validation_dict.json

{
    "empathielabs/mathinstruct_judge": {
        "list": [
            0.6532731741818228,
            0.6827826897136685,
            0.7001163519805342,
            0.7057138594272508,
            0.7314414375647146,
            0.7392402954598344,
            0.7466427053641944,
            0.7500713458150338,
            0.747936533652511,
            0.7619898544563874,
            0.7662807599515503,
            0.7680152728934999,
            0.7706196256230843,
            0.7715489611023946,
            0.7773229197458916,
            0.7773043717346646,
            0.7733321095010403,
            0.7775986327028765,
            0.7800441830653727,
            0.7837579138868496,
            0.7801122901658338,
            0.7837344653328987,
            0.786501516727352,
            0.7871143792930052,
            0.7866521672769715
        ]
    }
}