[ { "Model": "llama-3.1-8b", "CR": "7.5", "S-Acc": "1.2", "EM": "0.0", "PM-0.5": "0.0", "Tokens": "2782" }, { "Model": "llama-3.1-70b", "CR": "69.5", "S-Acc": "24.2", "EM": "1.0", "PM-0.5": "17.5", "Tokens": "1939" }, { "Model": "llama-3.3-70b", "CR": "93.5", "S-Acc": "34.8", "EM": "7.0", "PM-0.5": "22.5", "Tokens": "1061" }, { "Model": "mistral-7b", "CR": "84.0", "S-Acc": "11.9", "EM": "0.0", "PM-0.5": "1.5", "Tokens": "3108" }, { "Model": "mistral-small-22b", "CR": "89.0", "S-Acc": "20.5", "EM": "0.5", "PM-0.5": "7.5", "Tokens": "1968" }, { "Model": "mistral-large-123b", "CR": "85.5", "S-Acc": "39.5", "EM": "10.0", "PM-0.5": "33.5", "Tokens": "1955" }, { "Model": "qwen-2.5-7b", "CR": "94.5", "S-Acc": "30.2", "EM": "1.5", "PM-0.5": "15.0", "Tokens": "1486" }, { "Model": "qwen-2.5-32b", "CR": "100.0", "S-Acc": "42.8", "EM": "3.5", "PM-0.5": "30.5", "Tokens": "1201" }, { "Model": "qwen-2.5-72b", "CR": "97.5", "S-Acc": "43.0", "EM": "5.5", "PM-0.5": "34.0", "Tokens": "2013" }, { "Model": "qwq-32b", "CR": "54.5", "S-Acc": "40.1", "EM": "31.5", "PM-0.5": "35.5", "Tokens": "8381" }, { "Model": "deepseek-R1", "CR": "100.0", "S-Acc": "70.3", "EM": "50.0", "PM-0.5": "64.0", "Tokens": "8276" }, { "Model": "gemini-2.0-exp", "CR": "93.0", "S-Acc": "45.3", "EM": "12.5", "PM-0.5": "37.5", "Tokens": "2842" }, { "Model": "gemini-2.0-thinking", "CR": "79.5", "S-Acc": "46.5", "EM": "16.5", "PM-0.5": "41.0", "Tokens": "3852" }, { "Model": "openai-gpt-4o", "CR": "100.0", "S-Acc": "52.2", "EM": "14.5", "PM-0.5": "48.0", "Tokens": "1103" }, { "Model": "openai-o1-mini", "CR": "99.0", "S-Acc": "53.4", "EM": "27.0", "PM-0.5": "43.0", "Tokens": "3960" }, { "Model": "openai-o1-preview", "CR": "91.5", "S-Acc": "65.1", "EM": "50.0", "PM-0.5": "55.5", "Tokens": "8061" } ]