validation_dict.json

{
    "empathielabs/mathinstruct_judge": {
        "list": [
            0.012356160871901766,
            0.15461972547369457,
            0.48580542269155264,
            0.5262637506231544,
            0.5450907519617216,
            0.5426690585971847,
            0.5635065359404174,
            0.5506789258703952,
            0.574924521990181,
            0.5758008317098142,
            0.5712014323448747,
            0.5786318865430549,
            0.572295540765049,
            0.5747123874602763,
            0.571242627980021,
            0.5756114024214628,
            0.5732587800944239,
            0.5786399673519096,
            0.5806065718847804,
            0.5607240720069021,
            0.5822894774225912,
            0.5767564823654124,
            0.572368123430789,
            0.5715245728691419,
            0.5787839569757139,
            0.5740101795552708,
            0.5731954331805066,
            0.5769944259124559,
            0.5758143528690183,
            0.5825809721357939,
            0.5823101227300439,
            0.5773794875820962,
            0.5843074923244919,
            0.5768774442314133,
            0.5872149866468054,
            0.5804339090195613,
            0.578890068423215,
            0.5796248001054051,
            0.5825587657408986,
            0.57715393515016,
            0.5759083093848171,
            0.5766008513972042,
            0.5843901081358999,
            0.5800604047026576,
            0.5917740202046992
        ]
    }
}