{
  "contrasts": [
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.004975124378109541,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.08457711442786098
      },
      "delta": 0.039800995024875774,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 0.21757824217578242,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T01",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.31499999999999995,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.16999999999999998
      },
      "delta": -0.24,
      "family_size": 98,
      "metric": "acc_norm,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.009799020097990201,
      "p_raw": 9.999000099990002e-05,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T02",
      "unit": "score",
      "verdict": "worse"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 7.5,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 17.5
      },
      "delta": 12.5,
      "family_size": 98,
      "metric": "accuracy",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.009799020097990201,
      "p_raw": 9.999000099990002e-05,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T04",
      "unit": "percent",
      "verdict": "better"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.325,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.195
      },
      "delta": -0.26,
      "family_size": 98,
      "metric": "acc_norm,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.009799020097990201,
      "p_raw": 9.999000099990002e-05,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T05",
      "unit": "score",
      "verdict": "worse"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.17499999999999993,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.0050000000000000044
      },
      "delta": -0.08499999999999996,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 0.07949205079492051,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T07",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.09000000000000002,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.010000000000000009
      },
      "delta": -0.03999999999999998,
      "family_size": 98,
      "metric": "acc_norm,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 0.1653834616538346,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T08",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.25,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.135
      },
      "delta": -0.19000000000000006,
      "family_size": 98,
      "metric": "acc_norm,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.009799020097990201,
      "p_raw": 9.999000099990002e-05,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T09",
      "unit": "score",
      "verdict": "worse"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.16999999999999993,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.010000000000000009
      },
      "delta": -0.08000000000000007,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 0.10278972102789721,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T10",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.19999999999999996,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.04999999999999993
      },
      "delta": -0.12,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 100,
      "p_holm": 0.34283447265625,
      "p_raw": 0.004180908203125,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T11",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.32142857142857145,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.017857142857142794
      },
      "delta": -0.1607142857142857,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 56,
      "p_holm": 1.0,
      "p_raw": 0.10468953104689531,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T12",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.12,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.05999999999999994
      },
      "delta": -0.030000000000000027,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 0.5862413758624138,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T15",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.135,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.025000000000000022
      },
      "delta": -0.05499999999999999,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 0.22567743225677434,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T16",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.12,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.02999999999999997
      },
      "delta": -0.044999999999999984,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 0.28847115288471153,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T17",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.065,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.08499999999999996
      },
      "delta": 0.010000000000000009,
      "family_size": 98,
      "metric": "acc_norm,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 0.8964103589641036,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T20",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.37500000000000006,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.18000000000000005
      },
      "delta": -0.27999999999999997,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.009799020097990201,
      "p_raw": 9.999000099990002e-05,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T22",
      "unit": "score",
      "verdict": "worse"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.19499999999999995,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.04500000000000004
      },
      "delta": -0.11999999999999994,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.19318068193180682,
      "p_raw": 0.0022997700229977,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T23",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.09500000000000003,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.09499999999999997
      },
      "delta": 0.0,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 1.0,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T24",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.06895846122365462,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.10913930789707185
      },
      "delta": 0.022458628841607597,
      "family_size": 98,
      "metric": "f1,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": null,
      "p_raw": 0.5989401059894011,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": null,
        "artifact_sha256": null,
        "status": "pending"
      },
      "subset": "full",
      "task_id": "T25",
      "unit": "score",
      "verdict": "withheld"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.13883677298311442,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.4227642276422764
      },
      "delta": 0.15609756097560978,
      "family_size": 98,
      "metric": "f1,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 40,
      "p_holm": 1.0,
      "p_raw": 0.14908509149085092,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "hit_free_answer_fields",
      "task_id": "T25",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.275,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.05499999999999999
      },
      "delta": -0.16499999999999998,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.4079592040795921,
      "p_raw": 0.005099490050994901,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T26",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.185,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.015000000000000013
      },
      "delta": -0.10000000000000003,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 0.029197080291970802,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T29",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.10000000000000003,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.06000000000000005
      },
      "delta": -0.020000000000000018,
      "family_size": 98,
      "metric": "levenshtein,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": null,
      "p_raw": 0.7123287671232876,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": null,
        "artifact_sha256": null,
        "status": "pending"
      },
      "subset": "full",
      "task_id": "T30",
      "unit": "score",
      "verdict": "withheld"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.10050251256281406,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.06030150753768848
      },
      "delta": -0.02010050251256279,
      "family_size": 98,
      "metric": "levenshtein,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 199,
      "p_holm": 1.0,
      "p_raw": 0.7215278472152785,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "hit_free_answer_fields",
      "task_id": "T30",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.14500000000000002,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.0050000000000000044
      },
      "delta": -0.07499999999999998,
      "family_size": 98,
      "metric": "levenshtein,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": null,
      "p_raw": 0.05269473052694731,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": null,
        "artifact_sha256": null,
        "status": "pending"
      },
      "subset": "full",
      "task_id": "T31",
      "unit": "score",
      "verdict": "withheld"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.15822784810126583,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.006329113924050639
      },
      "delta": -0.08227848101265822,
      "family_size": 98,
      "metric": "levenshtein,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 158,
      "p_holm": 1.0,
      "p_raw": 0.05419458054194581,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "hit_free_answer_fields",
      "task_id": "T31",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.35736752328049853,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.20044404510648645
      },
      "delta": -0.28009458020496336,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.009799020097990201,
      "p_raw": 9.999000099990002e-05,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T32",
      "unit": "score",
      "verdict": "worse"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.53,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.39
      },
      "delta": -0.46,
      "family_size": 98,
      "metric": "exact_match,strict-match",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.009799020097990201,
      "p_raw": 9.999000099990002e-05,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T33",
      "unit": "score",
      "verdict": "worse"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.07330648059459305,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.016069938723859534
      },
      "delta": -0.028489426245139504,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 0.21237876212378762,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T35",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.08000000000000002,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.10000000000000003
      },
      "delta": 0.010000000000000009,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 0.9133086691330867,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T37",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -20.827199563427776,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -11.701103807248678
      },
      "delta": -16.089329668334905,
      "family_size": 98,
      "metric": "f1,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.009799020097990201,
      "p_raw": 9.999000099990002e-05,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T38",
      "unit": "percent",
      "verdict": "worse"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.0678,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.012799999999999992
      },
      "delta": -0.03894999999999999,
      "family_size": 98,
      "metric": "f1,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.38066193380661933,
      "p_raw": 0.0046995300469953,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T39",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.155,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.034999999999999976
      },
      "delta": -0.09499999999999997,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.32366763323667636,
      "p_raw": 0.0038996100389961006,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T40",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.2652249609511775,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.14013914178668174
      },
      "delta": -0.2023302388296171,
      "family_size": 98,
      "metric": "f1,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.009799020097990201,
      "p_raw": 9.999000099990002e-05,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T41",
      "unit": "score",
      "verdict": "worse"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.115,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.035
      },
      "delta": -0.075,
      "family_size": 98,
      "metric": "exact_match,remove_whitespace",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.024169921875,
      "p_raw": 0.000274658203125,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T42",
      "unit": "score",
      "verdict": "worse"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.36,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.23
      },
      "delta": -0.295,
      "family_size": 98,
      "metric": "exact_match,remove_whitespace",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.009799020097990201,
      "p_raw": 9.999000099990002e-05,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T43",
      "unit": "score",
      "verdict": "worse"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.25,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.07
      },
      "delta": -0.16000000000000003,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.152984701529847,
      "p_raw": 0.0017998200179982,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T45",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.05485222809750534,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.10008447243127017
      },
      "delta": 0.02331219694143971,
      "family_size": 98,
      "metric": "exact_match,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 0.5937406259374063,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T47",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.17000000000000004,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.0050000000000000044
      },
      "delta": -0.08499999999999996,
      "family_size": 98,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 1.0,
      "p_raw": 0.0744925507449255,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T48",
      "unit": "score",
      "verdict": "inconclusive"
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.19666666666666655,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": -0.053333333333333344
      },
      "delta": -0.125,
      "family_size": 98,
      "metric": "f1,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "model_ref": "LiquidAI/LFM2.5-1.2B-Base",
      "n": 200,
      "p_holm": 0.08612956935215323,
      "p_raw": 0.0009899950500247498,
      "protocol_id": "bench50-v0.3.1-panel200",
      "review": {
        "actor": "independent reviewer",
        "artifact_sha256": "bba5721493d2fc6e50ebe483ee24f70b145c9834b1ab0ff6326435ba6e0cb469",
        "status": "accepted"
      },
      "subset": "full",
      "task_id": "T50",
      "unit": "score",
      "verdict": "inconclusive"
    }
  ],
  "models": [
    {
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "repo": "LiquidAI/LFM2.5-1.2B-Base",
      "revision": "7453bca97ca1e67754c4035a4b4c584e1c9dd725",
      "weights_sha256": "7678ab9546a0c51c1fca161876b1efc4f0906277f170b5822045f40fdaf9eeff"
    },
    {
      "model_id": "SlayerLab/Slayer1B-1.06",
      "repo": "SlayerLab/Slayer1B-1.06",
      "revision": "ef81be58742544ad32e1afab38db011fec209609",
      "weights_sha256": "e10df6d50101d8424ff7338c97373ab9e151e359a6fcb23108023ac4a6400214"
    }
  ],
  "observations": [
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.7611940298507465,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.8407960199004979
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 67000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "30aa8b9159f8a599cd98c94cac594a99adba3ea9038b7c56df558ef3e24ccb2a",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "9d6d5c8b444b57a4043af8030c77daca572418a54ace83cd6c16ebe8945faf67",
      "state": "ok",
      "task_id": "T01",
      "unit": "score",
      "value": 0.8009950248756219,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.8059701492537314,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.8756218905472644
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.015,
      "hit_rate_answer_fields": 0.015,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 197,
      "n_hit_free_answer_fields": 197,
      "n_population": 67000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "fe310fa1e741d8e0e79cc574bdbf32123a32eb6575dc51f3276211d31075360a",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "9d6d5c8b444b57a4043af8030c77daca572418a54ace83cd6c16ebe8945faf67",
      "state": "ok",
      "task_id": "T01",
      "unit": "score",
      "value": 0.8407960199004975,
      "value_hit_free": 0.838308457711443,
      "value_hit_free_answer_fields": 0.838308457711443
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.42,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.56
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc_norm,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 1172,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.2495,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "a4e47cddfc2c871dc7498e25a7d382299aa78b23a783d34a64dbd43795662cae",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "1caa76f5cdf6f4d8a0a1eb8970d2537ccdb8eabe8d4a2eea6ca7aeb03835cbf9",
      "state": "ok",
      "task_id": "T02",
      "unit": "score",
      "value": 0.49,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.19,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.31
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.02,
      "hit_rate_answer_fields": 0.02,
      "metric": "acc_norm,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 196,
      "n_hit_free_answer_fields": 196,
      "n_population": 1172,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.2495,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "2b55bec70662eb301a59f68f4e0f49349c282c03290077c22f3f828ebbf8d3e6",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "1caa76f5cdf6f4d8a0a1eb8970d2537ccdb8eabe8d4a2eea6ca7aeb03835cbf9",
      "state": "ok",
      "task_id": "T02",
      "unit": "score",
      "value": 0.25,
      "value_hit_free": 0.25,
      "value_hit_free_answer_fields": 0.25
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 1.6937118906996582,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 1.7778426149146371
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": false,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "byte_perplexity,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 62,
      "n_scored": 62,
      "n_selected": 62,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "abdc9030b69e396fd1ccc227cdb1fd4ffe97bd3dfe3b28e6d31d24b73c19366e",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "e9225fcf908fbf2f54c7e46c0d421378ac4a0e9875e88634acfa376443aa53e6",
      "state": "ok",
      "task_id": "T03",
      "unit": "byte_ppl",
      "value": 1.7347771288213312,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 1.8152554368468583,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 1.878480340094036
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": false,
      "hit_rate": 0.33870967741935487,
      "hit_rate_answer_fields": 0.33870967741935487,
      "metric": "byte_perplexity,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 41,
      "n_hit_free_answer_fields": 41,
      "n_population": 62,
      "n_scored": 62,
      "n_selected": 62,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "99601259b1eb621878611a70aa332e6e2e93b43f508c358b8aa3e7e4e4d7b7d4",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "e9225fcf908fbf2f54c7e46c0d421378ac4a0e9875e88634acfa376443aa53e6",
      "state": "ok",
      "task_id": "T03",
      "unit": "byte_ppl",
      "value": 1.846283835597753,
      "value_hit_free": 1.84293133324836,
      "value_hit_free_answer_fields": 1.84293133324836
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 81.5,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 91.0
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "accuracy",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 3272,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 50.0,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "af27e95954e7260ca44f2186eb6fce69bac0c667a4f8ca90a012df0712fbbaa4",
      "runner_sha256": "51c21059fa8ac25d76dc611567a7390d72fc8541fd0020ef69995d8b9842903f",
      "source_id_digest": "c9aba49d565f69d046d6062b5947cc9f9e0dfc23747684f9564646393e2496fb",
      "state": "ok",
      "task_id": "T04",
      "unit": "percent",
      "value": 86.5,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 97.5,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 100.0
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.08,
      "hit_rate_answer_fields": 0.08,
      "metric": "accuracy",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 184,
      "n_hit_free_answer_fields": 184,
      "n_population": 3272,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 50.0,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "e4bdf566cbdbf3270a806147e90b3dac49faf82a7f3096b7991b1d799508e726",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "c9aba49d565f69d046d6062b5947cc9f9e0dfc23747684f9564646393e2496fb",
      "state": "ok",
      "task_id": "T04",
      "unit": "percent",
      "value": 99.0,
      "value_hit_free": 98.91304347826086,
      "value_hit_free_answer_fields": 98.91304347826086
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.555,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.695
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc_norm,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 10042,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "655e916e75563bd313f9a6f4ae79f2eab2d0e83e17495a999c443d811e164faf",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "927e1904dd5b604b52283e5af60a0a0e94e699c37917e47b488472560ce5b052",
      "state": "ok",
      "task_id": "T05",
      "unit": "score",
      "value": 0.625,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.295,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.435
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.0,
      "hit_rate_answer_fields": 0.0,
      "metric": "acc_norm,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 200,
      "n_hit_free_answer_fields": 200,
      "n_population": 10042,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "094d75e783c29cc45cd6a8f7689ee34d2fccdac156e434436d7dc3528756f8c0",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "927e1904dd5b604b52283e5af60a0a0e94e699c37917e47b488472560ce5b052",
      "state": "ok",
      "task_id": "T05",
      "unit": "score",
      "value": 0.365,
      "value_hit_free": 0.365,
      "value_hit_free_answer_fields": 0.365
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc_norm,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T06",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc_norm,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T06",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.57,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.7
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 1267,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "0162ffaac22b1036be4745bd187162f88b6bad141970841220701c979ed5b7b2",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "98a6dff8613cd375df50acd0a1882b7c9cbdccf4572795f6637a221b1723c64c",
      "state": "ok",
      "task_id": "T07",
      "unit": "score",
      "value": 0.635,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.48,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.615
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "clean",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 1267,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "f3b90322173e03680a88dbbdea5a0836471d30babb40193358b18ee47d9739ba",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "98a6dff8613cd375df50acd0a1882b7c9cbdccf4572795f6637a221b1723c64c",
      "state": "ok",
      "task_id": "T07",
      "unit": "score",
      "value": 0.55,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.295,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.425
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc_norm,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 500,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "0f3986c3f4ee25818862e9c9c611383a9422639aacd79979a9579f5852bf0056",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "f375631aa3eb26df7e84fec2e490bbfa06971ffdf4d986fc3d775cf9fd931f14",
      "state": "ok",
      "task_id": "T08",
      "unit": "score",
      "value": 0.36,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.255,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.385
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.12,
      "hit_rate_answer_fields": 0.12,
      "metric": "acc_norm,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 176,
      "n_hit_free_answer_fields": 176,
      "n_population": 500,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "f72e6b5d9bf5c2d9cd30c1d49fa308e85775f259578ada2643d55568e56a4e55",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "f375631aa3eb26df7e84fec2e490bbfa06971ffdf4d986fc3d775cf9fd931f14",
      "state": "ok",
      "task_id": "T08",
      "unit": "score",
      "value": 0.32,
      "value_hit_free": 0.3409090909090909,
      "value_hit_free_answer_fields": 0.3409090909090909
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.885,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.96
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc_norm,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 1000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "7fe0c5dd8ae879f0c8b50267ca1cf98544515a9f105e7d7b8426b1f52ec7e76b",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "7368315d4c7ec5b6e85d28a4f77f010c0790d5b7dd4b020aea7fd84524f74e70",
      "state": "ok",
      "task_id": "T09",
      "unit": "score",
      "value": 0.925,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.675,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.795
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.105,
      "hit_rate_answer_fields": 0.01,
      "metric": "acc_norm,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 179,
      "n_hit_free_answer_fields": 198,
      "n_population": 1000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "03b7a73657f80bf917844ca81a31762eae4aabcff3cd342ec8768bff5177f243",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "7368315d4c7ec5b6e85d28a4f77f010c0790d5b7dd4b020aea7fd84524f74e70",
      "state": "ok",
      "task_id": "T09",
      "unit": "score",
      "value": 0.735,
      "value_hit_free": 0.7262569832402235,
      "value_hit_free_answer_fields": 0.7323232323232324
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.605,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.735
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 3270,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": false,
        "majority_prediction_share": 0.595,
        "majority_value": 0.615
      },
      "results_sha256": "07a678d9f3c6604d19bdbe19e2bf8ac147c213fbbb6b79cbcc163317566d8170",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "35cec0f1fbcb80691dca498e533d6890402e8ede950e71e0be24da18f86398c0",
      "state": "ok",
      "task_id": "T10",
      "unit": "score",
      "value": 0.67,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.525,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.655
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.185,
      "hit_rate_answer_fields": 0.0,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 163,
      "n_hit_free_answer_fields": 200,
      "n_population": 3270,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": false,
        "majority_prediction_share": 0.915,
        "majority_value": 0.615
      },
      "results_sha256": "90a0c1a46df7289952d73ed759e85740105bc5424470fdf0225cccb96e6af66a",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "35cec0f1fbcb80691dca498e533d6890402e8ede950e71e0be24da18f86398c0",
      "state": "ok",
      "task_id": "T10",
      "unit": "score",
      "value": 0.59,
      "value_hit_free": 0.6012269938650306,
      "value_hit_free_answer_fields": 0.59
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.66,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.83
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 100,
      "n_scored": 100,
      "n_selected": 100,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "a77766212763171efd0de5927a41488aee1c2baf35e77134416a77699795ab6e",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "fa0b4f80a2cc898e96ca6c7d6f5ae0d2dbc46d59d712fc1c20cca88bebd2e114",
      "state": "ok",
      "task_id": "T11",
      "unit": "score",
      "value": 0.75,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.53,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.72
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.21,
      "hit_rate_answer_fields": 0.21,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 79,
      "n_hit_free_answer_fields": 79,
      "n_population": 100,
      "n_scored": 100,
      "n_selected": 100,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "bd2c9f0cd9bc0a75e507f30f08cc0abd802b0f7d6a1cef4d3ee6114551758dce",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "fa0b4f80a2cc898e96ca6c7d6f5ae0d2dbc46d59d712fc1c20cca88bebd2e114",
      "state": "ok",
      "task_id": "T11",
      "unit": "score",
      "value": 0.63,
      "value_hit_free": 0.6075949367088608,
      "value_hit_free_answer_fields": 0.6075949367088608
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.5,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.75
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 56,
      "n_scored": 56,
      "n_selected": 56,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.3333333333333333,
        "constant_prediction": false,
        "majority_prediction_share": 0.5,
        "majority_value": 0.5
      },
      "results_sha256": "017849e99eb758bca2e0132f7ffc036eeadbc4ee5bf292d544ee67c344decbee",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "310b71c1979b30e593368ced475893cd0ce5bee7693b9044e7d7a97ab4ec275e",
      "state": "ok",
      "task_id": "T12",
      "unit": "score",
      "value": 0.625,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.3392857142857143,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.5892857142857143
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "clean",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 56,
      "n_scored": 56,
      "n_selected": 56,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.3333333333333333,
        "constant_prediction": false,
        "majority_prediction_share": 0.9107142857142857,
        "majority_value": 0.5
      },
      "results_sha256": "db4886f81b8ff46e1b338c913c50cabf65473fc9b1d1e2f54dc2a41ee403bcb5",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "310b71c1979b30e593368ced475893cd0ce5bee7693b9044e7d7a97ab4ec275e",
      "state": "ok",
      "task_id": "T12",
      "unit": "score",
      "value": 0.4642857142857143,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.38,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.515
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 638,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": false,
        "majority_prediction_share": 0.98,
        "majority_value": 0.545
      },
      "results_sha256": "5f36bf460c59a4f0b887e8afff73ee0cbb21fbe41b9d29cf5a249568e7b7a4e9",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "bfd76be0170ef9f0a49cd916e9f91c699cef728a1f85d7d56c0cf4dfba70909f",
      "state": "ok",
      "task_id": "T13",
      "unit": "score",
      "value": 0.445,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.475,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.615
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.375,
      "hit_rate_answer_fields": 0.375,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 125,
      "n_hit_free_answer_fields": 125,
      "n_population": 638,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": true,
        "majority_prediction_share": 1.0,
        "majority_value": 0.545
      },
      "results_sha256": "04e2e8015e185a193e075d3353d05cc21d18d895fd0b99e6be2dde1147d71f34",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "bfd76be0170ef9f0a49cd916e9f91c699cef728a1f85d7d56c0cf4dfba70909f",
      "state": "ok",
      "task_id": "T13",
      "unit": "score",
      "value": 0.545,
      "value_hit_free": 0.52,
      "value_hit_free_answer_fields": 0.52
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T14",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T14",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.49,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.625
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 277,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": false,
        "majority_prediction_share": 0.55,
        "majority_value": 0.51
      },
      "results_sha256": "c5224c1ed0b0976f0211e727770b6740f7a9c544115d6e39dba5aae8b952182f",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "1b9b48d87ab1ba4f38607371c083043851102d89967b77b6ba9dd90afd9006c8",
      "state": "ok",
      "task_id": "T15",
      "unit": "score",
      "value": 0.56,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.46,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.6
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.03,
      "hit_rate_answer_fields": 0.03,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 194,
      "n_hit_free_answer_fields": 194,
      "n_population": 277,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": false,
        "majority_prediction_share": 0.85,
        "majority_value": 0.51
      },
      "results_sha256": "c12763b0ad64f58f673a86ce7b3766449b1c7e28bd10cb7da98358df8df7169c",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "1b9b48d87ab1ba4f38607371c083043851102d89967b77b6ba9dd90afd9006c8",
      "state": "ok",
      "task_id": "T15",
      "unit": "score",
      "value": 0.53,
      "value_hit_free": 0.5257731958762887,
      "value_hit_free_answer_fields": 0.5257731958762887
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.35,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.48
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 1200,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.33333333333333326,
        "constant_prediction": false,
        "majority_prediction_share": 0.515,
        "majority_value": 0.36
      },
      "results_sha256": "d5963aa9df00c8a449b72f6f23e8877bb2d09032c9698eb060792daa83ae3be0",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "8f6560a9a699b3ff3acdc03af4bc854ea2943e72c81900d4de63c9390a108b7b",
      "state": "ok",
      "task_id": "T16",
      "unit": "score",
      "value": 0.415,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.295,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.425
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.035,
      "hit_rate_answer_fields": 0.035,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 193,
      "n_hit_free_answer_fields": 193,
      "n_population": 1200,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.33333333333333326,
        "constant_prediction": false,
        "majority_prediction_share": 0.88,
        "majority_value": 0.36
      },
      "results_sha256": "b160634d004080dae805a808122eff52eab302e48d2163ccec7ab5ee17b63a9b",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "8f6560a9a699b3ff3acdc03af4bc854ea2943e72c81900d4de63c9390a108b7b",
      "state": "ok",
      "task_id": "T16",
      "unit": "score",
      "value": 0.36,
      "value_hit_free": 0.35233160621761656,
      "value_hit_free_answer_fields": 0.35233160621761656
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.375,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.515
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 5153,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "7c0fe44ca3892f4fe7183ba5a49cecd4ff64ba9d4abba1adb228bf394ba70b14",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "0e7f9f160515a068d8f2a4df367c0da497884bc3f6cd9d6b6835fcab2b3e7e9a",
      "state": "ok",
      "task_id": "T17",
      "unit": "score",
      "value": 0.445,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.33,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.465
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.0,
      "hit_rate_answer_fields": 0.0,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 200,
      "n_hit_free_answer_fields": 200,
      "n_population": 5153,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "971e414b21515002eff209781a4f85f9fbbb238c3567a89fc382655ecf680f9c",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "0e7f9f160515a068d8f2a4df367c0da497884bc3f6cd9d6b6835fcab2b3e7e9a",
      "state": "ok",
      "task_id": "T17",
      "unit": "score",
      "value": 0.4,
      "value_hit_free": 0.4,
      "value_hit_free_answer_fields": 0.4
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc_norm,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T18",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "clean",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc_norm,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T18",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc_norm,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T19",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc_norm,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T19",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.23,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.36
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc_norm,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 18736,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "297d36c7dfa3612fb6d9af241cc5a9a798913037ea0662339ca79953940f2107",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "7f70b2594a3b84097657ff33afb9c50ee8039883349049423d72efd7b22a7a0d",
      "state": "ok",
      "task_id": "T20",
      "unit": "score",
      "value": 0.295,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.24,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.37
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "clean",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc_norm,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 18736,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "d3651eaa70565fd77f1a7e1b5e1aa65797472522c43d8a36b3f04f205bb45c84",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "7f70b2594a3b84097657ff33afb9c50ee8039883349049423d72efd7b22a7a0d",
      "state": "ok",
      "task_id": "T20",
      "unit": "score",
      "value": 0.305,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.635,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.765
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 872,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": false,
        "majority_prediction_share": 0.755,
        "majority_value": 0.505
      },
      "results_sha256": "ac5d66fcb3238d55198246328baae16800a2e504f66dc8955b36847b8c6e8fb9",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "af3c3c75c608c2dec0e3ee4a324b77b53841a492804bdea7e67e4228bb6f81ba",
      "state": "ok",
      "task_id": "T21",
      "unit": "score",
      "value": 0.7,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.435,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.575
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.01,
      "hit_rate_answer_fields": 0.01,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 198,
      "n_hit_free_answer_fields": 198,
      "n_population": 872,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": true,
        "majority_prediction_share": 1.0,
        "majority_value": 0.505
      },
      "results_sha256": "cf21c35cecba72ed1be00e746fcf527fc93423efed29823640abb66b1b8b0634",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "af3c3c75c608c2dec0e3ee4a324b77b53841a492804bdea7e67e4228bb6f81ba",
      "state": "ok",
      "task_id": "T21",
      "unit": "score",
      "value": 0.505,
      "value_hit_free": 0.5,
      "value_hit_free_answer_fields": 0.5
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.535,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.675
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 722,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": false,
        "majority_prediction_share": 0.63,
        "majority_value": 0.4
      },
      "results_sha256": "b7e1d3c20c6f69f8f4dfc40d100cc3f74f746282f1f61f20b4aea983eea3b7e1",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "69155ab8e0f845beead50b746fb16e0b9ead60229de4fbd6e9013824fd46b3ec",
      "state": "ok",
      "task_id": "T22",
      "unit": "score",
      "value": 0.605,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.26,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.39
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.02,
      "hit_rate_answer_fields": 0.02,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 196,
      "n_hit_free_answer_fields": 196,
      "n_population": 722,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": false,
        "majority_prediction_share": 0.635,
        "majority_value": 0.4
      },
      "results_sha256": "228850f2134265180e7f82b33e7eab1aa1c986f2ad50dbc450d606b6e185faa3",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "69155ab8e0f845beead50b746fb16e0b9ead60229de4fbd6e9013824fd46b3ec",
      "state": "ok",
      "task_id": "T22",
      "unit": "score",
      "value": 0.325,
      "value_hit_free": 0.33163265306122447,
      "value_hit_free_answer_fields": 0.33163265306122447
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.505,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.64
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 4372,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.125,
        "constant_prediction": false,
        "majority_prediction_share": 0.225,
        "majority_value": 0.16
      },
      "results_sha256": "425db72aa75e03d32ac6232d910c8c0907a51cadef207bd7ebc5ad270a8daa7a",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "a599d193eb6aff848ee4b1a64e7fc781b3d617f35c2293319f29b2f10debe0d1",
      "state": "ok",
      "task_id": "T23",
      "unit": "score",
      "value": 0.575,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.385,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.525
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.03,
      "hit_rate_answer_fields": 0.03,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 194,
      "n_hit_free_answer_fields": 194,
      "n_population": 4372,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.125,
        "constant_prediction": false,
        "majority_prediction_share": 0.515,
        "majority_value": 0.16
      },
      "results_sha256": "ae22d8523778e9b2ab6f060a472800f4a43ff32d81d420fd2c9783cf92351e48",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "a599d193eb6aff848ee4b1a64e7fc781b3d617f35c2293319f29b2f10debe0d1",
      "state": "ok",
      "task_id": "T23",
      "unit": "score",
      "value": 0.455,
      "value_hit_free": 0.4536082474226804,
      "value_hit_free_answer_fields": 0.4536082474226804
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.24,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.365
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 900,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": false,
        "majority_prediction_share": 0.665,
        "majority_value": 0.3
      },
      "results_sha256": "e2316eba6c6191492b6aa63959ed7108528c3e2db41bd2c5724e132c7724cc3b",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "7bb3efeaa3ff8382b06c253e6542db63c16f500c4dc685fad1a188f7aa37cb36",
      "state": "ok",
      "task_id": "T24",
      "unit": "score",
      "value": 0.3,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.235,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.365
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "clean",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 900,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": false,
        "majority_prediction_share": 0.395,
        "majority_value": 0.3
      },
      "results_sha256": "aaca6733999a7b6198e172abb5d25d5b36183e38006faaecfb091254c1705875",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "7bb3efeaa3ff8382b06c253e6542db63c16f500c4dc685fad1a188f7aa37cb36",
      "state": "ok",
      "task_id": "T24",
      "unit": "score",
      "value": 0.3,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.17341040462427745,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.33980582524271846
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "f1,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 1029,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": false,
        "majority_prediction_share": 0.81,
        "majority_value": 0.0
      },
      "results_sha256": "443f1bacd8f8fa2c6a885eded4c9acc5fde305ad69ff1a82450e930c82705b43",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "5b74b752db2489e6e12fbc31c5f9e84c95bfae73bcdaf940df20789716c23091",
      "state": "ok",
      "task_id": "T25",
      "unit": "score",
      "value": 0.2553191489361702,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.16666666666666666,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.38596491228070173
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.8,
      "hit_rate_answer_fields": 0.8,
      "metric": "f1,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 40,
      "n_hit_free_answer_fields": 40,
      "n_population": 1029,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": false,
        "majority_prediction_share": 0.59,
        "majority_value": 0.0
      },
      "results_sha256": "f7274ddd1a5c1f5c7e6d110e17b41220179187f10945b740a4531ac199adc4ce",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "5b74b752db2489e6e12fbc31c5f9e84c95bfae73bcdaf940df20789716c23091",
      "state": "ok",
      "task_id": "T25",
      "unit": "score",
      "value": 0.2777777777777778,
      "value_hit_free": 0.4,
      "value_hit_free_answer_fields": 0.4
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.35,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.49
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 1000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.33333333333333326,
        "constant_prediction": false,
        "majority_prediction_share": 0.905,
        "majority_value": 0.41
      },
      "results_sha256": "68b61886b30ff397507e5e3707aaa22fb4a3f6a0a93ccc0dc38ece0458829e13",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "5ca16a5f491004ebd48a5e4a5a58f8dcb7622ad004812ca8ed84288a9d205a52",
      "state": "ok",
      "task_id": "T26",
      "unit": "score",
      "value": 0.42,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.195,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.315
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.28,
      "hit_rate_answer_fields": 0.28,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 144,
      "n_hit_free_answer_fields": 144,
      "n_population": 1000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.33333333333333326,
        "constant_prediction": false,
        "majority_prediction_share": 0.78,
        "majority_value": 0.41
      },
      "results_sha256": "1fa89ad0ddd3e617f9ab9f0c9f42b41bfbb548d4a455228b9380b0a245bb71c5",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "5ca16a5f491004ebd48a5e4a5a58f8dcb7622ad004812ca8ed84288a9d205a52",
      "state": "ok",
      "task_id": "T26",
      "unit": "score",
      "value": 0.255,
      "value_hit_free": 0.2777777777777778,
      "value_hit_free_answer_fields": 0.2777777777777778
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.0,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.0
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "f1,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 1078,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": true,
        "majority_prediction_share": 1.0,
        "majority_value": 0.0
      },
      "results_sha256": "ccfe333023cb4e7f855edaba59842373e8b412a91b98ffcea3e4f321e20e4728",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "1eacd6a6d277d2dab76f9f5b28de389edad6e1db1cfa8d130f315b70d37dc9c7",
      "state": "ok",
      "task_id": "T27",
      "unit": "score",
      "value": 0.0,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.0,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.0
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.0,
      "hit_rate_answer_fields": 0.0,
      "metric": "f1,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 200,
      "n_hit_free_answer_fields": 200,
      "n_population": 1078,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": true,
        "majority_prediction_share": 1.0,
        "majority_value": 0.0
      },
      "results_sha256": "181fd12e7c0953e8acb4a4ff9a6c6bd84db097c7ff99db1abc85d98451323715",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "1eacd6a6d277d2dab76f9f5b28de389edad6e1db1cfa8d130f315b70d37dc9c7",
      "state": "ok",
      "task_id": "T27",
      "unit": "score",
      "value": 0.0,
      "value_hit_free": 0.0,
      "value_hit_free_answer_fields": 0.0
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.17257142857142857,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.2296195652173913
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "f1,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 1000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": true,
        "majority_prediction_share": 1.0,
        "majority_value": 0.1796143250688705
      },
      "results_sha256": "09b3af79137a6b10a72300acda72a7c4b118bebb81eb42e4c6a7b7bb68d98679",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "2771592012d0ffcdb46329d21fd29eef4c2de815b45ab1c1b8b597b2510a5999",
      "state": "ok",
      "task_id": "T28",
      "unit": "score",
      "value": 0.1796143250688705,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.17257142857142857,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.2296195652173913
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.03,
      "hit_rate_answer_fields": 0.03,
      "metric": "f1,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 194,
      "n_hit_free_answer_fields": 194,
      "n_population": 1000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": true,
        "majority_prediction_share": 1.0,
        "majority_value": 0.1796143250688705
      },
      "results_sha256": "cf0bdd1e0eeabbb7ac112edbcd1436632f7b97f3dd5e68959d134e2268c1b0d9",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "2771592012d0ffcdb46329d21fd29eef4c2de815b45ab1c1b8b597b2510a5999",
      "state": "ok",
      "task_id": "T28",
      "unit": "score",
      "value": 0.1796143250688705,
      "value_hit_free": 0.1789173789173789,
      "value_hit_free_answer_fields": 0.1789173789173789
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.39,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.53
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 2058,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.16666666666666663,
        "constant_prediction": false,
        "majority_prediction_share": 0.645,
        "majority_value": 0.4
      },
      "results_sha256": "5cf79411fd7efef677ced342a27de65f06c1b2707e93e84400a9ae7645e2c860",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "4c8f630dd6836f78d765b618c19c61b3196bb5ea92207845a3ed90f1bdbf2e17",
      "state": "ok",
      "task_id": "T29",
      "unit": "score",
      "value": 0.46,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.295,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.425
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.065,
      "hit_rate_answer_fields": 0.065,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 187,
      "n_hit_free_answer_fields": 187,
      "n_population": 2058,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.16666666666666663,
        "constant_prediction": false,
        "majority_prediction_share": 0.74,
        "majority_value": 0.4
      },
      "results_sha256": "db013b473df2b6c545a0f877b8c39c75dbd17543b8f1379c550f7564b765ca5b",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "4c8f630dd6836f78d765b618c19c61b3196bb5ea92207845a3ed90f1bdbf2e17",
      "state": "ok",
      "task_id": "T29",
      "unit": "score",
      "value": 0.36,
      "value_hit_free": 0.3689839572192513,
      "value_hit_free_answer_fields": 0.3689839572192513
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.465,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.605
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "levenshtein,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 5928,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "22e7a8c2c57285918269570db982a9f97516ec280876e58a69c9851cf500a0ba",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "4315be85aabdd27ef0c986e1a461bea8afc6c4cfde4d6b65c7d6149ca3f5232d",
      "state": "ok",
      "task_id": "T30",
      "unit": "score",
      "value": 0.535,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.445,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.585
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.965,
      "hit_rate_answer_fields": 0.005,
      "metric": "levenshtein,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 7,
      "n_hit_free_answer_fields": 199,
      "n_population": 5928,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "5665d6ebaf66ff2109fc8dc3710622925b60255661faaa63be6bcd4285e2e40f",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "4315be85aabdd27ef0c986e1a461bea8afc6c4cfde4d6b65c7d6149ca3f5232d",
      "state": "ok",
      "task_id": "T30",
      "unit": "score",
      "value": 0.515,
      "value_hit_free": 0.14285714285714285,
      "value_hit_free_answer_fields": 0.5175879396984925
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.235,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.36
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "levenshtein,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 5764,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "ee82cea1101b51db90614e3b92b563d5abf3a2476acbc8fd754c4f9721aaebfb",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "4632428c7c6240e6c79251aa573e01c87e7ef25690691327016d6466f9e09ade",
      "state": "ok",
      "task_id": "T31",
      "unit": "score",
      "value": 0.295,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.165,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.28
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 1.0,
      "hit_rate_answer_fields": 0.21,
      "metric": "levenshtein,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 0,
      "n_hit_free_answer_fields": 158,
      "n_population": 5764,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "cfc97f65375c0d2e7c76457d44279946bfec77cda6c8b2076ba0ef8d516ca403",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "4632428c7c6240e6c79251aa573e01c87e7ef25690691327016d6466f9e09ade",
      "state": "ok",
      "task_id": "T31",
      "unit": "score",
      "value": 0.22,
      "value_hit_free": null,
      "value_hit_free_answer_fields": 0.22784810126582278
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.505538317021724,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.6212591970532435
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 14042,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": false,
        "majority_prediction_share": 0.295,
        "majority_value": 0.28
      },
      "results_sha256": "0772e15260978ed97bd54538472e6024412466a4007407cfdd0f8fa1ddcd57f7",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "6ddde1e4519a7233a2918746a5f446717f5d70a5d0d9031655ae3dd63da33d03",
      "state": "ok",
      "task_id": "T32",
      "unit": "score",
      "value": 0.5632786242386014,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.22895617414916666,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.33779775392437406
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.03,
      "hit_rate_answer_fields": 0.03,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 194,
      "n_hit_free_answer_fields": 194,
      "n_population": 14042,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": false,
        "majority_prediction_share": 0.405,
        "majority_value": 0.28
      },
      "results_sha256": "bba8b008a1bb7e6c6e00fa2e24fc0660a006bbf38714c6dd08edac8297e2800d",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "6ddde1e4519a7233a2918746a5f446717f5d70a5d0d9031655ae3dd63da33d03",
      "state": "ok",
      "task_id": "T32",
      "unit": "score",
      "value": 0.2831840440336381,
      "value_hit_free": 0.2935724829222906,
      "value_hit_free_answer_fields": 0.2935724829222906
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.43,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.57
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "exact_match,strict-match",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 1319,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "094eb4a8b53de271f91d10041183ed474d17f649d770ef3a857c5288b2d78ebf",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "fba1895da6a8db6929b9c1cae827cabc98f2ba813e4a4f3cff7857a12ac4a9db",
      "state": "ok",
      "task_id": "T33",
      "unit": "score",
      "value": 0.5,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.015,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.07
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "clean",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "exact_match,strict-match",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 1319,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "11b6515486899a13b8f6c0b7c79df2ba618bac8f9a74e2abf0d2f3a91d47db79",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "fba1895da6a8db6929b9c1cae827cabc98f2ba813e4a4f3cff7857a12ac4a9db",
      "state": "ok",
      "task_id": "T33",
      "unit": "score",
      "value": 0.04,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.0,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.0
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "exact_match,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 5000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "df351a29a6bbc0a1f69aa790bdd2a6a9de35313fd383020a687c493bb6e3bcaa",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "d76e87064d912355ac548e65ef497055c1708fbe759c8766a75f3ac83ebe0116",
      "state": "ok",
      "task_id": "T34",
      "unit": "score",
      "value": 0.0,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.0,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.0
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.02,
      "hit_rate_answer_fields": 0.02,
      "metric": "exact_match,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 196,
      "n_hit_free_answer_fields": 196,
      "n_population": 5000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "6c4606e511a9f2af2585e8c3e24d6f3b94e5b6ac8dd7d963ae0e7d7b5a34b0d5",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "d76e87064d912355ac548e65ef497055c1708fbe759c8766a75f3ac83ebe0116",
      "state": "ok",
      "task_id": "T34",
      "unit": "score",
      "value": 0.0,
      "value_hit_free": 0.0,
      "value_hit_free_answer_fields": 0.0
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.4170048281999294,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.530905281685233
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 817,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "22add0cb7a15606f4e6aa69726d44b46f92a6d577d61b33e2d188726dd424efe",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "93301a192273d3dc8865e3dcbd057c2e0d4c21f4946734a7c908a0a08e0afeb1",
      "state": "ok",
      "task_id": "T35",
      "unit": "score",
      "value": 0.4730409384330242,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.3838642958423616,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.5049122565642408
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.07,
      "hit_rate_answer_fields": 0.07,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 186,
      "n_hit_free_answer_fields": 186,
      "n_population": 817,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "2456ff951abaff77856d564abcb8234740438c3d0dba976c57d4b112fcf280a7",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "93301a192273d3dc8865e3dcbd057c2e0d4c21f4946734a7c908a0a08e0afeb1",
      "state": "ok",
      "task_id": "T35",
      "unit": "score",
      "value": 0.4445515121878845,
      "value_hit_free": 0.4358428387900334,
      "value_hit_free_answer_fields": 0.4358428387900334
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "prompt_level_strict_acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T36",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "prompt_level_strict_acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T36",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.23,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.355
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 18821,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": false,
        "majority_prediction_share": 0.58,
        "majority_value": 0.33
      },
      "results_sha256": "ebcd27776a5fc02060defd76f54723846ad47a73d46ba0ea7a14442b7ba4e253",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "afaf9c14c3bf753305551a43eac9044b6519c296b0d942f47fc0bf46126589a0",
      "state": "ok",
      "task_id": "T37",
      "unit": "score",
      "value": 0.29,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.24,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.365
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.045,
      "hit_rate_answer_fields": 0.045,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 191,
      "n_hit_free_answer_fields": 191,
      "n_population": 18821,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": false,
        "majority_prediction_share": 0.655,
        "majority_value": 0.33
      },
      "results_sha256": "114818a564af652284faf55d7ea1c16cc9d47508260de24f2e0760a4ef541568",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "afaf9c14c3bf753305551a43eac9044b6519c296b0d942f47fc0bf46126589a0",
      "state": "ok",
      "task_id": "T37",
      "unit": "score",
      "value": 0.3,
      "value_hit_free": 0.29842931937172773,
      "value_hit_free_answer_fields": 0.29842931937172773
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 16.506203016316007,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 26.492142050376877
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "f1,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 11873,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "0db861039b901f666f2ddbc7bc3b83887e88ea3b196504bf2d76844f8b224dda",
      "runner_sha256": "e86eef3297f7b06d6bbfe898b51089295d76b0dd1bd338c44401a092c5f62777",
      "source_id_digest": "e5657b168e2c261d949e63361a31d63c3033da08f0928141995e7311a34e3ed0",
      "state": "ok",
      "task_id": "T38",
      "unit": "percent",
      "value": 21.445553515509165,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 3.6476204762255002,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 7.287952243174164
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.24,
      "hit_rate_answer_fields": 0.0,
      "metric": "f1,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 152,
      "n_hit_free_answer_fields": 200,
      "n_population": 11873,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "44c2eabb849a2f1ae87b9ee2f512c86f68712a3b586cd5493284b8437fa74450",
      "runner_sha256": "e86eef3297f7b06d6bbfe898b51089295d76b0dd1bd338c44401a092c5f62777",
      "source_id_digest": "e5657b168e2c261d949e63361a31d63c3033da08f0928141995e7311a34e3ed0",
      "state": "ok",
      "task_id": "T38",
      "unit": "percent",
      "value": 5.3562238471742605,
      "value_hit_free": 5.432371239488894,
      "value_hit_free_answer_fields": 5.3562238471742605
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.03455,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.08659999999999998
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "f1,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 9536,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "34f7dc7e011e41bd854aa02889829bce550616693531dfeb55bee55dbbc97368",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "0c4c8e2a4d4aacf9628a2d34e3ccea3528592d25348621ef86b9606723ff5230",
      "state": "ok",
      "task_id": "T39",
      "unit": "score",
      "value": 0.05874999999999998,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.010049999999999998,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.0315
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.215,
      "hit_rate_answer_fields": 0.0,
      "metric": "f1,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 157,
      "n_hit_free_answer_fields": 200,
      "n_population": 9536,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "28bdce9d0db7286033b13bd49dcc9664fcf99954fcda1db417c83fa7c6d1b0d1",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "0c4c8e2a4d4aacf9628a2d34e3ccea3528592d25348621ef86b9606723ff5230",
      "state": "ok",
      "task_id": "T39",
      "unit": "score",
      "value": 0.019800000000000005,
      "value_hit_free": 0.02089171974522293,
      "value_hit_free_answer_fields": 0.019800000000000005
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.3,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.43
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 1045,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "642d7a629f4d89bd0b24957ac70522b92228e57c2af4fea2c8facfcb4de10439",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "aea2e006a0d1f85819b30a9ec7b9bae4c440d5fcd0c6cf1cdd90ff17121ce846",
      "state": "ok",
      "task_id": "T40",
      "unit": "score",
      "value": 0.365,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.21,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.33
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.06,
      "hit_rate_answer_fields": 0.0,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 188,
      "n_hit_free_answer_fields": 200,
      "n_population": 1045,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "9ea656b54c49116629416b0770dc8e9565addf0781a20e016ef0a06a9c8354af",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "aea2e006a0d1f85819b30a9ec7b9bae4c440d5fcd0c6cf1cdd90ff17121ce846",
      "state": "ok",
      "task_id": "T40",
      "unit": "score",
      "value": 0.27,
      "value_hit_free": 0.2712765957446808,
      "value_hit_free_answer_fields": 0.27
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.5735389344451844,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.6817363146113147
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "f1,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 500,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "b049b999de93cbec5e42aafadeb7fcb367a7e2dcaaab57d748cf81b3586387aa",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "559c0825839f86a7a9ac4ed6d7ba71c7b5ca647307dd92f0bf69ade0f67246fa",
      "state": "ok",
      "task_id": "T41",
      "unit": "score",
      "value": 0.6282142451555351,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.36869991408373765,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.48298872799313974
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.145,
      "hit_rate_answer_fields": 0.13,
      "metric": "f1,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 171,
      "n_hit_free_answer_fields": 174,
      "n_population": 500,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "1f3258492c7bf935aaaf7d1b44e907fca99ec7009ff4ca06fb0fa914609d1827",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "559c0825839f86a7a9ac4ed6d7ba71c7b5ca647307dd92f0bf69ade0f67246fa",
      "state": "ok",
      "task_id": "T41",
      "unit": "score",
      "value": 0.42588400632591805,
      "value_hit_free": 0.4262927797336123,
      "value_hit_free_answer_fields": 0.4242111034546802
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.045,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.12
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "exact_match,remove_whitespace",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 3610,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "149539c68af9bb5670f8a3e08b214d6124d8bb5a201ece10dca8ee84acd31bc8",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "ec49de267492e8a3f520ab4eb858fee702f0cecd891ed6098107ac6125d09ae8",
      "state": "ok",
      "task_id": "T42",
      "unit": "score",
      "value": 0.08,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.0,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.015
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.025,
      "hit_rate_answer_fields": 0.025,
      "metric": "exact_match,remove_whitespace",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 195,
      "n_hit_free_answer_fields": 195,
      "n_population": 3610,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "6e53646f5d5a0d2b5e750de5ba0ad9549c6997d516d69070b6fa3f02377be3b9",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "ec49de267492e8a3f520ab4eb858fee702f0cecd891ed6098107ac6125d09ae8",
      "state": "ok",
      "task_id": "T42",
      "unit": "score",
      "value": 0.005,
      "value_hit_free": 0.005128205128205128,
      "value_hit_free_answer_fields": 0.005128205128205128
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.245,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.375
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "exact_match,remove_whitespace",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 17944,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "cdb3ab12ab3ac90c0f8203f6160054efbcbe1e1bb669d34559e955d4d3ed8aba",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "bedecbef62fad879bc241807f3c27fafa9b346fe5323fe8d229763ff9f1da29e",
      "state": "ok",
      "task_id": "T43",
      "unit": "score",
      "value": 0.31,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.0,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.035
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.06,
      "hit_rate_answer_fields": 0.06,
      "metric": "exact_match,remove_whitespace",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 188,
      "n_hit_free_answer_fields": 188,
      "n_population": 17944,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "47488433a70f44dfc77501c583b0a3ed3b0b36c644c5c42f8214fd06534c6ae1",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "bedecbef62fad879bc241807f3c27fafa9b346fe5323fe8d229763ff9f1da29e",
      "state": "ok",
      "task_id": "T43",
      "unit": "score",
      "value": 0.015,
      "value_hit_free": 0.015957446808510637,
      "value_hit_free_answer_fields": 0.015957446808510637
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T44",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T44",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.34,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.475
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 4183,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": false,
        "majority_prediction_share": 0.325,
        "majority_value": 0.305
      },
      "results_sha256": "a12c9ed811d417e8204271dee963b944f022b7ba86299f9c2716f15e8ab203bd",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "42a9f00f0e8d3f1379b0c2ed8123a7fd7421b1789f19e048da83efdf4e3e5974",
      "state": "ok",
      "task_id": "T45",
      "unit": "score",
      "value": 0.405,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.185,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.305
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.045,
      "hit_rate_answer_fields": 0.045,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 191,
      "n_hit_free_answer_fields": 191,
      "n_population": 4183,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.25,
        "constant_prediction": false,
        "majority_prediction_share": 0.88,
        "majority_value": 0.305
      },
      "results_sha256": "05a7b7f84dff46ca0397722374e1b6c28b1bb15ba3b9fa09de667b9e9a3608aa",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "42a9f00f0e8d3f1379b0c2ed8123a7fd7421b1789f19e048da83efdf4e3e5974",
      "state": "ok",
      "task_id": "T45",
      "unit": "score",
      "value": 0.245,
      "value_hit_free": 0.24083769633507854,
      "value_hit_free_answer_fields": 0.24083769633507854
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T46",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": null,
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": null,
      "n_scored": 0,
      "n_selected": null,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": null,
      "runner_sha256": null,
      "source_id_digest": null,
      "state": "not_run",
      "task_id": "T46",
      "unit": "score",
      "value": null,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.22688563676123188,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.33455068711557734
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "exact_match,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 6511,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "1fff839251972c7ddd6255f840202fce4a8c9a3287765b596cfef0b03e0c2ef2",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "9a04e7f2f5f136a8ee825fa5d5361dfef11caa4f820f27532831751663aba1a4",
      "state": "ok",
      "task_id": "T47",
      "unit": "score",
      "value": 0.2810441670140641,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.24883603571977092,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.35990228989768236
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.005,
      "hit_rate_answer_fields": 0.005,
      "metric": "exact_match,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 199,
      "n_hit_free_answer_fields": 199,
      "n_population": 6511,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": null,
      "results_sha256": "03bd7c99f1d9ebc25c918b204edefc11e55683216e9d7e58b6c49084493acfe5",
      "runner_sha256": "703c2ef4ef693bdce95a12554c1dba2668bc3180c4a4c0835186481568e60367",
      "source_id_digest": "9a04e7f2f5f136a8ee825fa5d5361dfef11caa4f820f27532831751663aba1a4",
      "state": "ok",
      "task_id": "T47",
      "unit": "score",
      "value": 0.30435636395550386,
      "value_hit_free": 0.30641332250916037,
      "value_hit_free_answer_fields": 0.30641332250916037
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.54,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.675
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "acc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 2000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "3f40b74bc0c63160fcd3e0926850649dd213d5cfaa05a3e2ff10907032c6a46f",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "4d23ca378e5e33e57a2de8109efdd79e7307b1ad6c0ee69211e3791b22951acc",
      "state": "ok",
      "task_id": "T48",
      "unit": "score",
      "value": 0.605,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.45,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.59
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.0,
      "hit_rate_answer_fields": 0.0,
      "metric": "acc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 200,
      "n_hit_free_answer_fields": 200,
      "n_population": 2000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": 0.5,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "70eed0d17a57ba64f8820cc18d1db49510da9a3897675d5403630887d9f5db96",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "4d23ca378e5e33e57a2de8109efdd79e7307b1ad6c0ee69211e3791b22951acc",
      "state": "ok",
      "task_id": "T48",
      "unit": "score",
      "value": 0.52,
      "value_hit_free": 0.52,
      "value_hit_free_answer_fields": 0.52
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": -0.1443378222898272,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.13502895642910745
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "mcc,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 1043,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": false,
        "majority_prediction_share": 0.565,
        "majority_value": 0.0
      },
      "results_sha256": "8ace22c45466f17b57036a7a1e2921eefcdb9a0954fa1a6159739f97bacc7193",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "eadab46ac0944d7d2523092310eb29d4ad35d1f616f1e891e6da12a4dd392323",
      "state": "ok",
      "task_id": "T49",
      "unit": "score",
      "value": -0.005921616513673969,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.0,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.0
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.05,
      "hit_rate_answer_fields": 0.05,
      "metric": "mcc,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 190,
      "n_hit_free_answer_fields": 190,
      "n_population": 1043,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": true,
        "majority_prediction_share": 1.0,
        "majority_value": 0.0
      },
      "results_sha256": "1d69a9c7d726c33e19a08ac13d7b9a68e69382c5bd113e51070f88203ed44876",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "eadab46ac0944d7d2523092310eb29d4ad35d1f616f1e891e6da12a4dd392323",
      "state": "ok",
      "task_id": "T49",
      "unit": "score",
      "value": 0.0,
      "value_hit_free": 0.0,
      "value_hit_free_answer_fields": 0.0
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.73,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.8433333333333333
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "unknown",
      "contamination_receipt_sha256": null,
      "higher_is_better": true,
      "hit_rate": null,
      "hit_rate_answer_fields": null,
      "metric": "f1,none",
      "model_id": "LiquidAI/LFM2.5-1.2B-Base",
      "n_hit_free": null,
      "n_hit_free_answer_fields": null,
      "n_population": 10000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "848e591f595c9b19660c11a9040164249fbc495909947c3a65c11b7134eda80a",
      "runner_sha256": "8a8ccfd4c12c40c6a8d7420af941d40643bd76804c7ddf323fda09a713f5b98e",
      "source_id_digest": "1c58ffbef64a26dfb93ade5d03941dd28ad2e446e60bfc1dceac74f677aad687",
      "state": "ok",
      "task_id": "T50",
      "unit": "score",
      "value": 0.7883333333333333,
      "value_hit_free": null,
      "value_hit_free_answer_fields": null
    },
    {
      "ci95": {
        "level": 0.95,
        "lower": 0.5983333333333334,
        "method": "paired_cluster_bootstrap",
        "replicates": 10000,
        "seed": 20261028,
        "upper": 0.73
      },
      "ci95_hit_free": null,
      "ci95_hit_free_answer_fields": null,
      "contamination": "flagged",
      "contamination_receipt_sha256": "bc6c168cf6ef35bce80b70f3bf582853526b558cb7f37e30101e35f3c2342de7",
      "higher_is_better": true,
      "hit_rate": 0.02,
      "hit_rate_answer_fields": 0.005,
      "metric": "f1,none",
      "model_id": "SlayerLab/Slayer1B-1.06",
      "n_hit_free": 196,
      "n_hit_free_answer_fields": 199,
      "n_population": 10000,
      "n_scored": 200,
      "n_selected": 200,
      "protocol_id": "bench50-v0.3.1-panel200",
      "reason": null,
      "reference": {
        "chance_value": null,
        "constant_prediction": null,
        "majority_prediction_share": null,
        "majority_value": null
      },
      "results_sha256": "fa0f39fcc704ecc13207371df0623ef23819e5b77745c118e406cd111063cc90",
      "runner_sha256": "f5ebb45ca18f04f3586159837d071bc33e77e90101bec98146edd7139dd6f37d",
      "source_id_digest": "1c58ffbef64a26dfb93ade5d03941dd28ad2e446e60bfc1dceac74f677aad687",
      "state": "ok",
      "task_id": "T50",
      "unit": "score",
      "value": 0.6633333333333333,
      "value_hit_free": 0.66156462585034,
      "value_hit_free_answer_fields": 0.6616415410385259
    }
  ],
  "protocol_id": "bench50-v0.3.1-panel200",
  "provenance": {
    "amendments": [
      {
        "artifact_sha256": "8527661ff15643cbd87aac357f8373d076a29c3d5e4f88501d8d9c4b3e62727e",
        "base_items_sha256": "1fff839251972c7ddd6255f840202fce4a8c9a3287765b596cfef0b03e0c2ef2",
        "code": "answer_leading_space_strip",
        "model_id": "LiquidAI/LFM2.5-1.2B-Base",
        "task_id": "T47"
      },
      {
        "artifact_sha256": "88f3ce2ab865fffb183d51ae5441da3b2887322ca464e02b0efb04f77505694b",
        "base_items_sha256": "03bd7c99f1d9ebc25c918b204edefc11e55683216e9d7e58b6c49084493acfe5",
        "code": "answer_leading_space_strip",
        "model_id": "SlayerLab/Slayer1B-1.06",
        "task_id": "T47"
      }
    ],
    "contamination_sha256": "64852fdd8877ff1e2b561cffe929339f2fb8f0af1b070cdf9f6a19d372733ab9",
    "contrasts_sha256": "71ae966d5dc5865059b60214c7b9fce1d94c67d178681b8ebc661173cb7c410c",
    "inventory_commit": "18a798843d300fa82cceb50c8ba5032d17be36d5",
    "kind": "measurement",
    "qualification_sha256": "1e3d7aa218b6f9a518824aedd2cdb5f5551d5650f84944dc88a94cdf4bdef86c",
    "references_sha256": "2b9342900bbc4b50dfc593ad6df482c3898bcfd67a60e68587c902a67edbcb54",
    "sources": [
      {
        "results_sha256": "f2561721cdee840e3b1044ca4c02b8e37d04c85ea52a46f76baddc97f7cd27b2",
        "run_id": "run7-slayer-ll-en"
      },
      {
        "results_sha256": "c8710f48a37a155ae90722e469dbeba1032d1ead06db85dfe2fbfa27ff83ab22",
        "run_id": "run7-slayer-ll-pl"
      },
      {
        "results_sha256": "5e47cd22e81bbfd1c6a64ce09cd065736af5000751baca67f494d05067ce7e41",
        "run_id": "run8-lfm-bos-ll-en"
      },
      {
        "results_sha256": "3b7e13142929272c6c4b095c3f29353e08453315bc2d3b8c6799df92500c7c71",
        "run_id": "run8-lfm-bos-ll-pl"
      },
      {
        "results_sha256": "65dd6c459de1996f072793fbf9576a240a5cea517b1f51865b213d9e95d4c6dd",
        "run_id": "run9-lfm-bos-t04-en"
      },
      {
        "results_sha256": "780db12275e2b113b0769d7e830ddfe8391459a5174db2da776a8aad27730959",
        "run_id": "genA-slayer-en"
      },
      {
        "results_sha256": "61d48f6b13b8768c97f16abd2023dfe42ba37402b3797e85db4349cb01a6ef7b",
        "run_id": "genA-slayer-pl"
      },
      {
        "results_sha256": "78da5b851bc37e83c61890f01571cd82419a9b05780717376ec1fd797b894407",
        "run_id": "genB-slayer-en"
      },
      {
        "results_sha256": "c810accb57392072dd8f5384ec8742aeafb8c92d9e32826e169ebf238a496c46",
        "run_id": "genL-lfm-en"
      },
      {
        "results_sha256": "cf163e7ecec0e14a582b20daf24ecfbb2ca8cb256ccde434bea76760376eea90",
        "run_id": "genL-lfm-pl"
      },
      {
        "results_sha256": "dc5f78da8792fe36fb3884d4783f834f911624c630c783c1861e82fb27275bb0",
        "run_id": "sq2S-slayer-en"
      },
      {
        "results_sha256": "e670112e21ee99d55c2b6f090d10749b0157b2945e6e2fa625903e44d5e6d73b",
        "run_id": "sq2L-lfm-en"
      }
    ],
    "superseded": [
      {
        "by_results_sha256": "65dd6c459de1996f072793fbf9576a240a5cea517b1f51865b213d9e95d4c6dd",
        "model_id": "LiquidAI/LFM2.5-1.2B-Base",
        "replaced": [
          {
            "results_sha256": "5e47cd22e81bbfd1c6a64ce09cd065736af5000751baca67f494d05067ce7e41",
            "state": "not_run"
          }
        ],
        "task_id": "T04"
      },
      {
        "by_results_sha256": "cf163e7ecec0e14a582b20daf24ecfbb2ca8cb256ccde434bea76760376eea90",
        "model_id": "LiquidAI/LFM2.5-1.2B-Base",
        "replaced": [
          {
            "results_sha256": "3b7e13142929272c6c4b095c3f29353e08453315bc2d3b8c6799df92500c7c71",
            "state": "not_run"
          }
        ],
        "task_id": "T30"
      },
      {
        "by_results_sha256": "cf163e7ecec0e14a582b20daf24ecfbb2ca8cb256ccde434bea76760376eea90",
        "model_id": "LiquidAI/LFM2.5-1.2B-Base",
        "replaced": [
          {
            "results_sha256": "3b7e13142929272c6c4b095c3f29353e08453315bc2d3b8c6799df92500c7c71",
            "state": "not_run"
          }
        ],
        "task_id": "T31"
      },
      {
        "by_results_sha256": "c810accb57392072dd8f5384ec8742aeafb8c92d9e32826e169ebf238a496c46",
        "model_id": "LiquidAI/LFM2.5-1.2B-Base",
        "replaced": [
          {
            "results_sha256": "5e47cd22e81bbfd1c6a64ce09cd065736af5000751baca67f494d05067ce7e41",
            "state": "not_run"
          }
        ],
        "task_id": "T33"
      },
      {
        "by_results_sha256": "c810accb57392072dd8f5384ec8742aeafb8c92d9e32826e169ebf238a496c46",
        "model_id": "LiquidAI/LFM2.5-1.2B-Base",
        "replaced": [
          {
            "results_sha256": "5e47cd22e81bbfd1c6a64ce09cd065736af5000751baca67f494d05067ce7e41",
            "state": "not_run"
          }
        ],
        "task_id": "T34"
      },
      {
        "by_results_sha256": "e670112e21ee99d55c2b6f090d10749b0157b2945e6e2fa625903e44d5e6d73b",
        "model_id": "LiquidAI/LFM2.5-1.2B-Base",
        "replaced": [
          {
            "results_sha256": "5e47cd22e81bbfd1c6a64ce09cd065736af5000751baca67f494d05067ce7e41",
            "state": "not_run"
          }
        ],
        "task_id": "T38"
      },
      {
        "by_results_sha256": "c810accb57392072dd8f5384ec8742aeafb8c92d9e32826e169ebf238a496c46",
        "model_id": "LiquidAI/LFM2.5-1.2B-Base",
        "replaced": [
          {
            "results_sha256": "5e47cd22e81bbfd1c6a64ce09cd065736af5000751baca67f494d05067ce7e41",
            "state": "not_run"
          }
        ],
        "task_id": "T39"
      },
      {
        "by_results_sha256": "c810accb57392072dd8f5384ec8742aeafb8c92d9e32826e169ebf238a496c46",
        "model_id": "LiquidAI/LFM2.5-1.2B-Base",
        "replaced": [
          {
            "results_sha256": "5e47cd22e81bbfd1c6a64ce09cd065736af5000751baca67f494d05067ce7e41",
            "state": "not_run"
          }
        ],
        "task_id": "T41"
      },
      {
        "by_results_sha256": "c810accb57392072dd8f5384ec8742aeafb8c92d9e32826e169ebf238a496c46",
        "model_id": "LiquidAI/LFM2.5-1.2B-Base",
        "replaced": [
          {
            "results_sha256": "5e47cd22e81bbfd1c6a64ce09cd065736af5000751baca67f494d05067ce7e41",
            "state": "not_run"
          }
        ],
        "task_id": "T42"
      },
      {
        "by_results_sha256": "c810accb57392072dd8f5384ec8742aeafb8c92d9e32826e169ebf238a496c46",
        "model_id": "LiquidAI/LFM2.5-1.2B-Base",
        "replaced": [
          {
            "results_sha256": "5e47cd22e81bbfd1c6a64ce09cd065736af5000751baca67f494d05067ce7e41",
            "state": "not_run"
          }
        ],
        "task_id": "T43"
      },
      {
        "by_results_sha256": "c810accb57392072dd8f5384ec8742aeafb8c92d9e32826e169ebf238a496c46",
        "model_id": "LiquidAI/LFM2.5-1.2B-Base",
        "replaced": [
          {
            "results_sha256": "5e47cd22e81bbfd1c6a64ce09cd065736af5000751baca67f494d05067ce7e41",
            "state": "not_run"
          }
        ],
        "task_id": "T47"
      },
      {
        "by_results_sha256": "61d48f6b13b8768c97f16abd2023dfe42ba37402b3797e85db4349cb01a6ef7b",
        "model_id": "SlayerLab/Slayer1B-1.06",
        "replaced": [
          {
            "results_sha256": "c8710f48a37a155ae90722e469dbeba1032d1ead06db85dfe2fbfa27ff83ab22",
            "state": "not_run"
          }
        ],
        "task_id": "T30"
      },
      {
        "by_results_sha256": "61d48f6b13b8768c97f16abd2023dfe42ba37402b3797e85db4349cb01a6ef7b",
        "model_id": "SlayerLab/Slayer1B-1.06",
        "replaced": [
          {
            "results_sha256": "c8710f48a37a155ae90722e469dbeba1032d1ead06db85dfe2fbfa27ff83ab22",
            "state": "not_run"
          }
        ],
        "task_id": "T31"
      },
      {
        "by_results_sha256": "780db12275e2b113b0769d7e830ddfe8391459a5174db2da776a8aad27730959",
        "model_id": "SlayerLab/Slayer1B-1.06",
        "replaced": [
          {
            "results_sha256": "f2561721cdee840e3b1044ca4c02b8e37d04c85ea52a46f76baddc97f7cd27b2",
            "state": "not_run"
          }
        ],
        "task_id": "T33"
      },
      {
        "by_results_sha256": "78da5b851bc37e83c61890f01571cd82419a9b05780717376ec1fd797b894407",
        "model_id": "SlayerLab/Slayer1B-1.06",
        "replaced": [
          {
            "results_sha256": "f2561721cdee840e3b1044ca4c02b8e37d04c85ea52a46f76baddc97f7cd27b2",
            "state": "not_run"
          }
        ],
        "task_id": "T34"
      },
      {
        "by_results_sha256": "dc5f78da8792fe36fb3884d4783f834f911624c630c783c1861e82fb27275bb0",
        "model_id": "SlayerLab/Slayer1B-1.06",
        "replaced": [
          {
            "results_sha256": "f2561721cdee840e3b1044ca4c02b8e37d04c85ea52a46f76baddc97f7cd27b2",
            "state": "not_run"
          },
          {
            "results_sha256": "78da5b851bc37e83c61890f01571cd82419a9b05780717376ec1fd797b894407",
            "state": "unsupported"
          }
        ],
        "task_id": "T38"
      },
      {
        "by_results_sha256": "78da5b851bc37e83c61890f01571cd82419a9b05780717376ec1fd797b894407",
        "model_id": "SlayerLab/Slayer1B-1.06",
        "replaced": [
          {
            "results_sha256": "f2561721cdee840e3b1044ca4c02b8e37d04c85ea52a46f76baddc97f7cd27b2",
            "state": "not_run"
          }
        ],
        "task_id": "T39"
      },
      {
        "by_results_sha256": "78da5b851bc37e83c61890f01571cd82419a9b05780717376ec1fd797b894407",
        "model_id": "SlayerLab/Slayer1B-1.06",
        "replaced": [
          {
            "results_sha256": "f2561721cdee840e3b1044ca4c02b8e37d04c85ea52a46f76baddc97f7cd27b2",
            "state": "not_run"
          }
        ],
        "task_id": "T41"
      },
      {
        "by_results_sha256": "78da5b851bc37e83c61890f01571cd82419a9b05780717376ec1fd797b894407",
        "model_id": "SlayerLab/Slayer1B-1.06",
        "replaced": [
          {
            "results_sha256": "f2561721cdee840e3b1044ca4c02b8e37d04c85ea52a46f76baddc97f7cd27b2",
            "state": "not_run"
          }
        ],
        "task_id": "T42"
      },
      {
        "by_results_sha256": "78da5b851bc37e83c61890f01571cd82419a9b05780717376ec1fd797b894407",
        "model_id": "SlayerLab/Slayer1B-1.06",
        "replaced": [
          {
            "results_sha256": "f2561721cdee840e3b1044ca4c02b8e37d04c85ea52a46f76baddc97f7cd27b2",
            "state": "not_run"
          }
        ],
        "task_id": "T43"
      },
      {
        "by_results_sha256": "78da5b851bc37e83c61890f01571cd82419a9b05780717376ec1fd797b894407",
        "model_id": "SlayerLab/Slayer1B-1.06",
        "replaced": [
          {
            "results_sha256": "f2561721cdee840e3b1044ca4c02b8e37d04c85ea52a46f76baddc97f7cd27b2",
            "state": "not_run"
          }
        ],
        "task_id": "T47"
      }
    ]
  },
  "run_id": "public-export-a847dabc2eec0483",
  "schema_version": 2,
  "tasks": [
    {
      "family": "blimp",
      "language": "EN",
      "split": "train",
      "task": "blimp",
      "task_id": "T01"
    },
    {
      "family": "arc",
      "language": "EN",
      "split": "test",
      "task": "arc_challenge",
      "task_id": "T02"
    },
    {
      "family": "wikitext2",
      "language": "EN",
      "split": "test",
      "task": "wikitext",
      "task_id": "T03"
    },
    {
      "family": "multiblimp",
      "language": "PL",
      "split": "pol/data.tsv",
      "task": "multiblimp_pl",
      "task_id": "T04"
    },
    {
      "family": "hellaswag",
      "language": "EN",
      "split": "validation",
      "task": "hellaswag",
      "task_id": "T05"
    },
    {
      "family": "piqa",
      "language": "EN",
      "split": "validation",
      "task": "piqa",
      "task_id": "T06"
    },
    {
      "family": "winogrande",
      "language": "EN",
      "split": "validation",
      "task": "winogrande",
      "task_id": "T07"
    },
    {
      "family": "openbookqa",
      "language": "EN",
      "split": "test",
      "task": "openbookqa",
      "task_id": "T08"
    },
    {
      "family": "sciq",
      "language": "EN",
      "split": "test",
      "task": "sciq",
      "task_id": "T09"
    },
    {
      "family": "boolq",
      "language": "EN",
      "split": "validation",
      "task": "boolq",
      "task_id": "T10"
    },
    {
      "family": "copa",
      "language": "EN",
      "split": "validation",
      "task": "copa",
      "task_id": "T11"
    },
    {
      "family": "cb",
      "language": "EN",
      "split": "validation",
      "task": "cb",
      "task_id": "T12"
    },
    {
      "family": "wic",
      "language": "EN",
      "split": "validation",
      "task": "wic",
      "task_id": "T13"
    },
    {
      "family": "wsc273",
      "language": "EN",
      "split": "test",
      "task": "wsc273",
      "task_id": "T14"
    },
    {
      "family": "rte",
      "language": "EN",
      "split": "validation",
      "task": "sglue_rte",
      "task_id": "T15"
    },
    {
      "family": "anli",
      "language": "EN",
      "split": "test_r3",
      "task": "anli_r3",
      "task_id": "T16"
    },
    {
      "family": "lambada",
      "language": "EN",
      "split": "test",
      "task": "lambada_openai",
      "task_id": "T17"
    },
    {
      "family": "logiqa",
      "language": "EN",
      "split": "test",
      "task": "logiqa",
      "task_id": "T18"
    },
    {
      "family": "mathqa",
      "language": "EN",
      "split": "test",
      "task": "mathqa",
      "task_id": "T19"
    },
    {
      "family": "prost",
      "language": "EN",
      "split": "test",
      "task": "prost",
      "task_id": "T20"
    },
    {
      "family": "sst2",
      "language": "EN",
      "split": "validation",
      "task": "sst2",
      "task_id": "T21"
    },
    {
      "family": "polemo2",
      "language": "PL",
      "split": "test",
      "task": "polemo2_in_multiple_choice",
      "task_id": "T22"
    },
    {
      "family": "8tags",
      "language": "PL",
      "split": "test",
      "task": "polish_8tags_multiple_choice",
      "task_id": "T23"
    },
    {
      "family": "belebele",
      "language": "PL",
      "split": "test",
      "task": "polish_belebele_mc",
      "task_id": "T24"
    },
    {
      "family": "dyk",
      "language": "PL",
      "split": "test",
      "task": "polish_dyk_multiple_choice",
      "task_id": "T25"
    },
    {
      "family": "ppc",
      "language": "PL",
      "split": "test",
      "task": "polish_ppc_multiple_choice",
      "task_id": "T26"
    },
    {
      "family": "psc",
      "language": "PL",
      "split": "test",
      "task": "polish_psc_multiple_choice",
      "task_id": "T27"
    },
    {
      "family": "cbd",
      "language": "PL",
      "split": "test",
      "task": "polish_cbd_multiple_choice",
      "task_id": "T28"
    },
    {
      "family": "klej_ner",
      "language": "PL",
      "split": "test",
      "task": "polish_klej_ner_multiple_choice",
      "task_id": "T29"
    },
    {
      "family": "polqa",
      "language": "PL",
      "split": "test",
      "task": "polish_polqa_open_book",
      "task_id": "T30"
    },
    {
      "family": "poquad",
      "language": "PL",
      "split": "validation",
      "task": "polish_poquad_open_book",
      "task_id": "T31"
    },
    {
      "family": "mmlu",
      "language": "EN",
      "split": "test",
      "task": "mmlu",
      "task_id": "T32"
    },
    {
      "family": "gsm8k",
      "language": "EN",
      "split": "test",
      "task": "gsm8k",
      "task_id": "T33"
    },
    {
      "family": "math",
      "language": "EN",
      "split": "test",
      "task": "hendrycks_math",
      "task_id": "T34"
    },
    {
      "family": "truthfulqa",
      "language": "EN",
      "split": "validation",
      "task": "truthfulqa_mc2",
      "task_id": "T35"
    },
    {
      "family": "ifeval",
      "language": "EN",
      "split": "train",
      "task": "ifeval",
      "task_id": "T36"
    },
    {
      "family": "llmzszl",
      "language": "PL",
      "split": "test",
      "task": "n05_llmzszl",
      "task_id": "T37"
    },
    {
      "family": "squad2",
      "language": "EN",
      "split": "validation",
      "task": "squadv2",
      "task_id": "T38"
    },
    {
      "family": "drop",
      "language": "EN",
      "split": "validation",
      "task": "drop",
      "task_id": "T39"
    },
    {
      "family": "race",
      "language": "EN",
      "split": "test",
      "task": "race",
      "task_id": "T40"
    },
    {
      "family": "coqa",
      "language": "EN",
      "split": "validation",
      "task": "coqa",
      "task_id": "T41"
    },
    {
      "family": "nq_open",
      "language": "EN",
      "split": "validation",
      "task": "nq_open",
      "task_id": "T42"
    },
    {
      "family": "triviaqa",
      "language": "EN",
      "split": "validation",
      "task": "triviaqa",
      "task_id": "T43"
    },
    {
      "family": "pubmedqa",
      "language": "EN",
      "split": "test",
      "task": "pubmedqa",
      "task_id": "T44"
    },
    {
      "family": "medmcqa",
      "language": "EN",
      "split": "validation",
      "task": "medmcqa",
      "task_id": "T45"
    },
    {
      "family": "social_iqa",
      "language": "EN",
      "split": "validation",
      "task": "social_iqa",
      "task_id": "T46"
    },
    {
      "family": "bbh",
      "language": "EN",
      "split": "test",
      "task": "bbh_fewshot",
      "task_id": "T47"
    },
    {
      "family": "paws",
      "language": "EN",
      "split": "test",
      "task": "paws_en",
      "task_id": "T48"
    },
    {
      "family": "cola",
      "language": "EN",
      "split": "validation",
      "task": "cola",
      "task_id": "T49"
    },
    {
      "family": "record",
      "language": "EN",
      "split": "validation",
      "task": "record",
      "task_id": "T50"
    }
  ]
}