.851175308227539,
          "probe_belief_top1_on_support": 0.02734375,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.007471820339560509,
          "decoder_belief_target_nll": 6.887272357940674,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.30748385190963745,
          "query_target_nll": 2.8283579349517822,
          "query_top1_on_support": 0.3203125,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.043252043426036835,
          "probe_belief_target_nll": 5.345164775848389,
          "probe_belief_top1_on_support": 0.0703125,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.007447837386280298,
          "decoder_belief_target_nll": 6.887303352355957,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.30748385190963745,
          "query_target_nll": 2.8283579349517822,
          "query_top1_on_support": 0.3203125,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.043252043426036835,
          "probe_belief_target_nll": 5.345164775848389,
          "probe_belief_top1_on_support": 0.0703125,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.007447837386280298,
          "decoder_belief_target_nll": 6.887303352355957,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.30748385190963745,
          "query_target_nll": 2.8283579349517822,
          "query_top1_on_support": 0.3203125,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.043252043426036835,
          "probe_belief_target_nll": 5.345164775848389,
          "probe_belief_top1_on_support": 0.0703125,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.007447837386280298,
          "decoder_belief_target_nll": 6.887303352355957,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.30748385190963745,
          "query_target_nll": 2.8283579349517822,
          "query_top1_on_support": 0.3203125,
          "mean_query_support_size": 5.2890625,
          "probe_belief_target_mass": 0.043252043426036835,
          "probe_belief_target_nll": 5.345164775848389,
          "probe_belief_top1_on_support": 0.0703125,
          "mean_belief_support_size": 7.140625,
          "decoder_belief_target_mass": 0.007447837386280298,
          "decoder_belief_target_nll": 6.887303352355957,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.19766899943351746,
          "query_target_nll": 3.889392375946045,
          "query_top1_on_support": 0.20703125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.007634279318153858,
          "probe_belief_target_nll": 18.39246940612793,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.0073534827679395676,
          "decoder_belief_target_nll": 6.897350311279297,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.1982155442237854,
          "query_target_nll": 4.719602108001709,
          "query_top1_on_support": 0.1953125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.009770306758582592,
          "probe_belief_target_nll": 13.240630149841309,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007374165579676628,
          "decoder_belief_target_nll": 6.894089698791504,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.2012273073196411,
          "query_target_nll": 4.5574212074279785,
          "query_top1_on_support": 0.1953125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.013787575997412205,
          "probe_belief_target_nll": 10.99308967590332,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007387434132397175,
          "decoder_belief_target_nll": 6.889036178588867,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.22369906306266785,
          "query_target_nll": 4.171636581420898,
          "query_top1_on_support": 0.25390625,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.014343975111842155,
          "probe_belief_target_nll": 9.67086410522461,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007379623595625162,
          "decoder_belief_target_nll": 6.8886919021606445,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.3324320316314697,
          "query_target_nll": 2.8420939445495605,
          "query_top1_on_support": 0.39453125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.04348842054605484,
          "probe_belief_target_nll": 5.417312145233154,
          "probe_belief_top1_on_support": 0.05859375,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007441340945661068,
          "decoder_belief_target_nll": 6.879589557647705,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.3324320316314697,
          "query_target_nll": 2.8420939445495605,
          "query_top1_on_support": 0.39453125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.04348842054605484,
          "probe_belief_target_nll": 5.417312145233154,
          "probe_belief_top1_on_support": 0.05859375,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007441340945661068,
          "decoder_belief_target_nll": 6.879589557647705,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.3324320316314697,
          "query_target_nll": 2.8420939445495605,
          "query_top1_on_support": 0.39453125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.04348842054605484,
          "probe_belief_target_nll": 5.417312145233154,
          "probe_belief_top1_on_support": 0.05859375,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007441340945661068,
          "decoder_belief_target_nll": 6.879589557647705,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.3324320316314697,
          "query_target_nll": 2.8420939445495605,
          "query_top1_on_support": 0.39453125,
          "mean_query_support_size": 5.84765625,
          "probe_belief_target_mass": 0.04348842054605484,
          "probe_belief_target_nll": 5.417312145233154,
          "probe_belief_top1_on_support": 0.05859375,
          "mean_belief_support_size": 7.09765625,
          "decoder_belief_target_mass": 0.007441340945661068,
          "decoder_belief_target_nll": 6.879589557647705,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 7.09765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.17043952643871307,
          "query_target_nll": 3.7523679733276367,
          "query_top1_on_support": 0.1875,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.009995806030929089,
          "probe_belief_target_nll": 18.590633392333984,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.00601406954228878,
          "decoder_belief_target_nll": 6.883366107940674,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.1749255359172821,
          "query_target_nll": 3.783144950866699,
          "query_top1_on_support": 0.171875,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.0077660176903009415,
          "probe_belief_target_nll": 13.687175750732422,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.006001625210046768,
          "decoder_belief_target_nll": 6.886120319366455,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.17747895419597626,
          "query_target_nll": 3.748115062713623,
          "query_top1_on_support": 0.17578125,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.009571303613483906,
          "probe_belief_target_nll": 11.217192649841309,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.006005727685987949,
          "decoder_belief_target_nll": 6.8859758377075195,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.1775769293308258,
          "query_target_nll": 3.744523525238037,
          "query_top1_on_support": 0.19140625,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.018287722021341324,
          "probe_belief_target_nll": 9.891945838928223,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.005976315122097731,
          "decoder_belief_target_nll": 6.899434566497803,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.2240912914276123,
          "query_target_nll": 3.1704440116882324,
          "query_top1_on_support": 0.26171875,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.03581526502966881,
          "probe_belief_target_nll": 5.480973243713379,
          "probe_belief_top1_on_support": 0.046875,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.005949617363512516,
          "decoder_belief_target_nll": 6.89118766784668,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.2240912914276123,
          "query_target_nll": 3.1704440116882324,
          "query_top1_on_support": 0.26171875,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.03581526502966881,
          "probe_belief_target_nll": 5.480973243713379,
          "probe_belief_top1_on_support": 0.046875,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.005949617363512516,
          "decoder_belief_target_nll": 6.89118766784668,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.2240912914276123,
          "query_target_nll": 3.1704440116882324,
          "query_top1_on_support": 0.26171875,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.03581526502966881,
          "probe_belief_target_nll": 5.480973243713379,
          "probe_belief_top1_on_support": 0.046875,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.005949617363512516,
          "decoder_belief_target_nll": 6.89118766784668,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.2240912914276123,
          "query_target_nll": 3.1704440116882324,
          "query_top1_on_support": 0.26171875,
          "mean_query_support_size": 5.3515625,
          "probe_belief_target_mass": 0.03581526502966881,
          "probe_belief_target_nll": 5.480973243713379,
          "probe_belief_top1_on_support": 0.046875,
          "mean_belief_support_size": 5.765625,
          "decoder_belief_target_mass": 0.005949617363512516,
          "decoder_belief_target_nll": 6.89118766784668,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 5.765625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.17968064546585083,
          "query_target_nll": 12.312056541442871,
          "query_top1_on_support": 0.1796875,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.006099098362028599,
          "probe_belief_target_nll": 18.98821449279785,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007292720023542643,
          "decoder_belief_target_nll": 6.871679306030273,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.19227243959903717,
          "query_target_nll": 5.079844951629639,
          "query_top1_on_support": 0.203125,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.01553806010633707,
          "probe_belief_target_nll": 13.401219367980957,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007247653789818287,
          "decoder_belief_target_nll": 6.876961708068848,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.20042496919631958,
          "query_target_nll": 4.21448278427124,
          "query_top1_on_support": 0.19921875,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.016765184700489044,
          "probe_belief_target_nll": 11.335171699523926,
          "probe_belief_top1_on_support": 0.01953125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007271512411534786,
          "decoder_belief_target_nll": 6.88698673248291,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.20396770536899567,
          "query_target_nll": 3.8485336303710938,
          "query_top1_on_support": 0.1796875,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.012134723365306854,
          "probe_belief_target_nll": 9.906573295593262,
          "probe_belief_top1_on_support": 0.0390625,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007236457895487547,
          "decoder_belief_target_nll": 6.886918067932129,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.2652170956134796,
          "query_target_nll": 3.1494548320770264,
          "query_top1_on_support": 0.29296875,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.03742814064025879,
          "probe_belief_target_nll": 5.442135334014893,
          "probe_belief_top1_on_support": 0.05078125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007243808824568987,
          "decoder_belief_target_nll": 6.884702205657959,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.2652170956134796,
          "query_target_nll": 3.1494548320770264,
          "query_top1_on_support": 0.29296875,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.03742814064025879,
          "probe_belief_target_nll": 5.442135334014893,
          "probe_belief_top1_on_support": 0.05078125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007243808824568987,
          "decoder_belief_target_nll": 6.884702205657959,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.2652170956134796,
          "query_target_nll": 3.1494548320770264,
          "query_top1_on_support": 0.29296875,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.03742814064025879,
          "probe_belief_target_nll": 5.442135334014893,
          "probe_belief_top1_on_support": 0.05078125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007243808824568987,
          "decoder_belief_target_nll": 6.884702205657959,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 8,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.2652170956134796,
          "query_target_nll": 3.1494548320770264,
          "query_top1_on_support": 0.29296875,
          "mean_query_support_size": 6.21875,
          "probe_belief_target_mass": 0.03742814064025879,
          "probe_belief_target_nll": 5.442135334014893,
          "probe_belief_top1_on_support": 0.05078125,
          "mean_belief_support_size": 6.9609375,
          "decoder_belief_target_mass": 0.007243808824568987,
          "decoder_belief_target_nll": 6.884702205657959,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 6.9609375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.1127329170703888,
          "query_target_nll": 3.960693120956421,
          "query_top1_on_support": 0.12109375,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.004941415041685104,
          "probe_belief_target_nll": 18.538177490234375,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003913012333214283,
          "decoder_belief_target_nll": 6.898942470550537,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.11866828799247742,
          "query_target_nll": 4.8373122215271,
          "query_top1_on_support": 0.12890625,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.004133498761802912,
          "probe_belief_target_nll": 13.69979476928711,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003965060226619244,
          "decoder_belief_target_nll": 6.892287254333496,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.12383434176445007,
          "query_target_nll": 4.5756988525390625,
          "query_top1_on_support": 0.15234375,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.0051222899928689,
          "probe_belief_target_nll": 11.672182083129883,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.0039522103033959866,
          "decoder_belief_target_nll": 6.896213531494141,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.12227220833301544,
          "query_target_nll": 4.570303440093994,
          "query_top1_on_support": 0.12890625,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.004631858319044113,
          "probe_belief_target_nll": 10.645010948181152,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.0039694784209132195,
          "decoder_belief_target_nll": 6.889533996582031,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.13690683245658875,
          "query_target_nll": 4.3028154373168945,
          "query_top1_on_support": 0.14453125,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.006682506296783686,
          "probe_belief_target_nll": 9.165693283081055,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003908152692019939,
          "decoder_belief_target_nll": 6.9043474197387695,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.22544604539871216,
          "query_target_nll": 2.8613147735595703,
          "query_top1_on_support": 0.24609375,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.025360360741615295,
          "probe_belief_target_nll": 5.377765655517578,
          "probe_belief_top1_on_support": 0.02734375,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003961235284805298,
          "decoder_belief_target_nll": 6.890734672546387,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.22544604539871216,
          "query_target_nll": 2.8613147735595703,
          "query_top1_on_support": 0.24609375,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.025360360741615295,
          "probe_belief_target_nll": 5.377765655517578,
          "probe_belief_top1_on_support": 0.02734375,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003961235284805298,
          "decoder_belief_target_nll": 6.890734672546387,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.22544604539871216,
          "query_target_nll": 2.8613147735595703,
          "query_top1_on_support": 0.24609375,
          "mean_query_support_size": 3.5078125,
          "probe_belief_target_mass": 0.025360360741615295,
          "probe_belief_target_nll": 5.377765655517578,
          "probe_belief_top1_on_support": 0.02734375,
          "mean_belief_support_size": 3.8125,
          "decoder_belief_target_mass": 0.003961235284805298,
          "decoder_belief_target_nll": 6.890734672546387,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 3.8125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.12214994430541992,
          "query_target_nll": 3.916426181793213,
          "query_top1_on_support": 0.125,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.005535973235964775,
          "probe_belief_target_nll": 19.048948287963867,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004831121768802404,
          "decoder_belief_target_nll": 6.897492408752441,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.12080468982458115,
          "query_target_nll": 4.673115253448486,
          "query_top1_on_support": 0.13671875,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.005794924683868885,
          "probe_belief_target_nll": 13.310545921325684,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004836676642298698,
          "decoder_belief_target_nll": 6.887819290161133,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.11591262370347977,
          "query_target_nll": 4.788409233093262,
          "query_top1_on_support": 0.1484375,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.0089454660192132,
          "probe_belief_target_nll": 11.653854370117188,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004860647488385439,
          "decoder_belief_target_nll": 6.88640832901001,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.1298145353794098,
          "query_target_nll": 4.582821369171143,
          "query_top1_on_support": 0.1640625,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.007938312366604805,
          "probe_belief_target_nll": 10.345158576965332,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.0048925140872597694,
          "decoder_belief_target_nll": 6.875855445861816,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.15379562973976135,
          "query_target_nll": 4.060863494873047,
          "query_top1_on_support": 0.140625,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.008494874462485313,
          "probe_belief_target_nll": 9.084404945373535,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004855196923017502,
          "decoder_belief_target_nll": 6.882272243499756,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.24192126095294952,
          "query_target_nll": 2.8223557472229004,
          "query_top1_on_support": 0.2890625,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.026494011282920837,
          "probe_belief_target_nll": 5.444434642791748,
          "probe_belief_top1_on_support": 0.01953125,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004840375855565071,
          "decoder_belief_target_nll": 6.881380558013916,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.24192126095294952,
          "query_target_nll": 2.8223557472229004,
          "query_top1_on_support": 0.2890625,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.026494011282920837,
          "probe_belief_target_nll": 5.444434642791748,
          "probe_belief_top1_on_support": 0.01953125,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004840375855565071,
          "decoder_belief_target_nll": 6.881380558013916,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.24192126095294952,
          "query_target_nll": 2.8223557472229004,
          "query_top1_on_support": 0.2890625,
          "mean_query_support_size": 3.7265625,
          "probe_belief_target_mass": 0.026494011282920837,
          "probe_belief_target_nll": 5.444434642791748,
          "probe_belief_top1_on_support": 0.01953125,
          "mean_belief_support_size": 4.64453125,
          "decoder_belief_target_mass": 0.004840375855565071,
          "decoder_belief_target_nll": 6.881380558013916,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.13083091378211975,
          "query_target_nll": 3.717766761779785,
          "query_top1_on_support": 0.12109375,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.00358447153121233,
          "probe_belief_target_nll": 19.073143005371094,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004591030068695545,
          "decoder_belief_target_nll": 6.893551826477051,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.1251571625471115,
          "query_target_nll": 3.860501289367676,
          "query_top1_on_support": 0.12890625,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.006148833315819502,
          "probe_belief_target_nll": 13.867976188659668,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004582506604492664,
          "decoder_belief_target_nll": 6.8876848220825195,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.12689608335494995,
          "query_target_nll": 3.9027607440948486,
          "query_top1_on_support": 0.10546875,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.006514372304081917,
          "probe_belief_target_nll": 11.954669952392578,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004605238325893879,
          "decoder_belief_target_nll": 6.880591869354248,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.13241040706634521,
          "query_target_nll": 3.7551023960113525,
          "query_top1_on_support": 0.12890625,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.005720123182982206,
          "probe_belief_target_nll": 10.768220901489258,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004625948145985603,
          "decoder_belief_target_nll": 6.881738662719727,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.12623238563537598,
          "query_target_nll": 3.7817883491516113,
          "query_top1_on_support": 0.11328125,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.007192752789705992,
          "probe_belief_target_nll": 9.351972579956055,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.0045805685222148895,
          "decoder_belief_target_nll": 6.886223316192627,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.16828960180282593,
          "query_target_nll": 3.206430435180664,
          "query_top1_on_support": 0.19921875,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.026202604174613953,
          "probe_belief_target_nll": 5.386316299438477,
          "probe_belief_top1_on_support": 0.03515625,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004609733819961548,
          "decoder_belief_target_nll": 6.877379417419434,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.16828960180282593,
          "query_target_nll": 3.206430435180664,
          "query_top1_on_support": 0.19921875,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.026202604174613953,
          "probe_belief_target_nll": 5.386316299438477,
          "probe_belief_top1_on_support": 0.03515625,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004609733819961548,
          "decoder_belief_target_nll": 6.877379417419434,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.16828960180282593,
          "query_target_nll": 3.206430435180664,
          "query_top1_on_support": 0.19921875,
          "mean_query_support_size": 4.0,
          "probe_belief_target_mass": 0.026202604174613953,
          "probe_belief_target_nll": 5.386316299438477,
          "probe_belief_top1_on_support": 0.03515625,
          "mean_belief_support_size": 4.4140625,
          "decoder_belief_target_mass": 0.004609733819961548,
          "decoder_belief_target_nll": 6.877379417419434,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.4140625
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.15622791647911072,
          "query_target_nll": 12.036298751831055,
          "query_top1_on_support": 0.15625,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.0049526020884513855,
          "probe_belief_target_nll": 18.828533172607422,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004467158578336239,
          "decoder_belief_target_nll": 6.890968322753906,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.1472119837999344,
          "query_target_nll": 4.963583946228027,
          "query_top1_on_support": 0.1796875,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.00708037381991744,
          "probe_belief_target_nll": 13.109294891357422,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004505413584411144,
          "decoder_belief_target_nll": 6.881905555725098,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.1345149278640747,
          "query_target_nll": 4.28691291809082,
          "query_top1_on_support": 0.140625,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.004892526660114527,
          "probe_belief_target_nll": 11.455622673034668,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004498102702200413,
          "decoder_belief_target_nll": 6.883551120758057,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.12799084186553955,
          "query_target_nll": 3.978501081466675,
          "query_top1_on_support": 0.109375,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.005585440434515476,
          "probe_belief_target_nll": 10.700553894042969,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004464133642613888,
          "decoder_belief_target_nll": 6.8979973793029785,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.13156718015670776,
          "query_target_nll": 3.884058713912964,
          "query_top1_on_support": 0.12109375,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.007948249578475952,
          "probe_belief_target_nll": 9.646279335021973,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004440234042704105,
          "decoder_belief_target_nll": 6.9066948890686035,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.1823592185974121,
          "query_target_nll": 3.1739234924316406,
          "query_top1_on_support": 0.20703125,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.0295335054397583,
          "probe_belief_target_nll": 5.3053717613220215,
          "probe_belief_top1_on_support": 0.046875,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004479144234210253,
          "decoder_belief_target_nll": 6.885735988616943,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.1823592185974121,
          "query_target_nll": 3.1739234924316406,
          "query_top1_on_support": 0.20703125,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.0295335054397583,
          "probe_belief_target_nll": 5.3053717613220215,
          "probe_belief_top1_on_support": 0.046875,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004479144234210253,
          "decoder_belief_target_nll": 6.885735988616943,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 12,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.1823592185974121,
          "query_target_nll": 3.1739234924316406,
          "query_top1_on_support": 0.20703125,
          "mean_query_support_size": 4.078125,
          "probe_belief_target_mass": 0.0295335054397583,
          "probe_belief_target_nll": 5.3053717613220215,
          "probe_belief_top1_on_support": 0.046875,
          "mean_belief_support_size": 4.32421875,
          "decoder_belief_target_mass": 0.004479144234210253,
          "decoder_belief_target_nll": 6.885735988616943,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 4.32421875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.07796677201986313,
          "query_target_nll": 3.9148945808410645,
          "query_top1_on_support": 0.0859375,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.0030782646499574184,
          "probe_belief_target_nll": 19.416824340820312,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.0026605655439198017,
          "decoder_belief_target_nll": 6.903154373168945,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.07288013398647308,
          "query_target_nll": 4.964567184448242,
          "query_top1_on_support": 0.06640625,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.0032100947573781013,
          "probe_belief_target_nll": 14.3175048828125,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.002680230885744095,
          "decoder_belief_target_nll": 6.898764610290527,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.07316826283931732,
          "query_target_nll": 4.724486827850342,
          "query_top1_on_support": 0.078125,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.002710676984861493,
          "probe_belief_target_nll": 11.384746551513672,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.002703613368794322,
          "decoder_belief_target_nll": 6.8837385177612305,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.07149238884449005,
          "query_target_nll": 4.7049970626831055,
          "query_top1_on_support": 0.06640625,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.002240848960354924,
          "probe_belief_target_nll": 10.50899887084961,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.0027188523672521114,
          "decoder_belief_target_nll": 6.874066352844238,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.07822994887828827,
          "query_target_nll": 4.45457649230957,
          "query_top1_on_support": 0.0859375,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.0028842277824878693,
          "probe_belief_target_nll": 9.950956344604492,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.002682735910639167,
          "decoder_belief_target_nll": 6.891702651977539,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.09178243577480316,
          "query_target_nll": 4.246744155883789,
          "query_top1_on_support": 0.0703125,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.003827423322945833,
          "probe_belief_target_nll": 9.12129020690918,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.0027144351042807102,
          "decoder_belief_target_nll": 6.873781204223633,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.1533609926700592,
          "query_target_nll": 2.893183946609497,
          "query_top1_on_support": 0.14453125,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.01598162204027176,
          "probe_belief_target_nll": 5.450960636138916,
          "probe_belief_top1_on_support": 0.0234375,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.0027115033008158207,
          "decoder_belief_target_nll": 6.88720703125,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.1533609926700592,
          "query_target_nll": 2.893183946609497,
          "query_top1_on_support": 0.14453125,
          "mean_query_support_size": 2.296875,
          "probe_belief_target_mass": 0.01598162204027176,
          "probe_belief_target_nll": 5.450960636138916,
          "probe_belief_top1_on_support": 0.0234375,
          "mean_belief_support_size": 2.58984375,
          "decoder_belief_target_mass": 0.0027115033008158207,
          "decoder_belief_target_nll": 6.88720703125,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 2.58984375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.07609491050243378,
          "query_target_nll": 4.056446075439453,
          "query_top1_on_support": 0.0859375,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.003410288831219077,
          "probe_belief_target_nll": 19.170536041259766,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.002909217495471239,
          "decoder_belief_target_nll": 6.901834487915039,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.07888618856668472,
          "query_target_nll": 4.791328430175781,
          "query_top1_on_support": 0.09765625,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.002960572484880686,
          "probe_belief_target_nll": 13.900794982910156,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.002972414018586278,
          "decoder_belief_target_nll": 6.864245414733887,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.0765303522348404,
          "query_target_nll": 4.803711891174316,
          "query_top1_on_support": 0.0625,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.0032843034714460373,
          "probe_belief_target_nll": 11.842825889587402,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0029758652672171593,
          "decoder_belief_target_nll": 6.86671257019043,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.07738487422466278,
          "query_target_nll": 4.726408958435059,
          "query_top1_on_support": 0.078125,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.002561418805271387,
          "probe_belief_target_nll": 10.43895149230957,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0029507153667509556,
          "decoder_belief_target_nll": 6.88248348236084,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.07660073786973953,
          "query_target_nll": 4.688850402832031,
          "query_top1_on_support": 0.0859375,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.002743578515946865,
          "probe_belief_target_nll": 10.483691215515137,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0029825340025126934,
          "decoder_belief_target_nll": 6.8748884201049805,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.09137071669101715,
          "query_target_nll": 4.336093425750732,
          "query_top1_on_support": 0.08984375,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.003962407354265451,
          "probe_belief_target_nll": 9.764184951782227,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0029724985361099243,
          "decoder_belief_target_nll": 6.874992370605469,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.16480566561222076,
          "query_target_nll": 2.8717193603515625,
          "query_top1_on_support": 0.1796875,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.020051904022693634,
          "probe_belief_target_nll": 5.407364368438721,
          "probe_belief_top1_on_support": 0.02734375,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0029659937135875225,
          "decoder_belief_target_nll": 6.880322456359863,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.16480566561222076,
          "query_target_nll": 2.8717193603515625,
          "query_top1_on_support": 0.1796875,
          "mean_query_support_size": 2.44140625,
          "probe_belief_target_mass": 0.020051904022693634,
          "probe_belief_target_nll": 5.407364368438721,
          "probe_belief_top1_on_support": 0.02734375,
          "mean_belief_support_size": 2.828125,
          "decoder_belief_target_mass": 0.0029659937135875225,
          "decoder_belief_target_nll": 6.880322456359863,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.828125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.0752723217010498,
          "query_target_nll": 3.675793170928955,
          "query_top1_on_support": 0.08203125,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.0023179668933153152,
          "probe_belief_target_nll": 19.337644577026367,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.002578151412308216,
          "decoder_belief_target_nll": 6.881507873535156,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.07127004861831665,
          "query_target_nll": 3.892303943634033,
          "query_top1_on_support": 0.09765625,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.003746545407921076,
          "probe_belief_target_nll": 14.666508674621582,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.0025648646987974644,
          "decoder_belief_target_nll": 6.893634796142578,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.06755495071411133,
          "query_target_nll": 3.9433393478393555,
          "query_top1_on_support": 0.08203125,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.0031735512893646955,
          "probe_belief_target_nll": 12.248804092407227,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.002560135442763567,
          "decoder_belief_target_nll": 6.887178421020508,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.07924941182136536,
          "query_target_nll": 3.7144405841827393,
          "query_top1_on_support": 0.0859375,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.0022982414811849594,
          "probe_belief_target_nll": 10.798739433288574,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.002564901951700449,
          "decoder_belief_target_nll": 6.886247158050537,
          "decoder_belief_top1_on_support": 0.01171875,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.08330848067998886,
          "query_target_nll": 3.7375881671905518,
          "query_top1_on_support": 0.08984375,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.003733821213245392,
          "probe_belief_target_nll": 10.008598327636719,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.0025386675260961056,
          "decoder_belief_target_nll": 6.896212577819824,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.07289381325244904,
          "query_target_nll": 3.8154892921447754,
          "query_top1_on_support": 0.078125,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.0038078338839113712,
          "probe_belief_target_nll": 9.4683837890625,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.002542847767472267,
          "decoder_belief_target_nll": 6.899700164794922,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.1129687950015068,
          "query_target_nll": 3.1623117923736572,
          "query_top1_on_support": 0.13671875,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.020734883844852448,
          "probe_belief_target_nll": 5.332178115844727,
          "probe_belief_top1_on_support": 0.02734375,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.002574213780462742,
          "decoder_belief_target_nll": 6.877729892730713,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.1129687950015068,
          "query_target_nll": 3.1623117923736572,
          "query_top1_on_support": 0.13671875,
          "mean_query_support_size": 2.31640625,
          "probe_belief_target_mass": 0.020734883844852448,
          "probe_belief_target_nll": 5.332178115844727,
          "probe_belief_top1_on_support": 0.02734375,
          "mean_belief_support_size": 2.46484375,
          "decoder_belief_target_mass": 0.002574213780462742,
          "decoder_belief_target_nll": 6.877729892730713,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 2.46484375
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.09767062216997147,
          "query_target_nll": 12.261209487915039,
          "query_top1_on_support": 0.09765625,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.003074654610827565,
          "probe_belief_target_nll": 19.030813217163086,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.002822312992066145,
          "decoder_belief_target_nll": 6.881270408630371,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.08677688986063004,
          "query_target_nll": 5.358896255493164,
          "query_top1_on_support": 0.0703125,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.003066528122872114,
          "probe_belief_target_nll": 14.444900512695312,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.0028359456919133663,
          "decoder_belief_target_nll": 6.877339839935303,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.08567088842391968,
          "query_target_nll": 4.375212669372559,
          "query_top1_on_support": 0.08984375,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.0036841684486716986,
          "probe_belief_target_nll": 12.495094299316406,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.002821750473231077,
          "decoder_belief_target_nll": 6.887667655944824,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.07946346700191498,
          "query_target_nll": 4.03109884262085,
          "query_top1_on_support": 0.1015625,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.002167962957173586,
          "probe_belief_target_nll": 11.182050704956055,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.0028297381941229105,
          "decoder_belief_target_nll": 6.8809967041015625,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.08351743221282959,
          "query_target_nll": 3.9680535793304443,
          "query_top1_on_support": 0.0859375,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.00316959829069674,
          "probe_belief_target_nll": 10.316816329956055,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.002851146040484309,
          "decoder_belief_target_nll": 6.875732421875,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.07817612588405609,
          "query_target_nll": 3.9505667686462402,
          "query_top1_on_support": 0.0625,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.0047333575785160065,
          "probe_belief_target_nll": 9.303445816040039,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.002846138784661889,
          "decoder_belief_target_nll": 6.873347282409668,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.11342555284500122,
          "query_target_nll": 3.2204227447509766,
          "query_top1_on_support": 0.13671875,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.01580173522233963,
          "probe_belief_target_nll": 5.513126373291016,
          "probe_belief_top1_on_support": 0.0234375,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.002810610458254814,
          "decoder_belief_target_nll": 6.88219690322876,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 16,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.11342555284500122,
          "query_target_nll": 3.2204227447509766,
          "query_top1_on_support": 0.13671875,
          "mean_query_support_size": 2.5546875,
          "probe_belief_target_mass": 0.01580173522233963,
          "probe_belief_target_nll": 5.513126373291016,
          "probe_belief_top1_on_support": 0.0234375,
          "mean_belief_support_size": 2.69921875,
          "decoder_belief_target_mass": 0.002810610458254814,
          "decoder_belief_target_nll": 6.88219690322876,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 2.69921875
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.0545433834195137,
          "query_target_nll": 3.8925232887268066,
          "query_top1_on_support": 0.06640625,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.0017636767588555813,
          "probe_belief_target_nll": 18.87615203857422,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0016978500643745065,
          "decoder_belief_target_nll": 6.892560005187988,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.050416797399520874,
          "query_target_nll": 5.373021602630615,
          "query_top1_on_support": 0.05078125,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.0018305472331121564,
          "probe_belief_target_nll": 15.031593322753906,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.001673350459896028,
          "decoder_belief_target_nll": 6.896378040313721,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.050489701330661774,
          "query_target_nll": 4.8594279289245605,
          "query_top1_on_support": 0.0546875,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.0013959466014057398,
          "probe_belief_target_nll": 12.073762893676758,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0016932014841586351,
          "decoder_belief_target_nll": 6.883866310119629,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.053241241723299026,
          "query_target_nll": 4.699134826660156,
          "query_top1_on_support": 0.05859375,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.0018582199700176716,
          "probe_belief_target_nll": 10.513172149658203,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0016938212793320417,
          "decoder_belief_target_nll": 6.889729976654053,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.05504796653985977,
          "query_target_nll": 4.5955610275268555,
          "query_top1_on_support": 0.0625,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.002157565439119935,
          "probe_belief_target_nll": 10.281576156616211,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0017138629918918014,
          "decoder_belief_target_nll": 6.885958194732666,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.05082276463508606,
          "query_target_nll": 4.738102912902832,
          "query_top1_on_support": 0.0703125,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.002452777698636055,
          "probe_belief_target_nll": 10.281059265136719,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0017020186642184854,
          "decoder_belief_target_nll": 6.887733459472656,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.04877123981714249,
          "query_target_nll": 4.587276458740234,
          "query_top1_on_support": 0.046875,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.001769182737916708,
          "probe_belief_target_nll": 10.138912200927734,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.001711843884550035,
          "decoder_belief_target_nll": 6.886379718780518,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.10235182195901871,
          "query_target_nll": 2.895606279373169,
          "query_top1_on_support": 0.12890625,
          "mean_query_support_size": 1.56640625,
          "probe_belief_target_mass": 0.010676495730876923,
          "probe_belief_target_nll": 5.475768089294434,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 1.64453125,
          "decoder_belief_target_mass": 0.0017089392058551311,
          "decoder_belief_target_nll": 6.883117198944092,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.64453125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.05046955496072769,
          "query_target_nll": 3.944303512573242,
          "query_top1_on_support": 0.0390625,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.002537171356379986,
          "probe_belief_target_nll": 18.67137908935547,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018663487862795591,
          "decoder_belief_target_nll": 6.888060569763184,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.0495598167181015,
          "query_target_nll": 4.486264705657959,
          "query_top1_on_support": 0.04296875,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0022404640913009644,
          "probe_belief_target_nll": 13.584051132202148,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.001880388823337853,
          "decoder_belief_target_nll": 6.869832992553711,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.055749885737895966,
          "query_target_nll": 4.372976779937744,
          "query_top1_on_support": 0.0546875,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0029107648879289627,
          "probe_belief_target_nll": 11.364819526672363,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018495803233236074,
          "decoder_belief_target_nll": 6.885612964630127,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.05526028573513031,
          "query_target_nll": 4.240878105163574,
          "query_top1_on_support": 0.06640625,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.00212241243571043,
          "probe_belief_target_nll": 10.560112953186035,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018469550414010882,
          "decoder_belief_target_nll": 6.889764308929443,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.05614112690091133,
          "query_target_nll": 4.3783793449401855,
          "query_top1_on_support": 0.04296875,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0018042087322100997,
          "probe_belief_target_nll": 10.111257553100586,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018654151353985071,
          "decoder_belief_target_nll": 6.875096321105957,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.056102003902196884,
          "query_target_nll": 4.415143966674805,
          "query_top1_on_support": 0.078125,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0017401274526491761,
          "probe_belief_target_nll": 9.898533821105957,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018577547743916512,
          "decoder_belief_target_nll": 6.868346214294434,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.05023609474301338,
          "query_target_nll": 4.661558628082275,
          "query_top1_on_support": 0.0625,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.0020934033673256636,
          "probe_belief_target_nll": 10.25639533996582,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018599433824419975,
          "decoder_belief_target_nll": 6.8752312660217285,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.11320751905441284,
          "query_target_nll": 2.846771478652954,
          "query_top1_on_support": 0.1484375,
          "mean_query_support_size": 1.578125,
          "probe_belief_target_mass": 0.013075247406959534,
          "probe_belief_target_nll": 5.452033042907715,
          "probe_belief_top1_on_support": 0.0234375,
          "mean_belief_support_size": 1.78125,
          "decoder_belief_target_mass": 0.0018672222504392266,
          "decoder_belief_target_nll": 6.879985809326172,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.78125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.04805327206850052,
          "query_target_nll": 3.7452681064605713,
          "query_top1_on_support": 0.0625,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.001562569523230195,
          "probe_belief_target_nll": 19.639110565185547,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0016582157695665956,
          "decoder_belief_target_nll": 6.878258228302002,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.05149130895733833,
          "query_target_nll": 3.797159194946289,
          "query_top1_on_support": 0.06640625,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0008999384590424597,
          "probe_belief_target_nll": 14.234174728393555,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.001642600167542696,
          "decoder_belief_target_nll": 6.88369083404541,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.04774951934814453,
          "query_target_nll": 3.7304162979125977,
          "query_top1_on_support": 0.05859375,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0011481763795018196,
          "probe_belief_target_nll": 11.68929386138916,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0016632287297397852,
          "decoder_belief_target_nll": 6.872076034545898,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.050339192152023315,
          "query_target_nll": 3.74833345413208,
          "query_top1_on_support": 0.04296875,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.002387874759733677,
          "probe_belief_target_nll": 10.529827117919922,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0016582382377237082,
          "decoder_belief_target_nll": 6.877333641052246,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.04724283516407013,
          "query_target_nll": 3.8155853748321533,
          "query_top1_on_support": 0.046875,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0013702325522899628,
          "probe_belief_target_nll": 10.291004180908203,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0016408979427069426,
          "decoder_belief_target_nll": 6.8872575759887695,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.04881110414862633,
          "query_target_nll": 3.681596279144287,
          "query_top1_on_support": 0.05859375,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.0016018322203308344,
          "probe_belief_target_nll": 10.292238235473633,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0016570116858929396,
          "decoder_belief_target_nll": 6.880745887756348,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.05095507949590683,
          "query_target_nll": 3.7174391746520996,
          "query_top1_on_support": 0.05078125,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.001925286604091525,
          "probe_belief_target_nll": 9.933907508850098,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.0016283796867355704,
          "decoder_belief_target_nll": 6.897070407867432,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_PLUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.07574611902236938,
          "query_target_nll": 3.210456371307373,
          "query_top1_on_support": 0.125,
          "mean_query_support_size": 1.515625,
          "probe_belief_target_mass": 0.010490939021110535,
          "probe_belief_target_nll": 5.558432579040527,
          "probe_belief_top1_on_support": 0.01171875,
          "mean_belief_support_size": 1.578125,
          "decoder_belief_target_mass": 0.001636333647184074,
          "decoder_belief_target_nll": 6.890581130981445,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.578125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 0,
          "n": 256,
          "query_target_mass": 0.05857173353433609,
          "query_target_nll": 11.975107192993164,
          "query_top1_on_support": 0.05859375,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0017345352098345757,
          "probe_belief_target_nll": 18.655864715576172,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0016290670027956367,
          "decoder_belief_target_nll": 6.8749494552612305,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 1,
          "n": 256,
          "query_target_mass": 0.06017947569489479,
          "query_target_nll": 4.8631768226623535,
          "query_top1_on_support": 0.05078125,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.002939745783805847,
          "probe_belief_target_nll": 13.268908500671387,
          "probe_belief_top1_on_support": 0.0078125,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0016383202746510506,
          "decoder_belief_target_nll": 6.865511894226074,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 2,
          "n": 256,
          "query_target_mass": 0.05241341143846512,
          "query_target_nll": 4.173548221588135,
          "query_top1_on_support": 0.05859375,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0010552826570346951,
          "probe_belief_target_nll": 11.693658828735352,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.001640757080167532,
          "decoder_belief_target_nll": 6.8675031661987305,
          "decoder_belief_top1_on_support": 0.015625,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 4,
          "n": 256,
          "query_target_mass": 0.05089771747589111,
          "query_target_nll": 3.9172375202178955,
          "query_top1_on_support": 0.0546875,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.002121605910360813,
          "probe_belief_target_nll": 10.22000503540039,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0016125631518661976,
          "decoder_belief_target_nll": 6.886224269866943,
          "decoder_belief_top1_on_support": 0.0078125,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 8,
          "n": 256,
          "query_target_mass": 0.04816032573580742,
          "query_target_nll": 3.91409969329834,
          "query_top1_on_support": 0.046875,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0015802973648533225,
          "probe_belief_target_nll": 9.948531150817871,
          "probe_belief_top1_on_support": 0.0,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0016280205454677343,
          "decoder_belief_target_nll": 6.873702526092529,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 12,
          "n": 256,
          "query_target_mass": 0.052890703082084656,
          "query_target_nll": 3.8442678451538086,
          "query_top1_on_support": 0.0546875,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.0014868470607325435,
          "probe_belief_target_nll": 9.994505882263184,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0016273310175165534,
          "decoder_belief_target_nll": 6.872465133666992,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 16,
          "n": 256,
          "query_target_mass": 0.05423782020807266,
          "query_target_nll": 3.710984945297241,
          "query_top1_on_support": 0.0546875,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.001985232811421156,
          "probe_belief_target_nll": 9.856807708740234,
          "probe_belief_top1_on_support": 0.00390625,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0016131988959386945,
          "decoder_belief_target_nll": 6.878167629241943,
          "decoder_belief_top1_on_support": 0.0,
          "mean_decoder_belief_support_size": 1.55078125
        },
        {
          "model": "dense",
          "supervision": "sampled_final",
          "length": 24,
          "query_type": "A_MINUS_B",
          "k": 24,
          "n": 256,
          "query_target_mass": 0.08383125066757202,
          "query_target_nll": 3.1150407791137695,
          "query_top1_on_support": 0.09765625,
          "mean_query_support_size": 1.51953125,
          "probe_belief_target_mass": 0.013209330849349499,
          "probe_belief_target_nll": 5.460287570953369,
          "probe_belief_top1_on_support": 0.015625,
          "mean_belief_support_size": 1.55078125,
          "decoder_belief_target_mass": 0.0015847078757360578,
          "decoder_belief_target_nll": 6.897161960601807,
          "decoder_belief_top1_on_support": 0.00390625,
          "mean_decoder_belief_support_size": 1.55078125
        }
      ]
    }
  ],
  "checkpoints": [
    "large_artifacts/dense_supervision_ladder/checkpoints/main_sampled_final_mod31/checkpoint_final.pt"
  ]
}