Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
|
@@ -225,9 +225,113 @@ def aggregate_leaderboard_data():
|
|
| 225 |
"kk_biology_unt_mc": 0.22330729166666666,
|
| 226 |
"kk_human_society_rights_unt_mc": 0.242152466367713,
|
| 227 |
},
|
| 228 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 229 |
]
|
| 230 |
-
|
| 231 |
files_list = glob.glob("./m_data/model_data/external/*.json")
|
| 232 |
logging.info(f'FILES LIST: {files_list}')
|
| 233 |
|
|
|
|
| 225 |
"kk_biology_unt_mc": 0.22330729166666666,
|
| 226 |
"kk_human_society_rights_unt_mc": 0.242152466367713,
|
| 227 |
},
|
| 228 |
+
{
|
| 229 |
+
"model_dtype": "torch.float16",
|
| 230 |
+
"model": "gpt-4o-mini",
|
| 231 |
+
"ppl": 0,
|
| 232 |
+
"mmlu_translated_kk": 0.5623775310254735,
|
| 233 |
+
"kk_constitution_mc": 0.79,
|
| 234 |
+
"kk_dastur_mc": 0.755,
|
| 235 |
+
"kazakh_and_literature_unt_mc": 0.4953071672354949,
|
| 236 |
+
"kk_geography_unt_mc": 0.5675203725261933,
|
| 237 |
+
"kk_world_history_unt_mc": 0.6091205211726385,
|
| 238 |
+
"kk_history_of_kazakhstan_unt_mc": 0.47883435582822087,
|
| 239 |
+
"kk_english_unt_mc": 0.6763768775603095,
|
| 240 |
+
"kk_biology_unt_mc": 0.607421875,
|
| 241 |
+
"kk_human_society_rights_unt_mc": 0.7309417040358744,
|
| 242 |
+
},
|
| 243 |
+
{
|
| 244 |
+
"model_dtype": "api",
|
| 245 |
+
"model": "gpt-4o",
|
| 246 |
+
"ppl": 0,
|
| 247 |
+
"mmlu_translated_kk": 0.7419986936642717,
|
| 248 |
+
"kk_constitution_mc": 0.841,
|
| 249 |
+
"kk_dastur_mc": 0.798,
|
| 250 |
+
"kazakh_and_literature_unt_mc": 0.6785409556313993,
|
| 251 |
+
"kk_geography_unt_mc": 0.629802095459837,
|
| 252 |
+
"kk_world_history_unt_mc": 0.6783387622149837,
|
| 253 |
+
"kk_history_of_kazakhstan_unt_mc": 0.6785276073619632,
|
| 254 |
+
"kk_english_unt_mc": 0.7410104688211198,
|
| 255 |
+
"kk_biology_unt_mc": 0.6979166666666666,
|
| 256 |
+
"kk_human_society_rights_unt_mc": 0.7937219730941704,
|
| 257 |
+
},
|
| 258 |
+
{
|
| 259 |
+
"model_dtype": "torch.float16",
|
| 260 |
+
"model": "nova-pro-v1",
|
| 261 |
+
"ppl": 0,
|
| 262 |
+
"mmlu_translated_kk": 0.6792945787067276,
|
| 263 |
+
"kk_constitution_mc": 0.7753623188405797,
|
| 264 |
+
"kk_dastur_mc": 0.718407960199005,
|
| 265 |
+
"kazakh_and_literature_unt_mc": 0.4656569965870307,
|
| 266 |
+
"kk_geography_unt_mc": 0.5541327124563445,
|
| 267 |
+
"kk_world_history_unt_mc": 0.6425081433224755,
|
| 268 |
+
"kk_history_of_kazakhstan_unt_mc": 0.5,
|
| 269 |
+
"kk_english_unt_mc": 0.6845698680018206,
|
| 270 |
+
"kk_biology_unt_mc": 0.6197916666666666,
|
| 271 |
+
"kk_human_society_rights_unt_mc": 0.7713004484304933,
|
| 272 |
+
},
|
| 273 |
+
{
|
| 274 |
+
"model_dtype": "torch.float16",
|
| 275 |
+
"model": "gemini-1.5-pro",
|
| 276 |
+
"ppl": 0,
|
| 277 |
+
"mmlu_translated_kk": 0.7380796864794252,
|
| 278 |
+
"kk_constitution_mc": 0.8164251207729468,
|
| 279 |
+
"kk_dastur_mc": 0.7383084577114428,
|
| 280 |
+
"kazakh_and_literature_unt_mc": 0.5565273037542662,
|
| 281 |
+
"kk_geography_unt_mc": 0.6065192083818394,
|
| 282 |
+
"kk_world_history_unt_mc": 0.6669381107491856,
|
| 283 |
+
"kk_history_of_kazakhstan_unt_mc": 0.5791411042944785,
|
| 284 |
+
"kk_english_unt_mc": 0.7114246700045517,
|
| 285 |
+
"kk_biology_unt_mc": 0.6673177083333334,
|
| 286 |
+
"kk_human_society_rights_unt_mc": 0.7623318385650224,
|
| 287 |
+
},
|
| 288 |
+
{
|
| 289 |
+
"model_dtype": "torch.float16",
|
| 290 |
+
"model": "gemini-1.5-flash",
|
| 291 |
+
"ppl": 0,
|
| 292 |
+
"mmlu_translated_kk": 0.6335728282168517,
|
| 293 |
+
"kk_constitution_mc": 0.748792270531401,
|
| 294 |
+
"kk_dastur_mc": 0.7054726368159204,
|
| 295 |
+
"kazakh_and_literature_unt_mc": 0.4761092150170648,
|
| 296 |
+
"kk_geography_unt_mc": 0.5640279394644936,
|
| 297 |
+
"kk_world_history_unt_mc": 0.5838762214983714,
|
| 298 |
+
"kk_history_of_kazakhstan_unt_mc": 0.43374233128834355,
|
| 299 |
+
"kk_english_unt_mc": 0.6681838871187984,
|
| 300 |
+
"kk_biology_unt_mc": 0.6217447916666666,
|
| 301 |
+
"kk_human_society_rights_unt_mc": 0.7040358744394619,
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"model_dtype": "torch.float16",
|
| 305 |
+
"model": "claude-3-5-sonnet",
|
| 306 |
+
"ppl": 0,
|
| 307 |
+
"mmlu_translated_kk": 0.7335075114304376,
|
| 308 |
+
"kk_constitution_mc": 0.8623188405797102,
|
| 309 |
+
"kk_dastur_mc": 0.7950248756218905,
|
| 310 |
+
"kazakh_and_literature_unt_mc": 0.6548634812286689,
|
| 311 |
+
"kk_geography_unt_mc": 0.6431897555296857,
|
| 312 |
+
"kk_world_history_unt_mc": 0.6669381107491856,
|
| 313 |
+
"kk_history_of_kazakhstan_unt_mc": 0.6251533742331289,
|
| 314 |
+
"kk_english_unt_mc": 0.7291761492944925,
|
| 315 |
+
"kk_biology_unt_mc": 0.6686197916666666,
|
| 316 |
+
"kk_human_society_rights_unt_mc": 0.8026905829596412,
|
| 317 |
+
},
|
| 318 |
+
{
|
| 319 |
+
"model_dtype": "torch.float16",
|
| 320 |
+
"model": "yandex-gpt",
|
| 321 |
+
"ppl": 0,
|
| 322 |
+
"mmlu_translated_kk": 0.39777922926192033,
|
| 323 |
+
"kk_constitution_mc": 0.7028985507246377,
|
| 324 |
+
"kk_dastur_mc": 0.6159203980099502,
|
| 325 |
+
"kazakh_and_literature_unt_mc": 0.3914249146757679,
|
| 326 |
+
"kk_geography_unt_mc": 0.4912689173457509,
|
| 327 |
+
"kk_world_history_unt_mc": 0.5244299674267101,
|
| 328 |
+
"kk_history_of_kazakhstan_unt_mc": 0.4030674846625767,
|
| 329 |
+
"kk_english_unt_mc": 0.5844333181611289,
|
| 330 |
+
"kk_biology_unt_mc": 0.4368489583333333,
|
| 331 |
+
"kk_human_society_rights_unt_mc": 0.6995515695067265,
|
| 332 |
+
},
|
| 333 |
]
|
| 334 |
+
|
| 335 |
files_list = glob.glob("./m_data/model_data/external/*.json")
|
| 336 |
logging.info(f'FILES LIST: {files_list}')
|
| 337 |
|