Spaces:
AIR-Bench
/
Running on CPU Upgrade

nan commited on
Commit
bb09125
·
1 Parent(s): d4c9a2d

fix: minor fix

Browse files
Files changed (1) hide show
  1. app.py +31 -100
app.py CHANGED
@@ -107,6 +107,7 @@ def update_doc_metric(
107
  show_revision_and_timestamp,
108
  )
109
 
 
110
  def update_datastore(version):
111
  global datastore
112
  global ds_dict
@@ -117,6 +118,7 @@ def update_datastore(version):
117
  print(f"current data version: {datastore.version}")
118
  return datastore
119
 
 
120
  def update_qa_domains(version):
121
  datastore = update_datastore(version)
122
  domain_elem = get_domain_dropdown(QABenchmarks[datastore.slug])
@@ -128,6 +130,7 @@ def update_doc_domains(version):
128
  domain_elem = get_domain_dropdown(LongDocBenchmarks[datastore.slug])
129
  return domain_elem
130
 
 
131
  def update_qa_langs(version):
132
  datastore = update_datastore(version)
133
  lang_elem = get_language_dropdown(QABenchmarks[datastore.slug])
@@ -139,6 +142,7 @@ def update_doc_langs(version):
139
  lang_elem = get_language_dropdown(LongDocBenchmarks[datastore.slug])
140
  return lang_elem
141
 
 
142
  def update_qa_models(version):
143
  datastore = update_datastore(version)
144
  model_elem = get_reranking_dropdown(datastore.reranking_models)
@@ -149,6 +153,7 @@ def update_qa_df_ret_rerank(version):
149
  datastore = update_datastore(version)
150
  return get_leaderboard_table(datastore.qa_fmt_df, datastore.qa_types)
151
 
 
152
  def update_qa_hidden_df_ret_rerank(version):
153
  datastore = update_datastore(version)
154
  return get_leaderboard_table(datastore.qa_raw_df, datastore.qa_types, visible=False)
@@ -169,6 +174,7 @@ def filter_df_ret(df):
169
  df_ret = reset_rank(df_ret)
170
  return df_ret
171
 
 
172
  def update_qa_df_ret(version):
173
  datastore = update_datastore(version)
174
  df_ret = filter_df_ret(datastore.qa_fmt_df)
@@ -192,26 +198,31 @@ def update_doc_hidden_df_ret(version):
192
  df_ret_hidden = filter_df_ret(datastore.doc_raw_df)
193
  return get_leaderboard_table(df_ret_hidden, datastore.doc_types, visible=False)
194
 
 
195
  def filter_df_rerank(df):
196
  df_rerank = df[df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK]
197
  df_rerank = reset_rank(df_rerank)
198
  return df_rerank
199
 
 
200
  def update_qa_df_rerank(version):
201
  datastore = update_datastore(version)
202
  df_rerank = filter_df_rerank(datastore.qa_fmt_df)
203
  return get_leaderboard_table(df_rerank, datastore.qa_types)
204
 
 
205
  def update_qa_hidden_df_rerank(version):
206
  datastore = update_datastore(version)
207
  df_rerank_hidden = filter_df_rerank(datastore.qa_raw_df)
208
  return get_leaderboard_table(df_rerank_hidden, datastore.qa_types, visible=False)
209
 
 
210
  def update_doc_df_rerank(version):
211
  datastore = update_datastore(version)
212
  df_rerank = filter_df_rerank(datastore.doc_fmt_df)
213
  return get_leaderboard_table(df_rerank, datastore.doc_types)
214
 
 
215
  def update_doc_hidden_df_rerank(version):
216
  datastore = update_datastore(version)
217
  df_rerank_hidden = filter_df_rerank(datastore.doc_raw_df)
@@ -237,19 +248,11 @@ with demo:
237
  # select domain
238
  with gr.Row():
239
  domains = get_domain_dropdown(QABenchmarks[datastore.slug])
240
- version.change(
241
- update_qa_domains,
242
- version,
243
- domains
244
- )
245
  # select language
246
  with gr.Row():
247
  langs = get_language_dropdown(QABenchmarks[datastore.slug])
248
- version.change(
249
- update_qa_langs,
250
- version,
251
- langs
252
- )
253
  with gr.Column():
254
  # select the metric
255
  metric = get_metric_dropdown(METRIC_LIST, DEFAULT_METRIC_QA)
@@ -266,27 +269,15 @@ with demo:
266
  # select reranking models
267
  with gr.Column():
268
  models = get_reranking_dropdown(datastore.reranking_models)
269
- version.change(
270
- update_qa_models,
271
- version,
272
- models
273
- )
274
  # shown_table
275
  qa_df_elem_ret_rerank = get_leaderboard_table(datastore.qa_fmt_df, datastore.qa_types)
276
- version.change(
277
- update_qa_df_ret_rerank,
278
- version,
279
- qa_df_elem_ret_rerank
280
- )
281
  # Dummy leaderboard for handling the case when the user uses backspace key
282
  qa_df_elem_ret_rerank_hidden = get_leaderboard_table(
283
  datastore.qa_raw_df, datastore.qa_types, visible=False
284
  )
285
- version.change(
286
- update_qa_hidden_df_ret_rerank,
287
- version,
288
- qa_df_elem_ret_rerank_hidden
289
- )
290
 
291
  set_listeners(
292
  TaskType.qa,
@@ -315,29 +306,17 @@ with demo:
315
  search_bar_ret = get_search_bar()
316
  with gr.Column(scale=1):
317
  models_ret = get_noreranking_dropdown()
318
- version.change(
319
- update_qa_models,
320
- version,
321
- models_ret
322
- )
323
  _qa_df_ret = filter_df_ret(datastore.qa_fmt_df)
324
  qa_df_elem_ret = get_leaderboard_table(_qa_df_ret, datastore.qa_types)
325
- version.change(
326
- update_qa_df_ret,
327
- version,
328
- qa_df_elem_ret
329
- )
330
 
331
  # Dummy leaderboard for handling the case when the user uses backspace key
332
  _qa_df_ret_hidden = filter_df_ret(datastore.qa_raw_df)
333
  qa_df_elem_ret_hidden = get_leaderboard_table(
334
  _qa_df_ret_hidden, datastore.qa_types, visible=False
335
  )
336
- version.change(
337
- update_qa_hidden_df_ret,
338
- version,
339
- qa_df_elem_ret_hidden
340
- )
341
 
342
  set_listeners(
343
  TaskType.qa,
@@ -373,29 +352,17 @@ with demo:
373
  with gr.Row():
374
  with gr.Column(scale=1):
375
  qa_models_rerank = get_reranking_dropdown(qa_rerank_models)
376
- version.change(
377
- update_qa_models,
378
- version,
379
- qa_models_rerank
380
- )
381
  with gr.Column(scale=1):
382
  qa_search_bar_rerank = gr.Textbox(show_label=False, visible=False)
383
  qa_df_elem_rerank = get_leaderboard_table(_qa_df_rerank, datastore.qa_types)
384
- version.change(
385
- update_qa_df_rerank,
386
- version,
387
- qa_df_elem_rerank
388
- )
389
 
390
  _qa_df_rerank_hidden = filter_df_rerank(datastore.qa_raw_df)
391
  qa_df_elem_rerank_hidden = get_leaderboard_table(
392
  _qa_df_rerank_hidden, datastore.qa_types, visible=False
393
  )
394
- version.change(
395
- update_qa_hidden_df_rerank,
396
- version,
397
- qa_df_elem_rerank_hidden
398
- )
399
 
400
  set_listeners(
401
  TaskType.qa,
@@ -430,19 +397,11 @@ with demo:
430
  # select domain
431
  with gr.Row():
432
  domains = get_domain_dropdown(LongDocBenchmarks[datastore.slug])
433
- version.change(
434
- update_doc_domains,
435
- version,
436
- domains
437
- )
438
  # select language
439
  with gr.Row():
440
  langs = get_language_dropdown(LongDocBenchmarks[datastore.slug])
441
- version.change(
442
- update_doc_langs,
443
- version,
444
- langs
445
- )
446
  with gr.Column():
447
  # select the metric
448
  with gr.Row():
@@ -458,29 +417,17 @@ with demo:
458
  search_bar = get_search_bar()
459
  with gr.Column():
460
  models = get_reranking_dropdown(datastore.reranking_models)
461
- version.change(
462
- update_qa_models,
463
- version,
464
- models
465
- )
466
 
467
  doc_df_elem_ret_rerank = get_leaderboard_table(datastore.doc_fmt_df, datastore.doc_types)
468
 
469
- version.change(
470
- update_doc_df_ret_rerank,
471
- version,
472
- doc_df_elem_ret_rerank
473
- )
474
 
475
  doc_df_elem_ret_rerank_hidden = get_leaderboard_table(
476
  datastore.doc_raw_df, datastore.doc_types, visible=False
477
  )
478
 
479
- version.change(
480
- update_doc_hidden_df_ret_rerank,
481
- version,
482
- doc_df_elem_ret_rerank_hidden
483
- )
484
 
485
  set_listeners(
486
  TaskType.long_doc,
@@ -518,21 +465,13 @@ with demo:
518
  models_ret = get_noreranking_dropdown()
519
  _doc_df_ret = filter_df_ret(datastore.doc_fmt_df)
520
  doc_df_elem_ret = get_leaderboard_table(_doc_df_ret, datastore.doc_types)
521
- version.change(
522
- update_doc_df_ret,
523
- version,
524
- doc_df_elem_ret
525
- )
526
 
527
  _doc_df_ret_hidden = filter_df_ret(datastore.doc_raw_df)
528
  doc_df_elem_ret_hidden = get_leaderboard_table(
529
  _doc_df_ret_hidden, datastore.doc_types, visible=False
530
  )
531
- version.change(
532
- update_doc_hidden_df_ret,
533
- version,
534
- doc_df_elem_ret_hidden
535
- )
536
 
537
  set_listeners(
538
  TaskType.long_doc,
@@ -572,22 +511,14 @@ with demo:
572
  with gr.Column(scale=1):
573
  doc_search_bar_rerank = gr.Textbox(show_label=False, visible=False)
574
  doc_df_elem_rerank = get_leaderboard_table(_doc_df_rerank, datastore.doc_types)
575
- version.change(
576
- update_doc_df_rerank,
577
- version,
578
- doc_df_elem_rerank
579
- )
580
 
581
  _doc_df_rerank_hidden = filter_df_rerank(datastore.doc_raw_df)
582
  doc_df_elem_rerank_hidden = get_leaderboard_table(
583
  _doc_df_rerank_hidden, datastore.doc_types, visible=False
584
  )
585
 
586
- version.change(
587
- update_doc_hidden_df_rerank,
588
- version,
589
- doc_df_elem_rerank_hidden
590
- )
591
 
592
  set_listeners(
593
  TaskType.long_doc,
 
107
  show_revision_and_timestamp,
108
  )
109
 
110
+
111
  def update_datastore(version):
112
  global datastore
113
  global ds_dict
 
118
  print(f"current data version: {datastore.version}")
119
  return datastore
120
 
121
+
122
  def update_qa_domains(version):
123
  datastore = update_datastore(version)
124
  domain_elem = get_domain_dropdown(QABenchmarks[datastore.slug])
 
130
  domain_elem = get_domain_dropdown(LongDocBenchmarks[datastore.slug])
131
  return domain_elem
132
 
133
+
134
  def update_qa_langs(version):
135
  datastore = update_datastore(version)
136
  lang_elem = get_language_dropdown(QABenchmarks[datastore.slug])
 
142
  lang_elem = get_language_dropdown(LongDocBenchmarks[datastore.slug])
143
  return lang_elem
144
 
145
+
146
  def update_qa_models(version):
147
  datastore = update_datastore(version)
148
  model_elem = get_reranking_dropdown(datastore.reranking_models)
 
153
  datastore = update_datastore(version)
154
  return get_leaderboard_table(datastore.qa_fmt_df, datastore.qa_types)
155
 
156
+
157
  def update_qa_hidden_df_ret_rerank(version):
158
  datastore = update_datastore(version)
159
  return get_leaderboard_table(datastore.qa_raw_df, datastore.qa_types, visible=False)
 
174
  df_ret = reset_rank(df_ret)
175
  return df_ret
176
 
177
+
178
  def update_qa_df_ret(version):
179
  datastore = update_datastore(version)
180
  df_ret = filter_df_ret(datastore.qa_fmt_df)
 
198
  df_ret_hidden = filter_df_ret(datastore.doc_raw_df)
199
  return get_leaderboard_table(df_ret_hidden, datastore.doc_types, visible=False)
200
 
201
+
202
  def filter_df_rerank(df):
203
  df_rerank = df[df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK]
204
  df_rerank = reset_rank(df_rerank)
205
  return df_rerank
206
 
207
+
208
  def update_qa_df_rerank(version):
209
  datastore = update_datastore(version)
210
  df_rerank = filter_df_rerank(datastore.qa_fmt_df)
211
  return get_leaderboard_table(df_rerank, datastore.qa_types)
212
 
213
+
214
  def update_qa_hidden_df_rerank(version):
215
  datastore = update_datastore(version)
216
  df_rerank_hidden = filter_df_rerank(datastore.qa_raw_df)
217
  return get_leaderboard_table(df_rerank_hidden, datastore.qa_types, visible=False)
218
 
219
+
220
  def update_doc_df_rerank(version):
221
  datastore = update_datastore(version)
222
  df_rerank = filter_df_rerank(datastore.doc_fmt_df)
223
  return get_leaderboard_table(df_rerank, datastore.doc_types)
224
 
225
+
226
  def update_doc_hidden_df_rerank(version):
227
  datastore = update_datastore(version)
228
  df_rerank_hidden = filter_df_rerank(datastore.doc_raw_df)
 
248
  # select domain
249
  with gr.Row():
250
  domains = get_domain_dropdown(QABenchmarks[datastore.slug])
251
+ version.change(update_qa_domains, version, domains)
 
 
 
 
252
  # select language
253
  with gr.Row():
254
  langs = get_language_dropdown(QABenchmarks[datastore.slug])
255
+ version.change(update_qa_langs, version, langs)
 
 
 
 
256
  with gr.Column():
257
  # select the metric
258
  metric = get_metric_dropdown(METRIC_LIST, DEFAULT_METRIC_QA)
 
269
  # select reranking models
270
  with gr.Column():
271
  models = get_reranking_dropdown(datastore.reranking_models)
272
+ version.change(update_qa_models, version, models)
 
 
 
 
273
  # shown_table
274
  qa_df_elem_ret_rerank = get_leaderboard_table(datastore.qa_fmt_df, datastore.qa_types)
275
+ version.change(update_qa_df_ret_rerank, version, qa_df_elem_ret_rerank)
 
 
 
 
276
  # Dummy leaderboard for handling the case when the user uses backspace key
277
  qa_df_elem_ret_rerank_hidden = get_leaderboard_table(
278
  datastore.qa_raw_df, datastore.qa_types, visible=False
279
  )
280
+ version.change(update_qa_hidden_df_ret_rerank, version, qa_df_elem_ret_rerank_hidden)
 
 
 
 
281
 
282
  set_listeners(
283
  TaskType.qa,
 
306
  search_bar_ret = get_search_bar()
307
  with gr.Column(scale=1):
308
  models_ret = get_noreranking_dropdown()
309
+ version.change(update_qa_models, version, models_ret)
 
 
 
 
310
  _qa_df_ret = filter_df_ret(datastore.qa_fmt_df)
311
  qa_df_elem_ret = get_leaderboard_table(_qa_df_ret, datastore.qa_types)
312
+ version.change(update_qa_df_ret, version, qa_df_elem_ret)
 
 
 
 
313
 
314
  # Dummy leaderboard for handling the case when the user uses backspace key
315
  _qa_df_ret_hidden = filter_df_ret(datastore.qa_raw_df)
316
  qa_df_elem_ret_hidden = get_leaderboard_table(
317
  _qa_df_ret_hidden, datastore.qa_types, visible=False
318
  )
319
+ version.change(update_qa_hidden_df_ret, version, qa_df_elem_ret_hidden)
 
 
 
 
320
 
321
  set_listeners(
322
  TaskType.qa,
 
352
  with gr.Row():
353
  with gr.Column(scale=1):
354
  qa_models_rerank = get_reranking_dropdown(qa_rerank_models)
355
+ version.change(update_qa_models, version, qa_models_rerank)
 
 
 
 
356
  with gr.Column(scale=1):
357
  qa_search_bar_rerank = gr.Textbox(show_label=False, visible=False)
358
  qa_df_elem_rerank = get_leaderboard_table(_qa_df_rerank, datastore.qa_types)
359
+ version.change(update_qa_df_rerank, version, qa_df_elem_rerank)
 
 
 
 
360
 
361
  _qa_df_rerank_hidden = filter_df_rerank(datastore.qa_raw_df)
362
  qa_df_elem_rerank_hidden = get_leaderboard_table(
363
  _qa_df_rerank_hidden, datastore.qa_types, visible=False
364
  )
365
+ version.change(update_qa_hidden_df_rerank, version, qa_df_elem_rerank_hidden)
 
 
 
 
366
 
367
  set_listeners(
368
  TaskType.qa,
 
397
  # select domain
398
  with gr.Row():
399
  domains = get_domain_dropdown(LongDocBenchmarks[datastore.slug])
400
+ version.change(update_doc_domains, version, domains)
 
 
 
 
401
  # select language
402
  with gr.Row():
403
  langs = get_language_dropdown(LongDocBenchmarks[datastore.slug])
404
+ version.change(update_doc_langs, version, langs)
 
 
 
 
405
  with gr.Column():
406
  # select the metric
407
  with gr.Row():
 
417
  search_bar = get_search_bar()
418
  with gr.Column():
419
  models = get_reranking_dropdown(datastore.reranking_models)
420
+ version.change(update_qa_models, version, models)
 
 
 
 
421
 
422
  doc_df_elem_ret_rerank = get_leaderboard_table(datastore.doc_fmt_df, datastore.doc_types)
423
 
424
+ version.change(update_doc_df_ret_rerank, version, doc_df_elem_ret_rerank)
 
 
 
 
425
 
426
  doc_df_elem_ret_rerank_hidden = get_leaderboard_table(
427
  datastore.doc_raw_df, datastore.doc_types, visible=False
428
  )
429
 
430
+ version.change(update_doc_hidden_df_ret_rerank, version, doc_df_elem_ret_rerank_hidden)
 
 
 
 
431
 
432
  set_listeners(
433
  TaskType.long_doc,
 
465
  models_ret = get_noreranking_dropdown()
466
  _doc_df_ret = filter_df_ret(datastore.doc_fmt_df)
467
  doc_df_elem_ret = get_leaderboard_table(_doc_df_ret, datastore.doc_types)
468
+ version.change(update_doc_df_ret, version, doc_df_elem_ret)
 
 
 
 
469
 
470
  _doc_df_ret_hidden = filter_df_ret(datastore.doc_raw_df)
471
  doc_df_elem_ret_hidden = get_leaderboard_table(
472
  _doc_df_ret_hidden, datastore.doc_types, visible=False
473
  )
474
+ version.change(update_doc_hidden_df_ret, version, doc_df_elem_ret_hidden)
 
 
 
 
475
 
476
  set_listeners(
477
  TaskType.long_doc,
 
511
  with gr.Column(scale=1):
512
  doc_search_bar_rerank = gr.Textbox(show_label=False, visible=False)
513
  doc_df_elem_rerank = get_leaderboard_table(_doc_df_rerank, datastore.doc_types)
514
+ version.change(update_doc_df_rerank, version, doc_df_elem_rerank)
 
 
 
 
515
 
516
  _doc_df_rerank_hidden = filter_df_rerank(datastore.doc_raw_df)
517
  doc_df_elem_rerank_hidden = get_leaderboard_table(
518
  _doc_df_rerank_hidden, datastore.doc_types, visible=False
519
  )
520
 
521
+ version.change(update_doc_hidden_df_rerank, version, doc_df_elem_rerank_hidden)
 
 
 
 
522
 
523
  set_listeners(
524
  TaskType.long_doc,