Spaces:
AIR-Bench
/
Running on CPU Upgrade

nan commited on
Commit
6ae6c7f
·
1 Parent(s): 98af8d1

fix: fix the long-doc version updating bug

Browse files
Files changed (1) hide show
  1. app.py +94 -41
app.py CHANGED
@@ -122,17 +122,28 @@ def update_qa_domains(version):
122
  return domain_elem
123
 
124
 
 
 
 
 
 
125
  def update_qa_langs(version):
126
  datastore = update_datastore(version)
127
  lang_elem = get_language_dropdown(QABenchmarks[datastore.slug])
128
  return lang_elem
129
 
130
 
 
 
 
 
 
131
  def update_qa_models(version):
132
  datastore = update_datastore(version)
133
  model_elem = get_reranking_dropdown(datastore.reranking_models)
134
  return model_elem
135
 
 
136
  def update_qa_df_ret_rerank(version):
137
  datastore = update_datastore(version)
138
  return get_leaderboard_table(datastore.qa_fmt_df, datastore.qa_types)
@@ -142,6 +153,16 @@ def update_qa_hidden_df_ret_rerank(version):
142
  return get_leaderboard_table(datastore.qa_raw_df, datastore.qa_types, visible=False)
143
 
144
 
 
 
 
 
 
 
 
 
 
 
145
  def filter_df_ret(df):
146
  df_ret = df[df[COL_NAME_RERANKING_MODEL] == "NoReranker"]
147
  df_ret = reset_rank(df_ret)
@@ -152,12 +173,24 @@ def update_qa_df_ret(version):
152
  df_ret = filter_df_ret(datastore.qa_fmt_df)
153
  return get_leaderboard_table(df_ret, datastore.qa_types)
154
 
 
155
  def update_qa_hidden_df_ret(version):
156
  datastore = update_datastore(version)
157
  df_ret_hidden = filter_df_ret(datastore.qa_raw_df)
158
  return get_leaderboard_table(df_ret_hidden, datastore.qa_types, visible=False)
159
 
160
 
 
 
 
 
 
 
 
 
 
 
 
161
  def filter_df_rerank(df):
162
  df_rerank = df[df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK]
163
  df_rerank = reset_rank(df_rerank)
@@ -173,17 +206,26 @@ def update_qa_hidden_df_rerank(version):
173
  df_rerank_hidden = filter_df_rerank(datastore.qa_raw_df)
174
  return get_leaderboard_table(df_rerank_hidden, datastore.qa_types, visible=False)
175
 
 
 
 
 
176
 
177
- def update_doc_version(version):
178
- global datastore
179
- global ds_dict
180
- datastore = ds_dict[version]
181
- domain_elem = get_domain_dropdown(LongDocBenchmarks[datastore.slug])
182
- lang_elem = get_language_dropdown(LongDocBenchmarks[datastore.slug])
183
- model_elem = get_reranking_dropdown(datastore.reranking_models)
184
- df_elem = get_leaderboard_table(datastore.doc_fmt_df, datastore.doc_types)
185
- hidden_df_elem = get_leaderboard_table(datastore.doc_raw_df, datastore.doc_types, visible=False)
186
- return domain_elem, lang_elem, model_elem, df_elem, hidden_df_elem
 
 
 
 
 
187
 
188
 
189
  demo = gr.Blocks(css=custom_css)
@@ -409,9 +451,19 @@ with demo:
409
  # select domain
410
  with gr.Row():
411
  domains = get_domain_dropdown(LongDocBenchmarks[datastore.slug])
 
 
 
 
 
412
  # select language
413
  with gr.Row():
414
  langs = get_language_dropdown(LongDocBenchmarks[datastore.slug])
 
 
 
 
 
415
  with gr.Column():
416
  # select the metric
417
  with gr.Row():
@@ -427,18 +479,28 @@ with demo:
427
  search_bar = get_search_bar()
428
  with gr.Column():
429
  models = get_reranking_dropdown(datastore.reranking_models)
 
 
 
 
 
430
 
431
  doc_df_elem_ret_rerank = get_leaderboard_table(datastore.doc_fmt_df, datastore.doc_types)
432
 
433
- # Dummy leaderboard for handling the case when the user uses backspace key
 
 
 
 
 
434
  doc_df_elem_ret_rerank_hidden = get_leaderboard_table(
435
  datastore.doc_raw_df, datastore.doc_types, visible=False
436
  )
437
 
438
  version.change(
439
- update_doc_version,
440
  version,
441
- [domains, langs, models, doc_df_elem_ret_rerank, doc_df_elem_ret_rerank_hidden],
442
  )
443
 
444
  set_listeners(
@@ -475,31 +537,22 @@ with demo:
475
  search_bar_ret = get_search_bar()
476
  with gr.Column(scale=1):
477
  models_ret = get_noreranking_dropdown()
478
-
479
- _doc_df_ret = datastore.doc_fmt_df[
480
- datastore.doc_fmt_df[COL_NAME_RERANKING_MODEL] == "NoReranker"
481
- ]
482
- _doc_df_ret = reset_rank(_doc_df_ret)
483
  doc_df_elem_ret = get_leaderboard_table(_doc_df_ret, datastore.doc_types)
 
 
 
 
 
484
 
485
- _doc_df_ret_hidden = datastore.doc_raw_df[
486
- datastore.doc_raw_df[COL_NAME_RERANKING_MODEL] == "NoReranker"
487
- ]
488
- _doc_df_ret_hidden = reset_rank(_doc_df_ret_hidden)
489
  doc_df_elem_ret_hidden = get_leaderboard_table(
490
  _doc_df_ret_hidden, datastore.doc_types, visible=False
491
  )
492
-
493
  version.change(
494
- update_doc_version,
495
  version,
496
- [
497
- domains,
498
- langs,
499
- models_ret,
500
- doc_df_elem_ret,
501
- doc_df_elem_ret_hidden
502
- ],
503
  )
504
 
505
  set_listeners(
@@ -530,10 +583,7 @@ with demo:
530
  queue=True,
531
  )
532
  with gr.TabItem("Reranking Only", id=22):
533
- _doc_df_rerank = datastore.doc_fmt_df[
534
- datastore.doc_fmt_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
535
- ]
536
- _doc_df_rerank = reset_rank(_doc_df_rerank)
537
  doc_rerank_models = (
538
  _doc_df_rerank[COL_NAME_RERANKING_MODEL].apply(remove_html).unique().tolist()
539
  )
@@ -543,18 +593,21 @@ with demo:
543
  with gr.Column(scale=1):
544
  doc_search_bar_rerank = gr.Textbox(show_label=False, visible=False)
545
  doc_df_elem_rerank = get_leaderboard_table(_doc_df_rerank, datastore.doc_types)
546
- _doc_df_rerank_hidden = datastore.doc_raw_df[
547
- datastore.doc_raw_df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK
548
- ]
549
- _doc_df_rerank_hidden = reset_rank(_doc_df_rerank_hidden)
 
 
 
550
  doc_df_elem_rerank_hidden = get_leaderboard_table(
551
  _doc_df_rerank_hidden, datastore.doc_types, visible=False
552
  )
553
 
554
  version.change(
555
- update_doc_version,
556
  version,
557
- [domains, langs, doc_models_rerank, doc_df_elem_rerank, doc_df_elem_rerank_hidden],
558
  )
559
 
560
  set_listeners(
 
122
  return domain_elem
123
 
124
 
125
+ def update_doc_domains(version):
126
+ datastore = update_datastore(version)
127
+ domain_elem = get_domain_dropdown(LongDocBenchmarks[datastore.slug])
128
+ return domain_elem
129
+
130
  def update_qa_langs(version):
131
  datastore = update_datastore(version)
132
  lang_elem = get_language_dropdown(QABenchmarks[datastore.slug])
133
  return lang_elem
134
 
135
 
136
+ def update_doc_langs(version):
137
+ datastore = update_datastore(version)
138
+ lang_elem = get_language_dropdown(LongDocBenchmarks[datastore.slug])
139
+ return lang_elem
140
+
141
  def update_qa_models(version):
142
  datastore = update_datastore(version)
143
  model_elem = get_reranking_dropdown(datastore.reranking_models)
144
  return model_elem
145
 
146
+
147
  def update_qa_df_ret_rerank(version):
148
  datastore = update_datastore(version)
149
  return get_leaderboard_table(datastore.qa_fmt_df, datastore.qa_types)
 
153
  return get_leaderboard_table(datastore.qa_raw_df, datastore.qa_types, visible=False)
154
 
155
 
156
+ def update_doc_df_ret_rerank(version):
157
+ datastore = update_datastore(version)
158
+ return get_leaderboard_table(datastore.doc_fmt_df, datastore.doc_types)
159
+
160
+
161
+ def update_doc_hidden_df_ret_rerank(version):
162
+ datastore = update_datastore(version)
163
+ return get_leaderboard_table(datastore.doc_raw_df, datastore.doc_types, visible=False)
164
+
165
+
166
  def filter_df_ret(df):
167
  df_ret = df[df[COL_NAME_RERANKING_MODEL] == "NoReranker"]
168
  df_ret = reset_rank(df_ret)
 
173
  df_ret = filter_df_ret(datastore.qa_fmt_df)
174
  return get_leaderboard_table(df_ret, datastore.qa_types)
175
 
176
+
177
  def update_qa_hidden_df_ret(version):
178
  datastore = update_datastore(version)
179
  df_ret_hidden = filter_df_ret(datastore.qa_raw_df)
180
  return get_leaderboard_table(df_ret_hidden, datastore.qa_types, visible=False)
181
 
182
 
183
+ def update_doc_df_ret(version):
184
+ datastore = update_datastore(version)
185
+ df_ret = filter_df_ret(datastore.doc_fmt_df)
186
+ return get_leaderboard_table(df_ret, datastore.doc_types)
187
+
188
+
189
+ def update_doc_hidden_df_ret(version):
190
+ datastore = update_datastore(version)
191
+ df_ret_hidden = filter_df_ret(datastore.doc_raw_df)
192
+ return get_leaderboard_table(df_ret_hidden, datastore.doc_types, visible=False)
193
+
194
  def filter_df_rerank(df):
195
  df_rerank = df[df[COL_NAME_RETRIEVAL_MODEL] == BM25_LINK]
196
  df_rerank = reset_rank(df_rerank)
 
206
  df_rerank_hidden = filter_df_rerank(datastore.qa_raw_df)
207
  return get_leaderboard_table(df_rerank_hidden, datastore.qa_types, visible=False)
208
 
209
+ def update_doc_df_rerank(version):
210
+ datastore = update_datastore(version)
211
+ df_rerank = filter_df_rerank(datastore.doc_fmt_df)
212
+ return get_leaderboard_table(df_rerank, datastore.doc_types)
213
 
214
+ def update_doc_hidden_df_rerank(version):
215
+ datastore = update_datastore(version)
216
+ df_rerank_hidden = filter_df_rerank(datastore.doc_raw_df)
217
+ return get_leaderboard_table(df_rerank_hidden, datastore.doc_types, visible=False)
218
+
219
+ # def update_doc_version(version):
220
+ # global datastore
221
+ # global ds_dict
222
+ # datastore = ds_dict[version]
223
+ # # domain_elem = get_domain_dropdown(LongDocBenchmarks[datastore.slug])
224
+ # # lang_elem = get_language_dropdown(LongDocBenchmarks[datastore.slug])
225
+ # # model_elem = get_reranking_dropdown(datastore.reranking_models)
226
+ # df_elem = get_leaderboard_table(datastore.doc_fmt_df, datastore.doc_types)
227
+ # hidden_df_elem = get_leaderboard_table(datastore.doc_raw_df, datastore.doc_types, visible=False)
228
+ # return domain_elem, lang_elem, model_elem, df_elem, hidden_df_elem
229
 
230
 
231
  demo = gr.Blocks(css=custom_css)
 
451
  # select domain
452
  with gr.Row():
453
  domains = get_domain_dropdown(LongDocBenchmarks[datastore.slug])
454
+ version.change(
455
+ update_doc_domains,
456
+ version,
457
+ domains
458
+ )
459
  # select language
460
  with gr.Row():
461
  langs = get_language_dropdown(LongDocBenchmarks[datastore.slug])
462
+ version.change(
463
+ update_doc_langs,
464
+ version,
465
+ langs
466
+ )
467
  with gr.Column():
468
  # select the metric
469
  with gr.Row():
 
479
  search_bar = get_search_bar()
480
  with gr.Column():
481
  models = get_reranking_dropdown(datastore.reranking_models)
482
+ version.change(
483
+ update_qa_models,
484
+ version,
485
+ models
486
+ )
487
 
488
  doc_df_elem_ret_rerank = get_leaderboard_table(datastore.doc_fmt_df, datastore.doc_types)
489
 
490
+ version.change(
491
+ update_doc_df_ret_rerank,
492
+ version,
493
+ doc_df_elem_ret_rerank
494
+ )
495
+
496
  doc_df_elem_ret_rerank_hidden = get_leaderboard_table(
497
  datastore.doc_raw_df, datastore.doc_types, visible=False
498
  )
499
 
500
  version.change(
501
+ update_doc_hidden_df_ret_rerank,
502
  version,
503
+ doc_df_elem_ret_rerank_hidden
504
  )
505
 
506
  set_listeners(
 
537
  search_bar_ret = get_search_bar()
538
  with gr.Column(scale=1):
539
  models_ret = get_noreranking_dropdown()
540
+ _doc_df_ret = filter_df_ret(datastore.doc_fmt_df)
 
 
 
 
541
  doc_df_elem_ret = get_leaderboard_table(_doc_df_ret, datastore.doc_types)
542
+ version.change(
543
+ update_doc_df_ret,
544
+ version,
545
+ doc_df_elem_ret
546
+ )
547
 
548
+ _doc_df_ret_hidden = filter_df_ret(datastore.doc_raw_df)
 
 
 
549
  doc_df_elem_ret_hidden = get_leaderboard_table(
550
  _doc_df_ret_hidden, datastore.doc_types, visible=False
551
  )
 
552
  version.change(
553
+ update_doc_hidden_df_ret,
554
  version,
555
+ doc_df_elem_ret_hidden
 
 
 
 
 
 
556
  )
557
 
558
  set_listeners(
 
583
  queue=True,
584
  )
585
  with gr.TabItem("Reranking Only", id=22):
586
+ _doc_df_rerank = filter_df_rerank(datastore.doc_fmt_df)
 
 
 
587
  doc_rerank_models = (
588
  _doc_df_rerank[COL_NAME_RERANKING_MODEL].apply(remove_html).unique().tolist()
589
  )
 
593
  with gr.Column(scale=1):
594
  doc_search_bar_rerank = gr.Textbox(show_label=False, visible=False)
595
  doc_df_elem_rerank = get_leaderboard_table(_doc_df_rerank, datastore.doc_types)
596
+ version.change(
597
+ update_doc_df_rerank,
598
+ version,
599
+ doc_df_elem_rerank
600
+ )
601
+
602
+ _doc_df_rerank_hidden = filter_df_rerank(datastore.doc_raw_df)
603
  doc_df_elem_rerank_hidden = get_leaderboard_table(
604
  _doc_df_rerank_hidden, datastore.doc_types, visible=False
605
  )
606
 
607
  version.change(
608
+ update_doc_hidden_df_rerank,
609
  version,
610
+ doc_df_elem_rerank_hidden
611
  )
612
 
613
  set_listeners(