Spaces:
Running
Running
Add/update results for Gemma-3 12B (version 0.1.2, guid 613829d1b5f9412ab338d0a255e13173)
Browse files- results.json +435 -350
results.json
CHANGED
|
@@ -1,351 +1,436 @@
|
|
| 1 |
-
{
|
| 2 |
-
"items": {
|
| 3 |
-
"0.1.1": {
|
| 4 |
-
"b0cf95586f434818a31f12325d9eae66": {
|
| 5 |
-
"model_name": "Gemma-3 4B",
|
| 6 |
-
"timestamp": "2025-03-28T10:48:36",
|
| 7 |
-
"config": {
|
| 8 |
-
"embedding_model": "sentence-transformers/distiluse-base-multilingual-cased",
|
| 9 |
-
"retriever_type": "mmr",
|
| 10 |
-
"retrieval_config": {
|
| 11 |
-
"top_k": 5,
|
| 12 |
-
"chunk_size": 1000,
|
| 13 |
-
"chunk_overlap": 100
|
| 14 |
-
}
|
| 15 |
-
},
|
| 16 |
-
"metrics": {
|
| 17 |
-
"cond": {
|
| 18 |
-
"retrieval": {
|
| 19 |
-
"hit_rate": 0.74,
|
| 20 |
-
"mrr": 0.5920000000000001,
|
| 21 |
-
"precision": 0.16799999999999998
|
| 22 |
-
},
|
| 23 |
-
"generation": {
|
| 24 |
-
"rouge1": 0.11885237343301858,
|
| 25 |
-
"rougeL": 0.11885237343301858
|
| 26 |
-
}
|
| 27 |
-
},
|
| 28 |
-
"comp": {
|
| 29 |
-
"retrieval": {
|
| 30 |
-
"hit_rate": 0.72,
|
| 31 |
-
"mrr": 0.4933333333333333,
|
| 32 |
-
"precision": 0.2
|
| 33 |
-
},
|
| 34 |
-
"generation": {
|
| 35 |
-
"rouge1": 0.08331701939804853,
|
| 36 |
-
"rougeL": 0.08331701939804853
|
| 37 |
-
}
|
| 38 |
-
},
|
| 39 |
-
"set": {
|
| 40 |
-
"retrieval": {
|
| 41 |
-
"hit_rate": 0.52,
|
| 42 |
-
"mrr": 0.38066666666666665,
|
| 43 |
-
"precision": 0.132
|
| 44 |
-
},
|
| 45 |
-
"generation": {
|
| 46 |
-
"rouge1": 0.09937513859436432,
|
| 47 |
-
"rougeL": 0.0802916087780759
|
| 48 |
-
}
|
| 49 |
-
},
|
| 50 |
-
"mh": {
|
| 51 |
-
"retrieval": {
|
| 52 |
-
"hit_rate": 0.6,
|
| 53 |
-
"mrr": 0.4073333333333333,
|
| 54 |
-
"precision": 0.14800000000000002
|
| 55 |
-
},
|
| 56 |
-
"generation": {
|
| 57 |
-
"rouge1": 0.0643998533286459,
|
| 58 |
-
"rougeL": 0.0643998533286459
|
| 59 |
-
}
|
| 60 |
-
},
|
| 61 |
-
"simple": {
|
| 62 |
-
"retrieval": {
|
| 63 |
-
"hit_rate": 0.5,
|
| 64 |
-
"mrr": 0.3496666666666666,
|
| 65 |
-
"precision": 0.12
|
| 66 |
-
},
|
| 67 |
-
"generation": {
|
| 68 |
-
"rouge1": 0.05821997465191799,
|
| 69 |
-
"rougeL": 0.05821997465191799
|
| 70 |
-
}
|
| 71 |
-
},
|
| 72 |
-
"overall": {
|
| 73 |
-
"retrieval": {
|
| 74 |
-
"hit_rate": 0.616,
|
| 75 |
-
"mrr": 0.4446,
|
| 76 |
-
"precision": 0.15360000000000001
|
| 77 |
-
},
|
| 78 |
-
"generation": {
|
| 79 |
-
"rouge1": 0.08483287188119908,
|
| 80 |
-
"rougeL": 0.08101616591794138
|
| 81 |
-
}
|
| 82 |
-
}
|
| 83 |
-
},
|
| 84 |
-
"metadata": {
|
| 85 |
-
"n_questions": 250,
|
| 86 |
-
"submit_timestamp": ""
|
| 87 |
-
}
|
| 88 |
-
},
|
| 89 |
-
"73ae38971b0c48a2bedd143db7c25d3b": {
|
| 90 |
-
"model_name": "Llama 3.3 70B",
|
| 91 |
-
"timestamp": "2025-03-28T10:51:17",
|
| 92 |
-
"config": {
|
| 93 |
-
"embedding_model": "sentence-transformers/distiluse-base-multilingual-cased",
|
| 94 |
-
"retriever_type": "mmr",
|
| 95 |
-
"retrieval_config": {
|
| 96 |
-
"top_k": 5,
|
| 97 |
-
"chunk_size": 1000,
|
| 98 |
-
"chunk_overlap": 100
|
| 99 |
-
}
|
| 100 |
-
},
|
| 101 |
-
"metrics": {
|
| 102 |
-
"cond": {
|
| 103 |
-
"retrieval": {
|
| 104 |
-
"hit_rate": 0.64,
|
| 105 |
-
"mrr": 0.514,
|
| 106 |
-
"precision": 0.14400000000000002
|
| 107 |
-
},
|
| 108 |
-
"generation": {
|
| 109 |
-
"rouge1": 0.2493333333333333,
|
| 110 |
-
"rougeL": 0.2493333333333333
|
| 111 |
-
}
|
| 112 |
-
},
|
| 113 |
-
"comp": {
|
| 114 |
-
"retrieval": {
|
| 115 |
-
"hit_rate": 0.72,
|
| 116 |
-
"mrr": 0.4933333333333333,
|
| 117 |
-
"precision": 0.2
|
| 118 |
-
},
|
| 119 |
-
"generation": {
|
| 120 |
-
"rouge1": 0.23564204559660382,
|
| 121 |
-
"rougeL": 0.23564204559660382
|
| 122 |
-
}
|
| 123 |
-
},
|
| 124 |
-
"set": {
|
| 125 |
-
"retrieval": {
|
| 126 |
-
"hit_rate": 0.5,
|
| 127 |
-
"mrr": 0.3683333333333333,
|
| 128 |
-
"precision": 0.12400000000000003
|
| 129 |
-
},
|
| 130 |
-
"generation": {
|
| 131 |
-
"rouge1": 0.21594023633499185,
|
| 132 |
-
"rougeL": 0.17318924819269935
|
| 133 |
-
}
|
| 134 |
-
},
|
| 135 |
-
"mh": {
|
| 136 |
-
"retrieval": {
|
| 137 |
-
"hit_rate": 0.58,
|
| 138 |
-
"mrr": 0.3846666666666666,
|
| 139 |
-
"precision": 0.14800000000000002
|
| 140 |
-
},
|
| 141 |
-
"generation": {
|
| 142 |
-
"rouge1": 0.19785714285714284,
|
| 143 |
-
"rougeL": 0.19785714285714284
|
| 144 |
-
}
|
| 145 |
-
},
|
| 146 |
-
"simple": {
|
| 147 |
-
"retrieval": {
|
| 148 |
-
"hit_rate": 0.48,
|
| 149 |
-
"mrr": 0.33399999999999996,
|
| 150 |
-
"precision": 0.12
|
| 151 |
-
},
|
| 152 |
-
"generation": {
|
| 153 |
-
"rouge1": 0.16162049062049064,
|
| 154 |
-
"rougeL": 0.16162049062049064
|
| 155 |
-
}
|
| 156 |
-
},
|
| 157 |
-
"overall": {
|
| 158 |
-
"retrieval": {
|
| 159 |
-
"hit_rate": 0.584,
|
| 160 |
-
"mrr": 0.41886666666666666,
|
| 161 |
-
"precision": 0.1472
|
| 162 |
-
},
|
| 163 |
-
"generation": {
|
| 164 |
-
"rouge1": 0.21207864974851248,
|
| 165 |
-
"rougeL": 0.20352845212005397
|
| 166 |
-
}
|
| 167 |
-
}
|
| 168 |
-
},
|
| 169 |
-
"metadata": {
|
| 170 |
-
"n_questions": 250,
|
| 171 |
-
"submit_timestamp": ""
|
| 172 |
-
}
|
| 173 |
-
},
|
| 174 |
-
"b4e8183f6aab486681c2c6ea3a13165d": {
|
| 175 |
-
"model_name": "Gemma-3 12B",
|
| 176 |
-
"timestamp": "2025-03-28T10:52:25",
|
| 177 |
-
"config": {
|
| 178 |
-
"embedding_model": "sentence-transformers/distiluse-base-multilingual-cased",
|
| 179 |
-
"retriever_type": "mmr",
|
| 180 |
-
"retrieval_config": {
|
| 181 |
-
"top_k": 5,
|
| 182 |
-
"chunk_size": 1000,
|
| 183 |
-
"chunk_overlap": 100
|
| 184 |
-
}
|
| 185 |
-
},
|
| 186 |
-
"metrics": {
|
| 187 |
-
"cond": {
|
| 188 |
-
"retrieval": {
|
| 189 |
-
"hit_rate": 0.74,
|
| 190 |
-
"mrr": 0.5920000000000001,
|
| 191 |
-
"precision": 0.16799999999999998
|
| 192 |
-
},
|
| 193 |
-
"generation": {
|
| 194 |
-
"rouge1": 0.14584126984126983,
|
| 195 |
-
"rougeL": 0.14584126984126983
|
| 196 |
-
}
|
| 197 |
-
},
|
| 198 |
-
"comp": {
|
| 199 |
-
"retrieval": {
|
| 200 |
-
"hit_rate": 0.72,
|
| 201 |
-
"mrr": 0.4933333333333333,
|
| 202 |
-
"precision": 0.2
|
| 203 |
-
},
|
| 204 |
-
"generation": {
|
| 205 |
-
"rouge1": 0.17622510822510826,
|
| 206 |
-
"rougeL": 0.17622510822510826
|
| 207 |
-
}
|
| 208 |
-
},
|
| 209 |
-
"set": {
|
| 210 |
-
"retrieval": {
|
| 211 |
-
"hit_rate": 0.52,
|
| 212 |
-
"mrr": 0.38066666666666665,
|
| 213 |
-
"precision": 0.132
|
| 214 |
-
},
|
| 215 |
-
"generation": {
|
| 216 |
-
"rouge1": 0.13877635775154523,
|
| 217 |
-
"rougeL": 0.11396197213715961
|
| 218 |
-
}
|
| 219 |
-
},
|
| 220 |
-
"mh": {
|
| 221 |
-
"retrieval": {
|
| 222 |
-
"hit_rate": 0.6,
|
| 223 |
-
"mrr": 0.4073333333333333,
|
| 224 |
-
"precision": 0.14800000000000002
|
| 225 |
-
},
|
| 226 |
-
"generation": {
|
| 227 |
-
"rouge1": 0.047134453781512604,
|
| 228 |
-
"rougeL": 0.047134453781512604
|
| 229 |
-
}
|
| 230 |
-
},
|
| 231 |
-
"simple": {
|
| 232 |
-
"retrieval": {
|
| 233 |
-
"hit_rate": 0.5,
|
| 234 |
-
"mrr": 0.3496666666666666,
|
| 235 |
-
"precision": 0.12
|
| 236 |
-
},
|
| 237 |
-
"generation": {
|
| 238 |
-
"rouge1": 0.08229166315355574,
|
| 239 |
-
"rougeL": 0.08229166315355574
|
| 240 |
-
}
|
| 241 |
-
},
|
| 242 |
-
"overall": {
|
| 243 |
-
"retrieval": {
|
| 244 |
-
"hit_rate": 0.616,
|
| 245 |
-
"mrr": 0.4446,
|
| 246 |
-
"precision": 0.15360000000000001
|
| 247 |
-
},
|
| 248 |
-
"generation": {
|
| 249 |
-
"rouge1": 0.11805377055059833,
|
| 250 |
-
"rougeL": 0.11309089342772122
|
| 251 |
-
}
|
| 252 |
-
}
|
| 253 |
-
},
|
| 254 |
-
"metadata": {
|
| 255 |
-
"n_questions": 250,
|
| 256 |
-
"submit_timestamp": ""
|
| 257 |
-
}
|
| 258 |
-
}
|
| 259 |
-
},
|
| 260 |
-
"0.1.2": {
|
| 261 |
-
"73ae38971b0c48a2bedd143db7c25d3b": {
|
| 262 |
-
"model_name": "Llama 3.3 70B",
|
| 263 |
-
"timestamp": "2025-03-28T10:55:59",
|
| 264 |
-
"config": {
|
| 265 |
-
"embedding_model": "sentence-transformers/distiluse-base-multilingual-cased",
|
| 266 |
-
"retriever_type": "mmr",
|
| 267 |
-
"retrieval_config": {
|
| 268 |
-
"top_k": 5,
|
| 269 |
-
"chunk_size": 1000,
|
| 270 |
-
"chunk_overlap": 100
|
| 271 |
-
}
|
| 272 |
-
},
|
| 273 |
-
"metrics": {
|
| 274 |
-
"cond": {
|
| 275 |
-
"retrieval": {
|
| 276 |
-
"hit_rate": 0.64,
|
| 277 |
-
"mrr": 0.514,
|
| 278 |
-
"precision": 0.14400000000000002
|
| 279 |
-
},
|
| 280 |
-
"generation": {
|
| 281 |
-
"rouge1": 0.2493333333333333,
|
| 282 |
-
"rougeL": 0.2493333333333333
|
| 283 |
-
}
|
| 284 |
-
},
|
| 285 |
-
"comp": {
|
| 286 |
-
"retrieval": {
|
| 287 |
-
"hit_rate": 0.72,
|
| 288 |
-
"mrr": 0.4933333333333333,
|
| 289 |
-
"precision": 0.2
|
| 290 |
-
},
|
| 291 |
-
"generation": {
|
| 292 |
-
"rouge1": 0.23564204559660382,
|
| 293 |
-
"rougeL": 0.23564204559660382
|
| 294 |
-
}
|
| 295 |
-
},
|
| 296 |
-
"set": {
|
| 297 |
-
"retrieval": {
|
| 298 |
-
"hit_rate": 0.5,
|
| 299 |
-
"mrr": 0.3683333333333333,
|
| 300 |
-
"precision": 0.12400000000000003
|
| 301 |
-
},
|
| 302 |
-
"generation": {
|
| 303 |
-
"rouge1": 0.21594023633499185,
|
| 304 |
-
"rougeL": 0.17318924819269935
|
| 305 |
-
}
|
| 306 |
-
},
|
| 307 |
-
"mh": {
|
| 308 |
-
"retrieval": {
|
| 309 |
-
"hit_rate": 0.58,
|
| 310 |
-
"mrr": 0.3846666666666666,
|
| 311 |
-
"precision": 0.14800000000000002
|
| 312 |
-
},
|
| 313 |
-
"generation": {
|
| 314 |
-
"rouge1": 0.19785714285714284,
|
| 315 |
-
"rougeL": 0.19785714285714284
|
| 316 |
-
}
|
| 317 |
-
},
|
| 318 |
-
"simple": {
|
| 319 |
-
"retrieval": {
|
| 320 |
-
"hit_rate": 0.48,
|
| 321 |
-
"mrr": 0.33399999999999996,
|
| 322 |
-
"precision": 0.12
|
| 323 |
-
},
|
| 324 |
-
"generation": {
|
| 325 |
-
"rouge1": 0.16162049062049064,
|
| 326 |
-
"rougeL": 0.16162049062049064
|
| 327 |
-
}
|
| 328 |
-
},
|
| 329 |
-
"overall": {
|
| 330 |
-
"retrieval": {
|
| 331 |
-
"hit_rate": 0.584,
|
| 332 |
-
"mrr": 0.41886666666666666,
|
| 333 |
-
"precision": 0.1472
|
| 334 |
-
},
|
| 335 |
-
"generation": {
|
| 336 |
-
"rouge1": 0.21207864974851248,
|
| 337 |
-
"rougeL": 0.20352845212005397
|
| 338 |
-
}
|
| 339 |
-
}
|
| 340 |
-
},
|
| 341 |
-
"metadata": {
|
| 342 |
-
"n_questions": 250,
|
| 343 |
-
"submit_timestamp": ""
|
| 344 |
-
}
|
| 345 |
-
}
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
|
| 349 |
-
|
| 350 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 351 |
}
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"items": {
|
| 3 |
+
"0.1.1": {
|
| 4 |
+
"b0cf95586f434818a31f12325d9eae66": {
|
| 5 |
+
"model_name": "Gemma-3 4B",
|
| 6 |
+
"timestamp": "2025-03-28T10:48:36",
|
| 7 |
+
"config": {
|
| 8 |
+
"embedding_model": "sentence-transformers/distiluse-base-multilingual-cased",
|
| 9 |
+
"retriever_type": "mmr",
|
| 10 |
+
"retrieval_config": {
|
| 11 |
+
"top_k": 5,
|
| 12 |
+
"chunk_size": 1000,
|
| 13 |
+
"chunk_overlap": 100
|
| 14 |
+
}
|
| 15 |
+
},
|
| 16 |
+
"metrics": {
|
| 17 |
+
"cond": {
|
| 18 |
+
"retrieval": {
|
| 19 |
+
"hit_rate": 0.74,
|
| 20 |
+
"mrr": 0.5920000000000001,
|
| 21 |
+
"precision": 0.16799999999999998
|
| 22 |
+
},
|
| 23 |
+
"generation": {
|
| 24 |
+
"rouge1": 0.11885237343301858,
|
| 25 |
+
"rougeL": 0.11885237343301858
|
| 26 |
+
}
|
| 27 |
+
},
|
| 28 |
+
"comp": {
|
| 29 |
+
"retrieval": {
|
| 30 |
+
"hit_rate": 0.72,
|
| 31 |
+
"mrr": 0.4933333333333333,
|
| 32 |
+
"precision": 0.2
|
| 33 |
+
},
|
| 34 |
+
"generation": {
|
| 35 |
+
"rouge1": 0.08331701939804853,
|
| 36 |
+
"rougeL": 0.08331701939804853
|
| 37 |
+
}
|
| 38 |
+
},
|
| 39 |
+
"set": {
|
| 40 |
+
"retrieval": {
|
| 41 |
+
"hit_rate": 0.52,
|
| 42 |
+
"mrr": 0.38066666666666665,
|
| 43 |
+
"precision": 0.132
|
| 44 |
+
},
|
| 45 |
+
"generation": {
|
| 46 |
+
"rouge1": 0.09937513859436432,
|
| 47 |
+
"rougeL": 0.0802916087780759
|
| 48 |
+
}
|
| 49 |
+
},
|
| 50 |
+
"mh": {
|
| 51 |
+
"retrieval": {
|
| 52 |
+
"hit_rate": 0.6,
|
| 53 |
+
"mrr": 0.4073333333333333,
|
| 54 |
+
"precision": 0.14800000000000002
|
| 55 |
+
},
|
| 56 |
+
"generation": {
|
| 57 |
+
"rouge1": 0.0643998533286459,
|
| 58 |
+
"rougeL": 0.0643998533286459
|
| 59 |
+
}
|
| 60 |
+
},
|
| 61 |
+
"simple": {
|
| 62 |
+
"retrieval": {
|
| 63 |
+
"hit_rate": 0.5,
|
| 64 |
+
"mrr": 0.3496666666666666,
|
| 65 |
+
"precision": 0.12
|
| 66 |
+
},
|
| 67 |
+
"generation": {
|
| 68 |
+
"rouge1": 0.05821997465191799,
|
| 69 |
+
"rougeL": 0.05821997465191799
|
| 70 |
+
}
|
| 71 |
+
},
|
| 72 |
+
"overall": {
|
| 73 |
+
"retrieval": {
|
| 74 |
+
"hit_rate": 0.616,
|
| 75 |
+
"mrr": 0.4446,
|
| 76 |
+
"precision": 0.15360000000000001
|
| 77 |
+
},
|
| 78 |
+
"generation": {
|
| 79 |
+
"rouge1": 0.08483287188119908,
|
| 80 |
+
"rougeL": 0.08101616591794138
|
| 81 |
+
}
|
| 82 |
+
}
|
| 83 |
+
},
|
| 84 |
+
"metadata": {
|
| 85 |
+
"n_questions": 250,
|
| 86 |
+
"submit_timestamp": ""
|
| 87 |
+
}
|
| 88 |
+
},
|
| 89 |
+
"73ae38971b0c48a2bedd143db7c25d3b": {
|
| 90 |
+
"model_name": "Llama 3.3 70B",
|
| 91 |
+
"timestamp": "2025-03-28T10:51:17",
|
| 92 |
+
"config": {
|
| 93 |
+
"embedding_model": "sentence-transformers/distiluse-base-multilingual-cased",
|
| 94 |
+
"retriever_type": "mmr",
|
| 95 |
+
"retrieval_config": {
|
| 96 |
+
"top_k": 5,
|
| 97 |
+
"chunk_size": 1000,
|
| 98 |
+
"chunk_overlap": 100
|
| 99 |
+
}
|
| 100 |
+
},
|
| 101 |
+
"metrics": {
|
| 102 |
+
"cond": {
|
| 103 |
+
"retrieval": {
|
| 104 |
+
"hit_rate": 0.64,
|
| 105 |
+
"mrr": 0.514,
|
| 106 |
+
"precision": 0.14400000000000002
|
| 107 |
+
},
|
| 108 |
+
"generation": {
|
| 109 |
+
"rouge1": 0.2493333333333333,
|
| 110 |
+
"rougeL": 0.2493333333333333
|
| 111 |
+
}
|
| 112 |
+
},
|
| 113 |
+
"comp": {
|
| 114 |
+
"retrieval": {
|
| 115 |
+
"hit_rate": 0.72,
|
| 116 |
+
"mrr": 0.4933333333333333,
|
| 117 |
+
"precision": 0.2
|
| 118 |
+
},
|
| 119 |
+
"generation": {
|
| 120 |
+
"rouge1": 0.23564204559660382,
|
| 121 |
+
"rougeL": 0.23564204559660382
|
| 122 |
+
}
|
| 123 |
+
},
|
| 124 |
+
"set": {
|
| 125 |
+
"retrieval": {
|
| 126 |
+
"hit_rate": 0.5,
|
| 127 |
+
"mrr": 0.3683333333333333,
|
| 128 |
+
"precision": 0.12400000000000003
|
| 129 |
+
},
|
| 130 |
+
"generation": {
|
| 131 |
+
"rouge1": 0.21594023633499185,
|
| 132 |
+
"rougeL": 0.17318924819269935
|
| 133 |
+
}
|
| 134 |
+
},
|
| 135 |
+
"mh": {
|
| 136 |
+
"retrieval": {
|
| 137 |
+
"hit_rate": 0.58,
|
| 138 |
+
"mrr": 0.3846666666666666,
|
| 139 |
+
"precision": 0.14800000000000002
|
| 140 |
+
},
|
| 141 |
+
"generation": {
|
| 142 |
+
"rouge1": 0.19785714285714284,
|
| 143 |
+
"rougeL": 0.19785714285714284
|
| 144 |
+
}
|
| 145 |
+
},
|
| 146 |
+
"simple": {
|
| 147 |
+
"retrieval": {
|
| 148 |
+
"hit_rate": 0.48,
|
| 149 |
+
"mrr": 0.33399999999999996,
|
| 150 |
+
"precision": 0.12
|
| 151 |
+
},
|
| 152 |
+
"generation": {
|
| 153 |
+
"rouge1": 0.16162049062049064,
|
| 154 |
+
"rougeL": 0.16162049062049064
|
| 155 |
+
}
|
| 156 |
+
},
|
| 157 |
+
"overall": {
|
| 158 |
+
"retrieval": {
|
| 159 |
+
"hit_rate": 0.584,
|
| 160 |
+
"mrr": 0.41886666666666666,
|
| 161 |
+
"precision": 0.1472
|
| 162 |
+
},
|
| 163 |
+
"generation": {
|
| 164 |
+
"rouge1": 0.21207864974851248,
|
| 165 |
+
"rougeL": 0.20352845212005397
|
| 166 |
+
}
|
| 167 |
+
}
|
| 168 |
+
},
|
| 169 |
+
"metadata": {
|
| 170 |
+
"n_questions": 250,
|
| 171 |
+
"submit_timestamp": ""
|
| 172 |
+
}
|
| 173 |
+
},
|
| 174 |
+
"b4e8183f6aab486681c2c6ea3a13165d": {
|
| 175 |
+
"model_name": "Gemma-3 12B",
|
| 176 |
+
"timestamp": "2025-03-28T10:52:25",
|
| 177 |
+
"config": {
|
| 178 |
+
"embedding_model": "sentence-transformers/distiluse-base-multilingual-cased",
|
| 179 |
+
"retriever_type": "mmr",
|
| 180 |
+
"retrieval_config": {
|
| 181 |
+
"top_k": 5,
|
| 182 |
+
"chunk_size": 1000,
|
| 183 |
+
"chunk_overlap": 100
|
| 184 |
+
}
|
| 185 |
+
},
|
| 186 |
+
"metrics": {
|
| 187 |
+
"cond": {
|
| 188 |
+
"retrieval": {
|
| 189 |
+
"hit_rate": 0.74,
|
| 190 |
+
"mrr": 0.5920000000000001,
|
| 191 |
+
"precision": 0.16799999999999998
|
| 192 |
+
},
|
| 193 |
+
"generation": {
|
| 194 |
+
"rouge1": 0.14584126984126983,
|
| 195 |
+
"rougeL": 0.14584126984126983
|
| 196 |
+
}
|
| 197 |
+
},
|
| 198 |
+
"comp": {
|
| 199 |
+
"retrieval": {
|
| 200 |
+
"hit_rate": 0.72,
|
| 201 |
+
"mrr": 0.4933333333333333,
|
| 202 |
+
"precision": 0.2
|
| 203 |
+
},
|
| 204 |
+
"generation": {
|
| 205 |
+
"rouge1": 0.17622510822510826,
|
| 206 |
+
"rougeL": 0.17622510822510826
|
| 207 |
+
}
|
| 208 |
+
},
|
| 209 |
+
"set": {
|
| 210 |
+
"retrieval": {
|
| 211 |
+
"hit_rate": 0.52,
|
| 212 |
+
"mrr": 0.38066666666666665,
|
| 213 |
+
"precision": 0.132
|
| 214 |
+
},
|
| 215 |
+
"generation": {
|
| 216 |
+
"rouge1": 0.13877635775154523,
|
| 217 |
+
"rougeL": 0.11396197213715961
|
| 218 |
+
}
|
| 219 |
+
},
|
| 220 |
+
"mh": {
|
| 221 |
+
"retrieval": {
|
| 222 |
+
"hit_rate": 0.6,
|
| 223 |
+
"mrr": 0.4073333333333333,
|
| 224 |
+
"precision": 0.14800000000000002
|
| 225 |
+
},
|
| 226 |
+
"generation": {
|
| 227 |
+
"rouge1": 0.047134453781512604,
|
| 228 |
+
"rougeL": 0.047134453781512604
|
| 229 |
+
}
|
| 230 |
+
},
|
| 231 |
+
"simple": {
|
| 232 |
+
"retrieval": {
|
| 233 |
+
"hit_rate": 0.5,
|
| 234 |
+
"mrr": 0.3496666666666666,
|
| 235 |
+
"precision": 0.12
|
| 236 |
+
},
|
| 237 |
+
"generation": {
|
| 238 |
+
"rouge1": 0.08229166315355574,
|
| 239 |
+
"rougeL": 0.08229166315355574
|
| 240 |
+
}
|
| 241 |
+
},
|
| 242 |
+
"overall": {
|
| 243 |
+
"retrieval": {
|
| 244 |
+
"hit_rate": 0.616,
|
| 245 |
+
"mrr": 0.4446,
|
| 246 |
+
"precision": 0.15360000000000001
|
| 247 |
+
},
|
| 248 |
+
"generation": {
|
| 249 |
+
"rouge1": 0.11805377055059833,
|
| 250 |
+
"rougeL": 0.11309089342772122
|
| 251 |
+
}
|
| 252 |
+
}
|
| 253 |
+
},
|
| 254 |
+
"metadata": {
|
| 255 |
+
"n_questions": 250,
|
| 256 |
+
"submit_timestamp": ""
|
| 257 |
+
}
|
| 258 |
+
}
|
| 259 |
+
},
|
| 260 |
+
"0.1.2": {
|
| 261 |
+
"73ae38971b0c48a2bedd143db7c25d3b": {
|
| 262 |
+
"model_name": "Llama 3.3 70B",
|
| 263 |
+
"timestamp": "2025-03-28T10:55:59",
|
| 264 |
+
"config": {
|
| 265 |
+
"embedding_model": "sentence-transformers/distiluse-base-multilingual-cased",
|
| 266 |
+
"retriever_type": "mmr",
|
| 267 |
+
"retrieval_config": {
|
| 268 |
+
"top_k": 5,
|
| 269 |
+
"chunk_size": 1000,
|
| 270 |
+
"chunk_overlap": 100
|
| 271 |
+
}
|
| 272 |
+
},
|
| 273 |
+
"metrics": {
|
| 274 |
+
"cond": {
|
| 275 |
+
"retrieval": {
|
| 276 |
+
"hit_rate": 0.64,
|
| 277 |
+
"mrr": 0.514,
|
| 278 |
+
"precision": 0.14400000000000002
|
| 279 |
+
},
|
| 280 |
+
"generation": {
|
| 281 |
+
"rouge1": 0.2493333333333333,
|
| 282 |
+
"rougeL": 0.2493333333333333
|
| 283 |
+
}
|
| 284 |
+
},
|
| 285 |
+
"comp": {
|
| 286 |
+
"retrieval": {
|
| 287 |
+
"hit_rate": 0.72,
|
| 288 |
+
"mrr": 0.4933333333333333,
|
| 289 |
+
"precision": 0.2
|
| 290 |
+
},
|
| 291 |
+
"generation": {
|
| 292 |
+
"rouge1": 0.23564204559660382,
|
| 293 |
+
"rougeL": 0.23564204559660382
|
| 294 |
+
}
|
| 295 |
+
},
|
| 296 |
+
"set": {
|
| 297 |
+
"retrieval": {
|
| 298 |
+
"hit_rate": 0.5,
|
| 299 |
+
"mrr": 0.3683333333333333,
|
| 300 |
+
"precision": 0.12400000000000003
|
| 301 |
+
},
|
| 302 |
+
"generation": {
|
| 303 |
+
"rouge1": 0.21594023633499185,
|
| 304 |
+
"rougeL": 0.17318924819269935
|
| 305 |
+
}
|
| 306 |
+
},
|
| 307 |
+
"mh": {
|
| 308 |
+
"retrieval": {
|
| 309 |
+
"hit_rate": 0.58,
|
| 310 |
+
"mrr": 0.3846666666666666,
|
| 311 |
+
"precision": 0.14800000000000002
|
| 312 |
+
},
|
| 313 |
+
"generation": {
|
| 314 |
+
"rouge1": 0.19785714285714284,
|
| 315 |
+
"rougeL": 0.19785714285714284
|
| 316 |
+
}
|
| 317 |
+
},
|
| 318 |
+
"simple": {
|
| 319 |
+
"retrieval": {
|
| 320 |
+
"hit_rate": 0.48,
|
| 321 |
+
"mrr": 0.33399999999999996,
|
| 322 |
+
"precision": 0.12
|
| 323 |
+
},
|
| 324 |
+
"generation": {
|
| 325 |
+
"rouge1": 0.16162049062049064,
|
| 326 |
+
"rougeL": 0.16162049062049064
|
| 327 |
+
}
|
| 328 |
+
},
|
| 329 |
+
"overall": {
|
| 330 |
+
"retrieval": {
|
| 331 |
+
"hit_rate": 0.584,
|
| 332 |
+
"mrr": 0.41886666666666666,
|
| 333 |
+
"precision": 0.1472
|
| 334 |
+
},
|
| 335 |
+
"generation": {
|
| 336 |
+
"rouge1": 0.21207864974851248,
|
| 337 |
+
"rougeL": 0.20352845212005397
|
| 338 |
+
}
|
| 339 |
+
}
|
| 340 |
+
},
|
| 341 |
+
"metadata": {
|
| 342 |
+
"n_questions": 250,
|
| 343 |
+
"submit_timestamp": ""
|
| 344 |
+
}
|
| 345 |
+
},
|
| 346 |
+
"613829d1b5f9412ab338d0a255e13173": {
|
| 347 |
+
"model_name": "Gemma-3 12B",
|
| 348 |
+
"timestamp": "2025-04-01T07:07:53",
|
| 349 |
+
"config": {
|
| 350 |
+
"embedding_model": "sentence-transformers/distiluse-base-multilingual-cased",
|
| 351 |
+
"retriever_type": "mmr",
|
| 352 |
+
"retrieval_config": {
|
| 353 |
+
"top_k": 5,
|
| 354 |
+
"chunk_size": 1000,
|
| 355 |
+
"chunk_overlap": 100
|
| 356 |
+
}
|
| 357 |
+
},
|
| 358 |
+
"metrics": {
|
| 359 |
+
"cond": {
|
| 360 |
+
"retrieval": {
|
| 361 |
+
"hit_rate": 0.74,
|
| 362 |
+
"mrr": 0.5920000000000001,
|
| 363 |
+
"precision": 0.16799999999999998
|
| 364 |
+
},
|
| 365 |
+
"generation": {
|
| 366 |
+
"rouge1": 0.14584126984126983,
|
| 367 |
+
"rougeL": 0.14584126984126983
|
| 368 |
+
}
|
| 369 |
+
},
|
| 370 |
+
"comp": {
|
| 371 |
+
"retrieval": {
|
| 372 |
+
"hit_rate": 0.72,
|
| 373 |
+
"mrr": 0.4933333333333333,
|
| 374 |
+
"precision": 0.2
|
| 375 |
+
},
|
| 376 |
+
"generation": {
|
| 377 |
+
"rouge1": 0.17622510822510826,
|
| 378 |
+
"rougeL": 0.17622510822510826
|
| 379 |
+
}
|
| 380 |
+
},
|
| 381 |
+
"set": {
|
| 382 |
+
"retrieval": {
|
| 383 |
+
"hit_rate": 0.52,
|
| 384 |
+
"mrr": 0.38066666666666665,
|
| 385 |
+
"precision": 0.132
|
| 386 |
+
},
|
| 387 |
+
"generation": {
|
| 388 |
+
"rouge1": 0.13877635775154523,
|
| 389 |
+
"rougeL": 0.11396197213715961
|
| 390 |
+
}
|
| 391 |
+
},
|
| 392 |
+
"mh": {
|
| 393 |
+
"retrieval": {
|
| 394 |
+
"hit_rate": 0.6,
|
| 395 |
+
"mrr": 0.4073333333333333,
|
| 396 |
+
"precision": 0.14800000000000002
|
| 397 |
+
},
|
| 398 |
+
"generation": {
|
| 399 |
+
"rouge1": 0.047134453781512604,
|
| 400 |
+
"rougeL": 0.047134453781512604
|
| 401 |
+
}
|
| 402 |
+
},
|
| 403 |
+
"simple": {
|
| 404 |
+
"retrieval": {
|
| 405 |
+
"hit_rate": 0.5,
|
| 406 |
+
"mrr": 0.3496666666666666,
|
| 407 |
+
"precision": 0.12
|
| 408 |
+
},
|
| 409 |
+
"generation": {
|
| 410 |
+
"rouge1": 0.08229166315355574,
|
| 411 |
+
"rougeL": 0.08229166315355574
|
| 412 |
+
}
|
| 413 |
+
},
|
| 414 |
+
"overall": {
|
| 415 |
+
"retrieval": {
|
| 416 |
+
"hit_rate": 0.616,
|
| 417 |
+
"mrr": 0.4446,
|
| 418 |
+
"precision": 0.15360000000000001
|
| 419 |
+
},
|
| 420 |
+
"generation": {
|
| 421 |
+
"rouge1": 0.11805377055059833,
|
| 422 |
+
"rougeL": 0.11309089342772122
|
| 423 |
+
}
|
| 424 |
+
}
|
| 425 |
+
},
|
| 426 |
+
"metadata": {
|
| 427 |
+
"n_questions": 250,
|
| 428 |
+
"submit_timestamp": ""
|
| 429 |
+
}
|
| 430 |
+
}
|
| 431 |
+
}
|
| 432 |
+
},
|
| 433 |
+
"last_version": "0.1.2",
|
| 434 |
+
"n_questions": 250,
|
| 435 |
+
"date_title": "01 \u0430\u043f\u0440\u0435\u043b\u044f 2025"
|
| 436 |
}
|