validation_dict.json

{
    "empathielabs/mathinstruct_judge": {
        "list": [
            0.024569029600370856,
            0.5093165751274039,
            0.541806306047895,
            0.5638397473387602,
            0.5654554652549871,
            0.563291810640717,
            0.5717071706928858,
            0.5753602382976835,
            0.5739919801007926,
            0.5791001284298131,
            0.5783052380613195,
            0.5758259909848372,
            0.5791701558886505,
            0.5808276047188092,
            0.5777370982114377,
            0.5739802787277272,
            0.574000442888194,
            0.5825578023989996,
            0.5860628457936814,
            0.5799481567146817,
            0.5790629264118711,
            0.5852597256864959,
            0.5818849720000252,
            0.5854470470995923,
            0.5818715188016251,
            0.5835347656069733,
            0.5827314207466637,
            0.5776059131312177,
            0.5820925660007368,
            0.5811365405415616,
            0.579283438853132,
            0.5775143142941037,
            0.58587280928604,
            0.584526265727553,
            0.5795885451804331,
            0.5936778717681886,
            0.5897074193246965,
            0.5834238509094812,
            0.5802926459205829,
            0.5907918657532067
        ]
    }
}