Delete evals/mmlu_pro_test_1000.json
Browse files- evals/mmlu_pro_test_1000.json +0 -267
evals/mmlu_pro_test_1000.json
DELETED
|
@@ -1,267 +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": 1000,
|
| 7 |
-
"batch_size": 8,
|
| 8 |
-
"accuracy": 0.385,
|
| 9 |
-
"correct": 385,
|
| 10 |
-
"total": 1000,
|
| 11 |
-
"category_scores": {
|
| 12 |
-
"business": {
|
| 13 |
-
"accuracy": 0.37515842839036756,
|
| 14 |
-
"correct": 296,
|
| 15 |
-
"total": 789
|
| 16 |
-
},
|
| 17 |
-
"law": {
|
| 18 |
-
"accuracy": 0.4218009478672986,
|
| 19 |
-
"correct": 89,
|
| 20 |
-
"total": 211
|
| 21 |
-
}
|
| 22 |
-
},
|
| 23 |
-
"samples": [
|
| 24 |
-
{
|
| 25 |
-
"question_id": 70,
|
| 26 |
-
"category": "business",
|
| 27 |
-
"gold": "I",
|
| 28 |
-
"prediction": "I",
|
| 29 |
-
"raw_generation": "I",
|
| 30 |
-
"correct": true
|
| 31 |
-
},
|
| 32 |
-
{
|
| 33 |
-
"question_id": 71,
|
| 34 |
-
"category": "business",
|
| 35 |
-
"gold": "F",
|
| 36 |
-
"prediction": "F",
|
| 37 |
-
"raw_generation": "F",
|
| 38 |
-
"correct": true
|
| 39 |
-
},
|
| 40 |
-
{
|
| 41 |
-
"question_id": 72,
|
| 42 |
-
"category": "business",
|
| 43 |
-
"gold": "J",
|
| 44 |
-
"prediction": "J",
|
| 45 |
-
"raw_generation": "J",
|
| 46 |
-
"correct": true
|
| 47 |
-
},
|
| 48 |
-
{
|
| 49 |
-
"question_id": 73,
|
| 50 |
-
"category": "business",
|
| 51 |
-
"gold": "C",
|
| 52 |
-
"prediction": "H",
|
| 53 |
-
"raw_generation": "H",
|
| 54 |
-
"correct": false
|
| 55 |
-
},
|
| 56 |
-
{
|
| 57 |
-
"question_id": 74,
|
| 58 |
-
"category": "business",
|
| 59 |
-
"gold": "G",
|
| 60 |
-
"prediction": "G",
|
| 61 |
-
"raw_generation": "G",
|
| 62 |
-
"correct": true
|
| 63 |
-
},
|
| 64 |
-
{
|
| 65 |
-
"question_id": 75,
|
| 66 |
-
"category": "business",
|
| 67 |
-
"gold": "A",
|
| 68 |
-
"prediction": "A",
|
| 69 |
-
"raw_generation": "A",
|
| 70 |
-
"correct": true
|
| 71 |
-
},
|
| 72 |
-
{
|
| 73 |
-
"question_id": 76,
|
| 74 |
-
"category": "business",
|
| 75 |
-
"gold": "D",
|
| 76 |
-
"prediction": "D",
|
| 77 |
-
"raw_generation": "D",
|
| 78 |
-
"correct": true
|
| 79 |
-
},
|
| 80 |
-
{
|
| 81 |
-
"question_id": 77,
|
| 82 |
-
"category": "business",
|
| 83 |
-
"gold": "J",
|
| 84 |
-
"prediction": "J",
|
| 85 |
-
"raw_generation": "J",
|
| 86 |
-
"correct": true
|
| 87 |
-
},
|
| 88 |
-
{
|
| 89 |
-
"question_id": 78,
|
| 90 |
-
"category": "business",
|
| 91 |
-
"gold": "E",
|
| 92 |
-
"prediction": "E",
|
| 93 |
-
"raw_generation": "E",
|
| 94 |
-
"correct": true
|
| 95 |
-
},
|
| 96 |
-
{
|
| 97 |
-
"question_id": 79,
|
| 98 |
-
"category": "business",
|
| 99 |
-
"gold": "F",
|
| 100 |
-
"prediction": "F",
|
| 101 |
-
"raw_generation": "F",
|
| 102 |
-
"correct": true
|
| 103 |
-
},
|
| 104 |
-
{
|
| 105 |
-
"question_id": 80,
|
| 106 |
-
"category": "business",
|
| 107 |
-
"gold": "E",
|
| 108 |
-
"prediction": "E",
|
| 109 |
-
"raw_generation": "E",
|
| 110 |
-
"correct": true
|
| 111 |
-
},
|
| 112 |
-
{
|
| 113 |
-
"question_id": 81,
|
| 114 |
-
"category": "business",
|
| 115 |
-
"gold": "G",
|
| 116 |
-
"prediction": "C",
|
| 117 |
-
"raw_generation": "C",
|
| 118 |
-
"correct": false
|
| 119 |
-
},
|
| 120 |
-
{
|
| 121 |
-
"question_id": 82,
|
| 122 |
-
"category": "business",
|
| 123 |
-
"gold": "J",
|
| 124 |
-
"prediction": "J",
|
| 125 |
-
"raw_generation": "J",
|
| 126 |
-
"correct": true
|
| 127 |
-
},
|
| 128 |
-
{
|
| 129 |
-
"question_id": 83,
|
| 130 |
-
"category": "business",
|
| 131 |
-
"gold": "D",
|
| 132 |
-
"prediction": "B",
|
| 133 |
-
"raw_generation": "B",
|
| 134 |
-
"correct": false
|
| 135 |
-
},
|
| 136 |
-
{
|
| 137 |
-
"question_id": 84,
|
| 138 |
-
"category": "business",
|
| 139 |
-
"gold": "F",
|
| 140 |
-
"prediction": "F",
|
| 141 |
-
"raw_generation": "F",
|
| 142 |
-
"correct": true
|
| 143 |
-
},
|
| 144 |
-
{
|
| 145 |
-
"question_id": 85,
|
| 146 |
-
"category": "business",
|
| 147 |
-
"gold": "J",
|
| 148 |
-
"prediction": "J",
|
| 149 |
-
"raw_generation": "J",
|
| 150 |
-
"correct": true
|
| 151 |
-
},
|
| 152 |
-
{
|
| 153 |
-
"question_id": 86,
|
| 154 |
-
"category": "business",
|
| 155 |
-
"gold": "F",
|
| 156 |
-
"prediction": "I",
|
| 157 |
-
"raw_generation": "I",
|
| 158 |
-
"correct": false
|
| 159 |
-
},
|
| 160 |
-
{
|
| 161 |
-
"question_id": 87,
|
| 162 |
-
"category": "business",
|
| 163 |
-
"gold": "A",
|
| 164 |
-
"prediction": "D",
|
| 165 |
-
"raw_generation": "D",
|
| 166 |
-
"correct": false
|
| 167 |
-
},
|
| 168 |
-
{
|
| 169 |
-
"question_id": 88,
|
| 170 |
-
"category": "business",
|
| 171 |
-
"gold": "B",
|
| 172 |
-
"prediction": "B",
|
| 173 |
-
"raw_generation": "B",
|
| 174 |
-
"correct": true
|
| 175 |
-
},
|
| 176 |
-
{
|
| 177 |
-
"question_id": 89,
|
| 178 |
-
"category": "business",
|
| 179 |
-
"gold": "D",
|
| 180 |
-
"prediction": "D",
|
| 181 |
-
"raw_generation": "D",
|
| 182 |
-
"correct": true
|
| 183 |
-
},
|
| 184 |
-
{
|
| 185 |
-
"question_id": 90,
|
| 186 |
-
"category": "business",
|
| 187 |
-
"gold": "J",
|
| 188 |
-
"prediction": "C",
|
| 189 |
-
"raw_generation": "C",
|
| 190 |
-
"correct": false
|
| 191 |
-
},
|
| 192 |
-
{
|
| 193 |
-
"question_id": 91,
|
| 194 |
-
"category": "business",
|
| 195 |
-
"gold": "I",
|
| 196 |
-
"prediction": "B",
|
| 197 |
-
"raw_generation": "B",
|
| 198 |
-
"correct": false
|
| 199 |
-
},
|
| 200 |
-
{
|
| 201 |
-
"question_id": 92,
|
| 202 |
-
"category": "business",
|
| 203 |
-
"gold": "E",
|
| 204 |
-
"prediction": "A",
|
| 205 |
-
"raw_generation": "A",
|
| 206 |
-
"correct": false
|
| 207 |
-
},
|
| 208 |
-
{
|
| 209 |
-
"question_id": 93,
|
| 210 |
-
"category": "business",
|
| 211 |
-
"gold": "H",
|
| 212 |
-
"prediction": "C",
|
| 213 |
-
"raw_generation": "C",
|
| 214 |
-
"correct": false
|
| 215 |
-
},
|
| 216 |
-
{
|
| 217 |
-
"question_id": 94,
|
| 218 |
-
"category": "business",
|
| 219 |
-
"gold": "J",
|
| 220 |
-
"prediction": "I",
|
| 221 |
-
"raw_generation": "I",
|
| 222 |
-
"correct": false
|
| 223 |
-
},
|
| 224 |
-
{
|
| 225 |
-
"question_id": 95,
|
| 226 |
-
"category": "business",
|
| 227 |
-
"gold": "J",
|
| 228 |
-
"prediction": "J",
|
| 229 |
-
"raw_generation": "J",
|
| 230 |
-
"correct": true
|
| 231 |
-
},
|
| 232 |
-
{
|
| 233 |
-
"question_id": 96,
|
| 234 |
-
"category": "business",
|
| 235 |
-
"gold": "H",
|
| 236 |
-
"prediction": "C",
|
| 237 |
-
"raw_generation": "C",
|
| 238 |
-
"correct": false
|
| 239 |
-
},
|
| 240 |
-
{
|
| 241 |
-
"question_id": 97,
|
| 242 |
-
"category": "business",
|
| 243 |
-
"gold": "F",
|
| 244 |
-
"prediction": "F",
|
| 245 |
-
"raw_generation": "F",
|
| 246 |
-
"correct": true
|
| 247 |
-
},
|
| 248 |
-
{
|
| 249 |
-
"question_id": 98,
|
| 250 |
-
"category": "business",
|
| 251 |
-
"gold": "H",
|
| 252 |
-
"prediction": "D",
|
| 253 |
-
"raw_generation": "D",
|
| 254 |
-
"correct": false
|
| 255 |
-
},
|
| 256 |
-
{
|
| 257 |
-
"question_id": 99,
|
| 258 |
-
"category": "business",
|
| 259 |
-
"gold": "C",
|
| 260 |
-
"prediction": "C",
|
| 261 |
-
"raw_generation": "C",
|
| 262 |
-
"correct": true
|
| 263 |
-
}
|
| 264 |
-
],
|
| 265 |
-
"attention_impl": "sdpa",
|
| 266 |
-
"elapsed_seconds": 91.607788196
|
| 267 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|