validation_dict.json

{
    "empathielabs/mathinstruct_judge": {
        "list": [
            0.22320180573152054,
            0.48842203011881946,
            0.5435712020280961,
            0.5727889217646079,
            0.5885552823294937
        ]
    },
    "empathielabs/imdb": {
        "list": [
            0.635898318529129,
            0.688334094516933,
            0.7036623126752675,
            0.7129036727994681,
            0.7191759207546711
        ]
    }
}