small (1).json

[
    {
        "model": "OPT (1.3B)",
        "Average": 7.84,
        "MMLU": 7.4,
        "WinoGrande": 12.47,
        "PiQA": 4.45,
        "CommonsenseQA": 7.61,
        "Race": 13.61,
        "MedMCQA": 1.25,
        "OpenkookQA": 4.48
    },
    {
        "model": "SlimPajama",
        "Average": 9.54,
        "MMLU": 9.22,
        "WinoGrande": 14.76,
        "PiQA": 5.32,
        "CommonsenseQA": 9.01,
        "Race": 16.19,
        "MedMCQA": 1.68,
        "OpenkookQA": 5.7
    },
    {
        "model": "OLMo (1B)",
        "Average": 8.8,
        "MMLU": 8.54,
        "WinoGrande": 6.16,
        "PiQA": 8.05,
        "CommonsenseQA": 13.1,
        "Race": 13.61,
        "MedMCQA": 2.1,
        "OpenkookQA": 6.11
    },
    {
        "model": "GPT-Neo (1.3B)",
        "Average": 7.38,
        "MMLU": 6.94,
        "WinoGrande": 10.81,
        "PiQA": 4.31,
        "CommonsenseQA": 6.34,
        "Race": 13.75,
        "MedMCQA": 2.63,
        "OpenkookQA": 4.89
    },
    {
        "model": "Cerebras-GPT (1.3B)",
        "Average": 4.84,
        "MMLU": 5.37,
        "WinoGrande": 9.31,
        "PiQA": 2.16,
        "CommonsenseQA": 6.2,
        "Race": 6.9,
        "MedMCQA": 1.04,
        "OpenkookQA": 3.46
    },
    {
        "model": "RedPajama (1B)",
        "Average": 9.01,
        "MMLU": 9.21,
        "WinoGrande": 16.97,
        "PiQA": 1.39,
        "CommonsenseQA": 11.41,
        "Race": 14.35,
        "MedMCQA": 1.86,
        "OpenkookQA": 3.87
    },
    {
        "model": "Pythia (1.4B)",
        "Average": 8.73,
        "MMLU": 9.66,
        "WinoGrande": 11.52,
        "PiQA": 4.17,
        "CommonsenseQA": 9.01,
        "Race": 12.76,
        "MedMCQA": 3.19,
        "OpenkookQA": 5.3
    },
    {
        "model": "TinyLLama (1.1B)",
        "Average": 8.39,
        "MMLU": 8.94,
        "WinoGrande": 12.23,
        "PiQA": 3.59,
        "CommonsenseQA": 6.06,
        "Race": 16.7,
        "MedMCQA": 2.07,
        "OpenkookQA": 4.68
    },
    {
        "model": "OELM (1B)",
        "Average": 8.99,
        "MMLU": 9.03,
        "WinoGrande": 10.18,
        "PiQA": 9.05,
        "CommonsenseQA": 7.75,
        "Race": 12.78,
        "MedMCQA": 2.5,
        "OpenkookQA": 6.31
    },
    {
        "model": "Phi-3-mini-128k-instruct  (3.8B)",
        "Average": 39.73,
        "MMLU": 36.97,
        "WinoGrande": 46.88,
        "PiQA": 32.04,
        "CommonsenseQA": 49.15,
        "Race": 37.81,
        "MedMCQA": 22.61,
        "OpenkookQA": 33.6
    },
    {
        "model": "Gemma (2B)",
        "Average": 17.37,
        "MMLU": 17.52,
        "WinoGrande": 22.68,
        "PiQA": 15.09,
        "CommonsenseQA": 27.46,
        "Race": 14.32,
        "MedMCQA": 4.57,
        "OpenkookQA": 14.26
    },
    {
        "model": "Qwen (1.8B)",
        "Average": 21.61,
        "MMLU": 10.0,
        "WinoGrande": 40.97,
        "PiQA": 15.52,
        "CommonsenseQA": 31.13,
        "Race": 34.91,
        "MedMCQA": 4.7,
        "OpenkookQA": 20.37
    }
]