Delete evals/mmlu_pro_test_20.json
Browse files- evals/mmlu_pro_test_20.json +0 -181
evals/mmlu_pro_test_20.json
DELETED
|
@@ -1,181 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"benchmark": "MMLU-Pro",
|
| 3 |
-
"dataset": "TIGER-Lab/MMLU-Pro",
|
| 4 |
-
"split": "test",
|
| 5 |
-
"checkpoint": "/artifacts/20260518-164838/full-safetensors",
|
| 6 |
-
"samples_evaluated": 20,
|
| 7 |
-
"accuracy": 0.75,
|
| 8 |
-
"correct": 15,
|
| 9 |
-
"total": 20,
|
| 10 |
-
"category_scores": {
|
| 11 |
-
"business": {
|
| 12 |
-
"accuracy": 0.75,
|
| 13 |
-
"correct": 15,
|
| 14 |
-
"total": 20
|
| 15 |
-
}
|
| 16 |
-
},
|
| 17 |
-
"samples": [
|
| 18 |
-
{
|
| 19 |
-
"question_id": 70,
|
| 20 |
-
"category": "business",
|
| 21 |
-
"gold": "I",
|
| 22 |
-
"prediction": "I",
|
| 23 |
-
"raw_generation": "I",
|
| 24 |
-
"correct": true
|
| 25 |
-
},
|
| 26 |
-
{
|
| 27 |
-
"question_id": 71,
|
| 28 |
-
"category": "business",
|
| 29 |
-
"gold": "F",
|
| 30 |
-
"prediction": "F",
|
| 31 |
-
"raw_generation": "F",
|
| 32 |
-
"correct": true
|
| 33 |
-
},
|
| 34 |
-
{
|
| 35 |
-
"question_id": 72,
|
| 36 |
-
"category": "business",
|
| 37 |
-
"gold": "J",
|
| 38 |
-
"prediction": "J",
|
| 39 |
-
"raw_generation": "J",
|
| 40 |
-
"correct": true
|
| 41 |
-
},
|
| 42 |
-
{
|
| 43 |
-
"question_id": 73,
|
| 44 |
-
"category": "business",
|
| 45 |
-
"gold": "C",
|
| 46 |
-
"prediction": "H",
|
| 47 |
-
"raw_generation": "H",
|
| 48 |
-
"correct": false
|
| 49 |
-
},
|
| 50 |
-
{
|
| 51 |
-
"question_id": 74,
|
| 52 |
-
"category": "business",
|
| 53 |
-
"gold": "G",
|
| 54 |
-
"prediction": "G",
|
| 55 |
-
"raw_generation": "G",
|
| 56 |
-
"correct": true
|
| 57 |
-
},
|
| 58 |
-
{
|
| 59 |
-
"question_id": 75,
|
| 60 |
-
"category": "business",
|
| 61 |
-
"gold": "A",
|
| 62 |
-
"prediction": "A",
|
| 63 |
-
"raw_generation": "A",
|
| 64 |
-
"correct": true
|
| 65 |
-
},
|
| 66 |
-
{
|
| 67 |
-
"question_id": 76,
|
| 68 |
-
"category": "business",
|
| 69 |
-
"gold": "D",
|
| 70 |
-
"prediction": "D",
|
| 71 |
-
"raw_generation": "D",
|
| 72 |
-
"correct": true
|
| 73 |
-
},
|
| 74 |
-
{
|
| 75 |
-
"question_id": 77,
|
| 76 |
-
"category": "business",
|
| 77 |
-
"gold": "J",
|
| 78 |
-
"prediction": "J",
|
| 79 |
-
"raw_generation": "J",
|
| 80 |
-
"correct": true
|
| 81 |
-
},
|
| 82 |
-
{
|
| 83 |
-
"question_id": 78,
|
| 84 |
-
"category": "business",
|
| 85 |
-
"gold": "E",
|
| 86 |
-
"prediction": "E",
|
| 87 |
-
"raw_generation": "E",
|
| 88 |
-
"correct": true
|
| 89 |
-
},
|
| 90 |
-
{
|
| 91 |
-
"question_id": 79,
|
| 92 |
-
"category": "business",
|
| 93 |
-
"gold": "F",
|
| 94 |
-
"prediction": "F",
|
| 95 |
-
"raw_generation": "F",
|
| 96 |
-
"correct": true
|
| 97 |
-
},
|
| 98 |
-
{
|
| 99 |
-
"question_id": 80,
|
| 100 |
-
"category": "business",
|
| 101 |
-
"gold": "E",
|
| 102 |
-
"prediction": "E",
|
| 103 |
-
"raw_generation": "E",
|
| 104 |
-
"correct": true
|
| 105 |
-
},
|
| 106 |
-
{
|
| 107 |
-
"question_id": 81,
|
| 108 |
-
"category": "business",
|
| 109 |
-
"gold": "G",
|
| 110 |
-
"prediction": "C",
|
| 111 |
-
"raw_generation": "C",
|
| 112 |
-
"correct": false
|
| 113 |
-
},
|
| 114 |
-
{
|
| 115 |
-
"question_id": 82,
|
| 116 |
-
"category": "business",
|
| 117 |
-
"gold": "J",
|
| 118 |
-
"prediction": "J",
|
| 119 |
-
"raw_generation": "J",
|
| 120 |
-
"correct": true
|
| 121 |
-
},
|
| 122 |
-
{
|
| 123 |
-
"question_id": 83,
|
| 124 |
-
"category": "business",
|
| 125 |
-
"gold": "D",
|
| 126 |
-
"prediction": "B",
|
| 127 |
-
"raw_generation": "B",
|
| 128 |
-
"correct": false
|
| 129 |
-
},
|
| 130 |
-
{
|
| 131 |
-
"question_id": 84,
|
| 132 |
-
"category": "business",
|
| 133 |
-
"gold": "F",
|
| 134 |
-
"prediction": "F",
|
| 135 |
-
"raw_generation": "F",
|
| 136 |
-
"correct": true
|
| 137 |
-
},
|
| 138 |
-
{
|
| 139 |
-
"question_id": 85,
|
| 140 |
-
"category": "business",
|
| 141 |
-
"gold": "J",
|
| 142 |
-
"prediction": "J",
|
| 143 |
-
"raw_generation": "J",
|
| 144 |
-
"correct": true
|
| 145 |
-
},
|
| 146 |
-
{
|
| 147 |
-
"question_id": 86,
|
| 148 |
-
"category": "business",
|
| 149 |
-
"gold": "F",
|
| 150 |
-
"prediction": "I",
|
| 151 |
-
"raw_generation": "I",
|
| 152 |
-
"correct": false
|
| 153 |
-
},
|
| 154 |
-
{
|
| 155 |
-
"question_id": 87,
|
| 156 |
-
"category": "business",
|
| 157 |
-
"gold": "A",
|
| 158 |
-
"prediction": "D",
|
| 159 |
-
"raw_generation": "D",
|
| 160 |
-
"correct": false
|
| 161 |
-
},
|
| 162 |
-
{
|
| 163 |
-
"question_id": 88,
|
| 164 |
-
"category": "business",
|
| 165 |
-
"gold": "B",
|
| 166 |
-
"prediction": "B",
|
| 167 |
-
"raw_generation": "B",
|
| 168 |
-
"correct": true
|
| 169 |
-
},
|
| 170 |
-
{
|
| 171 |
-
"question_id": 89,
|
| 172 |
-
"category": "business",
|
| 173 |
-
"gold": "D",
|
| 174 |
-
"prediction": "D",
|
| 175 |
-
"raw_generation": "D",
|
| 176 |
-
"correct": true
|
| 177 |
-
}
|
| 178 |
-
],
|
| 179 |
-
"attention_impl": "sdpa",
|
| 180 |
-
"elapsed_seconds": 51.712759704999996
|
| 181 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|